From 26043ae88f2789fb6372255c8c56aac852234d51 Mon Sep 17 00:00:00 2001 From: Zihua Wu <13583761+lucifer1004@users.noreply.github.com> Date: Fri, 25 Sep 2026 20:02:07 -0700 Subject: [PATCH 01/17] feat(cuda-device): add SharedPtr, a CTA-shared pointer typed addrspace(3) `SharedPtr` is a raw pointer into the issuing CTA's shared memory that carries its address space in its type. It wraps a `*mut SharedSlot`, which the importer types as an `addrspace(3)` pointer, as it already does for pointers to `SharedArray` and `Barrier`, and the importer lowers `SharedPtr` itself to that pointer. Parameters, returns, locals, struct fields and generic code of type `SharedPtr` are therefore CTA-shared pointers by type, across calls to non-inlined helpers, with no dataflow inference. - Constructors: `SharedArray::shared_ptr(&raw mut S)` and `DynamicSharedArray::shared_ptr()` / `shared_ptr_at(bytes)` keep the allocation's shared address; `unsafe SharedPtr::from_raw` is the one explicit locality assertion. - `add` / `offset` / `byte_add` / `cast` keep the type; `as_ptr` widens to a generic pointer; `addr_u32` gives the 32-bit `.shared::cta` address for the `_shared_u32` forms. - The importer drops projections onto `SharedPtr`'s field, which are the identity for this scalar-lowered newtype, so a `SharedPtr` nested in another place is read as a pointer value. - A kernel entry cannot take a `SharedPtr`: the exporter already rejects `addrspace(3)` kernel parameters. - `shared_memory_helper` runs its probe a second time through a non-inlined generic helper that takes the `SharedPtr` in a struct; `verify-code-shape.sh` checks the helper's `ptr addrspace(3)` parameter and that no PTX converts a generic address to shared. Refs #1337, #1238. Co-Authored-By: Claude Opus 5.5 (1M context) Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com> --- crates/cuda-device/src/lib.rs | 2 +- crates/cuda-device/src/shared.rs | 167 ++++++++++++++++++ crates/mir-importer/src/translator/facts.rs | 9 + .../src/translator/rvalue/place_addr.rs | 1 + .../src/translator/rvalue/place_read.rs | 46 +++++ crates/mir-importer/src/translator/types.rs | 16 ++ .../examples/shared_memory_helper/README.md | 23 +++ .../examples/shared_memory_helper/src/main.rs | 128 ++++++++++---- .../shared_memory_helper/verify-code-shape.sh | 30 ++++ .../shared-memory-and-synchronization.md | 42 +++++ .../appendix/api-quick-reference.md | 3 +- 11 files changed, 431 insertions(+), 36 deletions(-) create mode 100755 crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh diff --git a/crates/cuda-device/src/lib.rs b/crates/cuda-device/src/lib.rs index a3171afbcf..ea8b7528f5 100644 --- a/crates/cuda-device/src/lib.rs +++ b/crates/cuda-device/src/lib.rs @@ -82,7 +82,7 @@ pub use cusimd::{CuSimd, Float2, Float4, TmemRegs4, TmemRegs32}; pub use disjoint::{__LaunchContractDisjointSlice, __LaunchContractDisjointSliceAbi}; pub use disjoint::{DisjointSlice, SpaceLayout}; pub use fence::*; -pub use shared::{DynamicSharedArray, SharedArray}; +pub use shared::{DynamicSharedArray, SharedArray, SharedPtr}; pub use tcgen05::{ TensorMemoryHandle, TmemAddress, TmemDeallocated, TmemF32x4, TmemF32x32, TmemGuard, TmemReady, TmemUninit, diff --git a/crates/cuda-device/src/shared.rs b/crates/cuda-device/src/shared.rs index 5467ee7d30..9d1102551a 100644 --- a/crates/cuda-device/src/shared.rs +++ b/crates/cuda-device/src/shared.rs @@ -220,6 +220,23 @@ impl SharedArray { let _ = shared; unreachable!("SharedArray::as_raw_mut_ptr called outside CUDA kernel context") } + + /// Returns a [`SharedPtr`] to the first element of a shared array. + /// + /// Unlike [`Self::as_raw_mut_ptr`], which returns a generic pointer, the + /// result keeps the array's CTA-shared address space in its type, so it + /// stays a shared-memory pointer when it is passed to helper functions. + /// + /// ```rust,ignore + /// static mut TILE: SharedArray = SharedArray::UNINIT; + /// let tile: SharedPtr = SharedArray::shared_ptr(&raw mut TILE); + /// ``` + #[inline(always)] + pub fn shared_ptr(shared: *mut Self) -> SharedPtr { + SharedPtr { + ptr: shared as *mut SharedSlot, + } + } } impl Index for SharedArray { @@ -493,6 +510,24 @@ impl DynamicSharedArray { let _ = byte_offset; unreachable!("DynamicSharedArray::offset called outside CUDA kernel context") } + + /// Like [`Self::get`], but returns a [`SharedPtr`], which keeps the + /// CTA-shared address space in its type. + #[inline(always)] + pub fn shared_ptr() -> SharedPtr { + SharedPtr { + ptr: Self::get() as *mut SharedSlot, + } + } + + /// Like [`Self::offset`], but returns a [`SharedPtr`], which keeps the + /// CTA-shared address space in its type. + #[inline(always)] + pub fn shared_ptr_at(byte_offset: usize) -> SharedPtr { + SharedPtr { + ptr: Self::offset(byte_offset) as *mut SharedSlot, + } + } } /// Convert a generic-address pointer into its raw `.shared` window offset. @@ -546,4 +581,136 @@ pub unsafe fn cvta_generic_to_shared_u32(ptr: *const u8) -> u32 { unreachable!("cvta_generic_to_shared_u32 called outside CUDA kernel context") } +// ============================================================================ +// SharedPtr - a pointer into the issuing CTA's shared memory, typed as such +// ============================================================================ + +/// The pointee of [`SharedPtr`]'s field. A pointer to `SharedSlot` is a +/// pointer into CTA-shared memory (`addrspace(3)`) by its type, as a pointer +/// to [`SharedArray`] is; `repr(transparent)` gives it `T`'s size, so pointer +/// arithmetic strides by `T`. +#[doc(hidden)] +#[repr(transparent)] +pub struct SharedSlot(T); + +/// A raw pointer into the shared memory of the issuing CTA (`.shared::cta`). +/// +/// A `*mut T` in device code is a generic pointer, which may address global, +/// local or shared memory, including another CTA's shared memory in the +/// cluster. `SharedPtr` instead carries its address space in its type: the +/// compiler knows it is a CTA-shared pointer wherever it goes, including +/// through calls to helper functions, and emits the `.shared::cta` forms of +/// instructions that take it without converting a generic address. +/// +/// It is created from shared memory the CTA owns, with +/// [`SharedArray::shared_ptr`] or [`DynamicSharedArray::shared_ptr`], or from +/// a generic pointer with the `unsafe` [`SharedPtr::from_raw`]. Like a raw +/// pointer, it may dangle; dereferencing it goes through [`SharedPtr::as_ptr`] +/// and `unsafe` code. +/// +/// Pointers from [`crate::cluster::map_shared_rank`], which may name another +/// CTA's shared memory, are not `SharedPtr`s. +/// +/// # Example +/// +/// ```rust,ignore +/// static mut TILE: SharedArray = SharedArray::UNINIT; +/// +/// fn store(tile: SharedPtr, i: usize, value: f32) { +/// // `tile` is still a shared pointer here, although `store` is a +/// // separate function. +/// unsafe { tile.add(i).as_ptr().write(value) }; +/// } +/// +/// let tile = SharedArray::shared_ptr(&raw mut TILE); +/// store(tile, thread::threadIdx_x() as usize, 1.0); +/// ``` +#[repr(transparent)] +pub struct SharedPtr { + ptr: *mut SharedSlot, +} + +impl Clone for SharedPtr { + #[inline(always)] + fn clone(&self) -> Self { + *self + } +} + +impl Copy for SharedPtr {} + +impl SharedPtr { + /// Converts a generic pointer to a `SharedPtr`, as CUDA C++'s + /// `__cvta_generic_to_shared` or CuTe's `make_smem_ptr` do. + /// + /// # Safety + /// + /// `ptr` must point into the shared memory of the issuing CTA. A pointer + /// to global or local memory, or to another CTA's shared memory, gives an + /// unspecified address. + #[inline(always)] + pub unsafe fn from_raw(ptr: *mut T) -> Self { + Self { + ptr: ptr as *mut SharedSlot, + } + } + + /// The generic pointer to the same memory, for ordinary loads and stores + /// and for APIs that take raw pointers. + #[inline(always)] + pub fn as_ptr(self) -> *mut T { + self.ptr as *mut T + } + + /// The pointer's 32-bit `.shared::cta` address, as taken by the + /// `_shared_u32` forms of CTA-shared instructions. + #[inline(always)] + pub fn addr_u32(self) -> u32 { + // SAFETY: `self` points into the issuing CTA's shared memory. + unsafe { cvta_generic_to_shared_u32(self.as_ptr() as *const u8) } + } + + /// Reinterprets the pointee type, keeping the address space. + #[inline(always)] + pub fn cast(self) -> SharedPtr { + SharedPtr { + ptr: self.ptr as *mut SharedSlot, + } + } + + /// The pointer `count` elements of `T` further on, as `<*mut T>::add`. + /// + /// # Safety + /// + /// As for `<*mut T>::add`. + #[inline(always)] + pub unsafe fn add(self, count: usize) -> Self { + Self { + ptr: unsafe { self.ptr.add(count) }, + } + } + + /// The pointer `count` elements of `T` away, as `<*mut T>::offset`. + /// + /// # Safety + /// + /// As for `<*mut T>::offset`. + #[inline(always)] + pub unsafe fn offset(self, count: isize) -> Self { + Self { + ptr: unsafe { self.ptr.offset(count) }, + } + } + + /// The pointer `count` bytes further on, as `<*mut T>::byte_add`. + /// + /// # Safety + /// + /// As for `<*mut T>::byte_add`. + #[inline(always)] + pub unsafe fn byte_add(self, count: usize) -> Self { + unsafe { self.cast::().add(count).cast::() } + } +} + include!("generated/shared_sreg.rs"); diff --git a/crates/mir-importer/src/translator/facts.rs b/crates/mir-importer/src/translator/facts.rs index 2be0ceafe8..e585fac95d 100644 --- a/crates/mir-importer/src/translator/facts.rs +++ b/crates/mir-importer/src/translator/facts.rs @@ -389,6 +389,15 @@ mod pointer_origin { } } + /// ABI RULE: `cuda_device::SharedPtr` stores a `*mut SharedSlot`, so + /// the shared pointer it lowers to is always `RawMut`. + pub(crate) fn abi_shared_ptr_field() -> PointerOrigin { + PointerOrigin { + kind: MirPointerKind::RawMut, + mutable: true, + } + } + /// ABI RULE: `cuda_device::DisjointSlice<'_, T>` stores its data pointer /// as `*mut T`, so its data/element pointers are always `RawMut`. pub(crate) fn abi_disjoint_slice_data_ptr() -> PointerOrigin { diff --git a/crates/mir-importer/src/translator/rvalue/place_addr.rs b/crates/mir-importer/src/translator/rvalue/place_addr.rs index 0b09daa48c..f48b6d74b0 100644 --- a/crates/mir-importer/src/translator/rvalue/place_addr.rs +++ b/crates/mir-importer/src/translator/rvalue/place_addr.rs @@ -64,6 +64,7 @@ pub(crate) fn translate_place_address( prev_op: Option>, loc: Location, ) -> TranslationResult>)>> { + let place = &*super::place_read::strip_shared_ptr_field_projections(body, place); let Some(slot) = value_map.get_slot(place.local) else { return Ok(None); }; diff --git a/crates/mir-importer/src/translator/rvalue/place_read.rs b/crates/mir-importer/src/translator/rvalue/place_read.rs index d1e6a0f392..b0fc482f02 100644 --- a/crates/mir-importer/src/translator/rvalue/place_read.rs +++ b/crates/mir-importer/src/translator/rvalue/place_read.rs @@ -80,6 +80,7 @@ pub fn translate_place( prev_op: Option>, loc: Location, ) -> TranslationResult<(Value, Option>)> { + let place = &*strip_shared_ptr_field_projections(body, place); match classify_place_read_strategy(ctx, place, value_map)? { PlaceReadStrategy::Address => { if let Some((value, last_op)) = translate_place_load_from_address( @@ -109,6 +110,51 @@ pub fn translate_place( } } +/// Drop the projections of a place onto the field of a `cuda_device::SharedPtr`. +/// +/// `SharedPtr { ptr: *mut SharedSlot }` is a scalar-lowered newtype: its +/// value is its one field, a CTA-shared pointer (see `types::translate_type`). +/// Projecting that field is therefore the identity. Without this, the place +/// walkers would read the pointer value of a `SharedPtr` nested in another +/// place, such as `(region.base).ptr`, as the address of an aggregate. +pub(crate) fn strip_shared_ptr_field_projections<'a>( + body: &mir::Body, + place: &'a mir::Place, +) -> std::borrow::Cow<'a, mir::Place> { + use rustc_public::ty::{RigidTy, TyKind}; + + if !place + .projection + .iter() + .any(|elem| matches!(elem, mir::ProjectionElem::Field(0, _))) + { + return std::borrow::Cow::Borrowed(place); + } + let mut place_ty = body.locals()[place.local].ty; + let mut kept = Vec::with_capacity(place.projection.len()); + for elem in &place.projection { + let is_shared_ptr = matches!( + place_ty.kind(), + TyKind::RigidTy(RigidTy::Adt(adt_def, _)) if facts::is_cuda_device_adt(&adt_def, "SharedPtr") + ); + let Ok(next_ty) = elem.ty(place_ty) else { + return std::borrow::Cow::Borrowed(place); + }; + if !(is_shared_ptr && matches!(elem, mir::ProjectionElem::Field(0, _))) { + kept.push(elem.clone()); + } + place_ty = next_ty; + } + if kept.len() == place.projection.len() { + std::borrow::Cow::Borrowed(place) + } else { + std::borrow::Cow::Owned(mir::Place { + local: place.local, + projection: kept, + }) + } +} + // ============================================================================ // Place Read Strategy and Address Path // ============================================================================ diff --git a/crates/mir-importer/src/translator/types.rs b/crates/mir-importer/src/translator/types.rs index ee5c23b8e7..f860dbe82e 100644 --- a/crates/mir-importer/src/translator/types.rs +++ b/crates/mir-importer/src/translator/types.rs @@ -364,6 +364,14 @@ fn translate_pointer_like( let elem = shared_array_element_type(ctx, &substs, "SharedArray")?; Ok(facts::mint_shared_ptr_type(ctx, elem, origin).into()) } + rustc_public::ty::TyKind::RigidTy(rustc_public::ty::RigidTy::Adt(adt_def, substs)) + if is_cuda_device_adt(&adt_def, "SharedSlot") => + { + // `*mut SharedSlot` is the field of `SharedPtr`: a pointer + // to a `T` in CTA-shared memory, by its type. + let elem = shared_array_element_type(ctx, &substs, "SharedSlot")?; + Ok(facts::mint_shared_ptr_type(ctx, elem, origin).into()) + } rustc_public::ty::TyKind::RigidTy(rustc_public::ty::RigidTy::Adt(adt_def, _substs)) if is_cuda_device_adt(&adt_def, "Barrier") => { @@ -718,6 +726,14 @@ pub fn translate_type( // ThreadIndex is a newtype around usize - translate to usize // The type safety is enforced at the Rust level, not the IR level Ok(get_usize_type(ctx).into()) + } else if is_cuda_device_adt(&adt_def, "SharedPtr") { + // `SharedPtr { ptr: *mut SharedSlot }` is a scalar-lowered + // newtype: the value is its field, a CTA-shared (`addrspace(3)`) + // pointer to `T`. Params, returns, locals and fields of type + // `SharedPtr` are therefore shared pointers by type, across + // function boundaries. + let elem = shared_array_element_type(ctx, &substs, "SharedPtr")?; + Ok(facts::mint_shared_ptr_type(ctx, elem, facts::abi_shared_ptr_field()).into()) } else if is_cuda_device_adt(&adt_def, "RangeToken") { Ok(dialect_iket::types::IketRangeTokenType::get(ctx).into()) } else if is_cuda_device_adt(&adt_def, "SharedArray") { diff --git a/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md b/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md index 8947f5a809..b764d8a0f3 100644 --- a/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md +++ b/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md @@ -13,6 +13,26 @@ unsafe fn helper(smem: *mut u8, offset: usize) -> u64 { } ``` +Every build also runs the same probe through `SharedPtr`, a CTA-shared +pointer typed as such: + +```rust +let smem = DynamicSharedArray::::shared_ptr(); +shared_ptr_probe(Region { base: smem, offset }); + +#[inline(never)] +unsafe fn shared_ptr_probe(region: Region) -> u64 { + let full = region.base.cast::().byte_add(1024).cast::().as_ptr(); + // ... +} +``` + +`SharedPtr` lowers to an `addrspace(3)` pointer by its type, so the +generic helper's parameter, nested in a struct, is one too, and the call +needs no address-space adaptation. `verify-code-shape.sh` checks that the +helper takes `ptr addrspace(3)` and that no PTX converts a generic address to a +shared one. + The launch supplies 2048 dynamic shared bytes. Four independent CTAs each use one thread. Nine launches cover three data offsets and three input values, including wrapping arithmetic. Every CTA checks both a volatile shared-memory @@ -79,6 +99,9 @@ RTX 5090, NVIDIA driver 580.173.02, and CUDA tools 13.3. | Local helper, all three inline policies | MIR verification rejects the call | | Dependency helper, all three inline policies | Same rejection | +With no features, the `SharedPtr` probe runs after the direct control and +reports `PASS shared-ptr: 9 launches, 36 CTA data/barrier checks`. + Both Compute Sanitizer runs of the direct control reported: ```text diff --git a/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs b/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs index 3abc8dc978..44c1a0fd3a 100644 --- a/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs +++ b/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs @@ -9,7 +9,8 @@ use cuda_core::simt::LaunchConfig; use cuda_core::{CudaContext, DeviceBuffer}; -use cuda_device::{DynamicSharedArray, cuda_module, kernel, launch_bounds, thread}; +use cuda_device::{DynamicSharedArray, SharedPtr, barrier::Barrier}; +use cuda_device::{cuda_module, kernel, launch_bounds, thread}; #[cfg(all(feature = "local-helper", feature = "cross-crate-helper"))] compile_error!("Select at most one helper feature; no features selects the direct control"); @@ -27,6 +28,45 @@ unsafe fn local_probe(smem: *mut u8, offset: usize) -> u64 { shared_memory_helper_lib::shared_probe_body!(smem, offset) } +/// The shared memory a typed probe works on, as a struct field. +#[derive(Clone, Copy)] +struct Region { + base: SharedPtr, + offset: usize, +} + +/// The probe body over a typed shared pointer, in a generic helper that is not +/// inlined. Unlike the raw `*mut u8` helpers above, it compiles: `SharedPtr` +/// is a CTA-shared pointer by its type, so the helper's parameter is one too, +/// and the barrier is addressed without converting a generic pointer. +/// +/// # Safety +/// As for `shared_probe_body!`: 2048 live shared bytes aligned to 1024, one +/// thread per CTA, and an initialized u64 at `offset`, below byte 1024. +#[inline(never)] +unsafe fn shared_ptr_probe(region: Region) -> u64 { + unsafe { + let bytes = region.base.cast::(); + let full: *mut Barrier = bytes.byte_add(1024).cast::().as_ptr(); + let data = bytes.byte_add(region.offset).cast::().as_ptr(); + cuda_device::barrier::mbarrier_init(full, 1); + let pending = !cuda_device::barrier::mbarrier_try_wait_parity(full, 0); + let _token = cuda_device::barrier::mbarrier_arrive(full); + let mut complete = 0u64; + let mut attempts = 0u32; + while attempts < 1024 { + if cuda_device::barrier::mbarrier_try_wait_parity(full, 0) { + complete = if pending { 1 } else { 2 }; + break; + } + attempts += 1; + } + let value = core::ptr::read_volatile(data); + core::ptr::write_volatile(data, value.wrapping_mul(3) ^ 0x1234_5678); + complete + } +} + #[cuda_module] mod kernels { use super::*; @@ -52,6 +92,23 @@ mod kernels { *output.add(block * 2 + 1) = completed; } } + + /// The same probe through `SharedPtr`, a typed CTA-shared pointer, passed + /// to a non-inlined generic helper inside a struct. + #[kernel] + #[launch_bounds(1)] + pub fn shared_ptr_memory_probe(output: *mut u64, seed: u64, byte_offset: u64) { + let smem = DynamicSharedArray::::shared_ptr(); + let block = thread::blockIdx_x() as usize; + let offset = byte_offset as usize; + unsafe { + let data = smem.byte_add(offset).cast::().as_ptr(); + core::ptr::write_volatile(data, seed.wrapping_add(block as u64)); + let completed = shared_ptr_probe(Region { base: smem, offset }); + *output.add(block * 2) = core::ptr::read_volatile(data); + *output.add(block * 2 + 1) = completed; + } + } } fn main() -> Result<(), Box> { @@ -79,41 +136,44 @@ fn main() -> Result<(), Box> { block_dim: (1, 1, 1), shared_mem_bytes: 2048, }; - let mut cases = 0; - for seed in [0u64, 17, u64::MAX] { - for offset in [0u64, 64, 248] { - // SAFETY: four independent CTAs, one thread each; output holds two - // u64s per CTA. Data offsets are aligned and disjoint from the - // barrier. Every launch receives a fresh shared-memory allocation. - unsafe { - module.shared_memory_probe( - &stream, - config, - output.cu_deviceptr() as *mut u64, - seed, - offset, - )?; - } - let got = output.to_host_vec(&stream)?; - for block in 0..BLOCKS { - let expected = seed.wrapping_add(block as u64).wrapping_mul(3) ^ 0x1234_5678; - assert_eq!( - got[block * 2], - expected, - "{variant}: data, seed={seed}, offset={offset}, block={block}" - ); - assert_eq!( - got[block * 2 + 1], - 1, - "{variant}: barrier, seed={seed}, offset={offset}, block={block}" - ); + for typed in [false, true] { + let variant = if typed { "shared-ptr" } else { variant }; + let mut cases = 0; + for seed in [0u64, 17, u64::MAX] { + for offset in [0u64, 64, 248] { + let out = output.cu_deviceptr() as *mut u64; + // SAFETY: four independent CTAs, one thread each; output holds + // two u64s per CTA. Data offsets are aligned and disjoint from + // the barrier. Every launch receives a fresh shared-memory + // allocation. + unsafe { + if typed { + module.shared_ptr_memory_probe(&stream, config, out, seed, offset)?; + } else { + module.shared_memory_probe(&stream, config, out, seed, offset)?; + } + } + let got = output.to_host_vec(&stream)?; + for block in 0..BLOCKS { + let expected = seed.wrapping_add(block as u64).wrapping_mul(3) ^ 0x1234_5678; + assert_eq!( + got[block * 2], + expected, + "{variant}: data, seed={seed}, offset={offset}, block={block}" + ); + assert_eq!( + got[block * 2 + 1], + 1, + "{variant}: barrier, seed={seed}, offset={offset}, block={block}" + ); + } + cases += 1; } - cases += 1; } + println!( + "PASS {variant}: {cases} launches, {} CTA data/barrier checks", + cases * BLOCKS + ); } - println!( - "PASS {variant}: {cases} launches, {} CTA data/barrier checks", - cases * BLOCKS - ); Ok(()) } diff --git a/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh b/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh new file mode 100755 index 0000000000..8ffe705a18 --- /dev/null +++ b/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh @@ -0,0 +1,30 @@ +#!/usr/bin/env bash +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +# +# `SharedPtr` keeps a CTA-shared pointer typed `addrspace(3)` across a call: +# the non-inlined `shared_ptr_probe` helper takes one as its parameter, and +# no PTX converts a generic address to a shared one (`cvta.to.shared`). +set -euo pipefail + +dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +ll="${dir}/shared_memory_helper.ll" +ptx="${dir}/shared_memory_helper.ptx" + +for file in "${ll}" "${ptx}"; do + [[ -f "${file}" ]] || { echo "FAIL: ${file} not found; build the example first" >&2; exit 1; } +done + +if ! grep -Eq '^define .*shared_ptr_probe.*\(ptr addrspace\(3\) ' "${ll}"; then + echo "FAIL: shared_ptr_probe does not take a ptr addrspace(3) parameter" >&2 + grep -E '^define .*shared_ptr_probe' "${ll}" >&2 || true + exit 1 +fi + +if grep -q 'cvta\.to\.shared' "${ptx}"; then + echo "FAIL: PTX converts a generic address to a shared one" >&2 + grep -n 'cvta\.to\.shared' "${ptx}" >&2 + exit 1 +fi + +echo "PASS: SharedPtr stays addrspace(3) across shared_ptr_probe, no cvta.to.shared" diff --git a/cuda-oxide-book/advanced/shared-memory-and-synchronization.md b/cuda-oxide-book/advanced/shared-memory-and-synchronization.md index c068c8af4f..8ffa7d0097 100644 --- a/cuda-oxide-book/advanced/shared-memory-and-synchronization.md +++ b/cuda-oxide-book/advanced/shared-memory-and-synchronization.md @@ -233,6 +233,48 @@ total does not exceed `shared_mem_bytes` — there are no runtime guards. (Hopper+), use `DynamicSharedArray` to get the required 128-byte alignment. The alignment is encoded in the PTX `.align` directive. +### Typed shared pointers: `SharedPtr` + +`DynamicSharedArray::get()` and `SharedArray::as_mut_ptr()` return `*mut T`, +which in device code is a *generic* pointer: it may address global, local or +shared memory, including another block's shared memory in a cluster. Once +such a pointer is passed to a helper function, the compiler no longer knows it +points into this block's shared memory. + +`SharedPtr` is a raw pointer into the issuing block's shared memory +(`.shared::cta`) that carries that fact in its type. It lowers to an +`addrspace(3)` pointer wherever it goes, including helper parameters and +return values, struct fields and generic code, so instructions that take it +use their shared-memory forms without converting a generic address. + +```rust +use cuda_device::{DynamicSharedArray, SharedArray, SharedPtr}; + +#[inline(never)] +unsafe fn fill(tile: SharedPtr, i: usize, value: f32) { + // Still a shared-memory pointer here, although `fill` is not inlined. + unsafe { tile.add(i).as_ptr().write(value) }; +} + +static mut TILE: SharedArray = SharedArray::UNINIT; +let tile: SharedPtr = SharedArray::shared_ptr(&raw mut TILE); +let pool: SharedPtr = DynamicSharedArray::::shared_ptr_at(1024); +unsafe { fill(tile, thread::threadIdx_x() as usize, 1.0) }; +``` + +| Operation | Result | +|:---------------------------------------|:----------------------------------------------------| +| `SharedArray::shared_ptr(&raw mut S)` | `SharedPtr` to the first element of `S` | +| `DynamicSharedArray::::shared_ptr()` / `shared_ptr_at(bytes)` | Like `get()` / `offset(bytes)` | +| `unsafe SharedPtr::from_raw(p)` | Asserts that generic `p` points into this block's shared memory | +| `add`, `offset`, `byte_add`, `cast` | Pointer arithmetic that keeps the type | +| `as_ptr()` | The generic `*mut T`, for loads, stores and raw-pointer APIs | +| `addr_u32()` | The 32-bit `.shared::cta` address, for `_shared_u32` forms | + +`SharedPtr` cannot be a kernel parameter: the host has no shared-memory +address to pass. Pointers from `cluster::map_shared_rank` may name another +block's shared memory and are not `SharedPtr`s. + --- ## Synchronization: sync_threads() diff --git a/cuda-oxide-book/appendix/api-quick-reference.md b/cuda-oxide-book/appendix/api-quick-reference.md index 6a986c9192..64ab9ee8ff 100644 --- a/cuda-oxide-book/appendix/api-quick-reference.md +++ b/cuda-oxide-book/appendix/api-quick-reference.md @@ -310,6 +310,7 @@ pub fn dynamic(data: &[f32]) { | `SharedArray` | Compile-time sized, block-scoped shared memory | | `SharedArray` | With 128-byte alignment (required for TMA destinations) | | `DynamicSharedArray` | Runtime-sized shared memory (set via `LaunchConfig`) | +| `SharedPtr` | Pointer into this block's shared memory, `addrspace(3)` by type | Both are `!Sync` — concurrent access requires explicit barriers. @@ -647,7 +648,7 @@ debug::prof_trigger::<7>(); // Nsight profiler trigger | `thread` | Thread/block IDs, `index_1d`, `sync_threads` | All | | `config` | Compile-time policies, shapes, tiles, atoms, layouts, memory spaces, and scopes | All | | `disjoint` | `DisjointSlice` — typed writes completed by a launch proof | All | -| `shared` | `SharedArray`, `DynamicSharedArray` | All | +| `shared` | `SharedArray`, `DynamicSharedArray`, `SharedPtr` | All | | `warp` | Shuffle, vote, match, lane/warp ID | All | | `atomic` | Scoped atomics (device/block/system) | sm_70+ | | `debug` | `clock64`, `trap`, `breakpoint`, `gpu_printf!` | All | From 8287263db389f47929b6483b659744f363ce6a3d Mon Sep 17 00:00:00 2001 From: Zihua Wu <13583761+lucifer1004@users.noreply.github.com> Date: Fri, 25 Sep 2026 16:50:44 -0700 Subject: [PATCH 02/17] feat(tma): add CTA-local G2S tile intrinsics Add cp_async_bulk_tensor_{1..5}d_g2s_cta, the TMA tile load into the issuing CTA's own shared memory (cp.async.bulk.tensor.Nd.shared::cta.global...). Its destination is a `SharedPtr`, a CTA-shared pointer by type (#1337), so the locality the `.shared::cta` form needs comes from the type rather than from the safety contract, including through helper functions. The raw ABI keeps `*mut u8`, as `*mut Barrier` keeps `*mut u64`. The existing cluster-shared G2S family is unchanged. On SM12x, ptxas implements a non-multicast `.shared::cluster` tensor load it cannot prove CTA-local with a runtime check and a syscall, which makes it ignore `setmaxnreg` (#1238); this form avoids that. - LLVM NVPTX lowers to the typed llvm.nvvm.cp.async.bulk.tensor.g2s.cta intrinsics; libNVVM uses the exact inline PTX, as the cluster forms do. - A destination or barrier known to be cluster shared (addrspace 7) is a located compile error rather than a local offset. - The PTX version detector also recognizes the typed intrinsics, so modules using them get PTX 8.6. - Tests: mir-lower conversion tests for both backends and a negative addrspace(7) test, a detector test, the extended TMA admission ordering assertion, and a tma_copy kernel whose code-shape check requires the `.shared::cta` form with no address conversion. Co-Authored-By: Claude Opus 5.5 (1M context) Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com> --- .../cuda-device/src/generated/async_copy.rs | 2 +- crates/cuda-device/src/generated/atomic.rs | 2 +- crates/cuda-device/src/generated/bf16.rs | 2 +- crates/cuda-device/src/generated/bf16x2.rs | 2 +- crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- crates/cuda-device/src/generated/convert.rs | 2 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- crates/cuda-device/src/generated/dotprod.rs | 2 +- crates/cuda-device/src/generated/f16.rs | 2 +- crates/cuda-device/src/generated/f16x2.rs | 2 +- crates/cuda-device/src/generated/f32x2.rs | 2 +- crates/cuda-device/src/generated/fence.rs | 2 +- crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- crates/cuda-device/src/generated/grid_sreg.rs | 2 +- crates/cuda-device/src/generated/i16x2.rs | 2 +- crates/cuda-device/src/generated/int.rs | 2 +- crates/cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- crates/cuda-device/src/generated/movmatrix.rs | 2 +- crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- crates/cuda-device/src/generated/sreg.rs | 2 +- crates/cuda-device/src/generated/stmatrix.rs | 2 +- crates/cuda-device/src/generated/tcgen05.rs | 2 +- crates/cuda-device/src/generated/tma.rs | 100 +- crates/cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- crates/cuda-device/src/tma.rs | 1 + .../src/model/contracts/tma.rs | 25 +- .../cuda-intrinsics-gen/src/model/overlay.rs | 2 + .../cuda-intrinsics-gen/src/render/compat.rs | 19 +- .../src/render/importer.rs | 12 + .../src/render/lowering.rs | 13 + .../cuda-intrinsics-gen/src/render/probes.rs | 61 +- .../src/render/tests/mma.rs | 2 +- .../src/render/tests/mod.rs | 2 +- .../src/render/tests/tensor.rs | 10 +- .../src/render/validate.rs | 7 +- .../cuda-intrinsics-gen/src/resolve/driver.rs | 2 + .../src/resolve/families/tma.rs | 124 +- .../cuda-intrinsics-gen/src/resolve/guards.rs | 69 +- .../src/resolve/tests/fixtures.rs | 2 + .../src/resolve/tests/overlay_schema.rs | 2 +- .../src/resolve/tests/tma_clc_cluster.rs | 7 +- .../cuda-intrinsics/src/generated/abi_v1.rs | 108 +- crates/cuda-intrinsics/src/generated/mod.rs | 47 +- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 2 +- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 347 ++- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 292 ++- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- .../cuda-oxide-codegen/src/target/detect.rs | 10 +- .../target/tests/detection_async_memory.rs | 9 + .../src/ops/generated/active_mask.rs | 2 +- crates/dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- crates/dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 2 +- crates/dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- crates/dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- crates/dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- crates/dialect-nvvm/src/ops/generated/tma.rs | 82 +- crates/dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 27 +- .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 2 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 2 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 182 +- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 37 +- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../translator/terminator/intrinsics/tma.rs | 111 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 2 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 72 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../mir-lower/src/convert/intrinsics/tma.rs | 164 +- crates/mir-lower/tests/lowering_test/tma.rs | 144 ++ .../examples/tma_copy/README.md | 12 +- .../examples/tma_copy/src/main.rs | 108 +- .../examples/tma_copy/verify-code-shape.sh | 75 + .../src/generated_intrinsics.rs | 22 +- .../advanced/tensor-memory-accelerator.md | 21 + .../appendix/api-quick-reference.md | 1 + cuda-oxide-book/compiler/mlir-dialects.md | 10 +- intrinsics/abi-v1.toml | 35 + intrinsics/catalog.json | 2248 +++++++++++++---- .../cuda-13.3-libnvvm-13.3.73-tma-cta.json | 109 + .../rust-llvm-23.1.0-16696adc-tma-cta.json | 116 + intrinsics/generated-reference.md | 72 +- intrinsics/overlay/tma.toml | 22 + intrinsics/probes/abs_bf16x2.ll | 2 +- intrinsics/probes/abs_f16x2.ll | 2 +- intrinsics/probes/active_mask.ll | 2 +- intrinsics/probes/add_bf16x2.ll | 2 +- intrinsics/probes/add_f16x2.ll | 2 +- intrinsics/probes/add_f32x2.ll | 2 +- intrinsics/probes/add_ftz_f32x2.ll | 2 +- intrinsics/probes/add_rm_f32.ll | 2 +- intrinsics/probes/add_rm_f64.ll | 2 +- intrinsics/probes/add_rm_ftz_f32.ll | 2 +- intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rm_sat_f32.ll | 2 +- intrinsics/probes/add_rn_f32.ll | 2 +- intrinsics/probes/add_rn_f64.ll | 2 +- intrinsics/probes/add_rn_ftz_f32.ll | 2 +- intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rn_sat_f32.ll | 2 +- intrinsics/probes/add_rp_f32.ll | 2 +- intrinsics/probes/add_rp_f64.ll | 2 +- intrinsics/probes/add_rp_ftz_f32.ll | 2 +- intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rp_sat_f32.ll | 2 +- intrinsics/probes/add_rz_f32.ll | 2 +- intrinsics/probes/add_rz_f64.ll | 2 +- intrinsics/probes/add_rz_ftz_f32.ll | 2 +- intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rz_sat_f32.ll | 2 +- intrinsics/probes/all_sync.ll | 2 +- intrinsics/probes/any_sync.ll | 2 +- intrinsics/probes/ballot_sync.ll | 2 +- intrinsics/probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- intrinsics/probes/barrier_cta_sync.ll | 2 +- intrinsics/probes/barrier_cta_sync_aligned.ll | 2 +- intrinsics/probes/block_dim_x.ll | 2 +- intrinsics/probes/block_dim_y.ll | 2 +- intrinsics/probes/block_dim_z.ll | 2 +- intrinsics/probes/block_idx_x.ll | 2 +- intrinsics/probes/block_idx_y.ll | 2 +- intrinsics/probes/block_idx_z.ll | 2 +- intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- intrinsics/probes/clc_query_is_canceled.ll | 2 +- intrinsics/probes/clc_try_cancel.ll | 2 +- intrinsics/probes/clc_try_cancel_multicast.ll | 2 +- intrinsics/probes/clock.lib_nvvm.ll | 2 +- intrinsics/probes/clock.llvm_nvptx.ll | 2 +- intrinsics/probes/clock64.lib_nvvm.ll | 2 +- intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- intrinsics/probes/cluster_block_count.ll | 2 +- intrinsics/probes/cluster_block_idx_x.ll | 2 +- intrinsics/probes/cluster_block_idx_y.ll | 2 +- intrinsics/probes/cluster_block_idx_z.ll | 2 +- intrinsics/probes/cluster_block_rank.ll | 2 +- intrinsics/probes/cluster_dim_x.ll | 2 +- intrinsics/probes/cluster_dim_y.ll | 2 +- intrinsics/probes/cluster_dim_z.ll | 2 +- intrinsics/probes/cluster_grid_dim_x.ll | 2 +- intrinsics/probes/cluster_grid_dim_y.ll | 2 +- intrinsics/probes/cluster_grid_dim_z.ll | 2 +- intrinsics/probes/cluster_idx_x.ll | 2 +- intrinsics/probes/cluster_idx_y.ll | 2 +- intrinsics/probes/cluster_idx_z.ll | 2 +- intrinsics/probes/cos_approx_f32.ll | 2 +- intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s_cta.ll | 18 + .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_cta.ll | 18 + .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s_cta.ll | 18 + .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s_cta.ll | 18 + .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s_cta.ll | 18 + .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- intrinsics/probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- intrinsics/probes/cp_async_ca_16.ll | 2 +- intrinsics/probes/cp_async_ca_4.ll | 2 +- intrinsics/probes/cp_async_ca_8.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- intrinsics/probes/cp_async_cg_16.ll | 2 +- intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- intrinsics/probes/cp_async_commit_group.ll | 2 +- intrinsics/probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- intrinsics/probes/cp_async_wait_all.ll | 2 +- intrinsics/probes/cp_async_wait_group.ll | 2 +- intrinsics/probes/cvt_f16x2_f32.ll | 2 +- intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- intrinsics/probes/div_rm_f32.ll | 2 +- intrinsics/probes/div_rm_f64.ll | 2 +- intrinsics/probes/div_rm_ftz_f32.ll | 2 +- intrinsics/probes/div_rn_f32.ll | 2 +- intrinsics/probes/div_rn_f64.ll | 2 +- intrinsics/probes/div_rn_ftz_f32.ll | 2 +- intrinsics/probes/div_rp_f32.ll | 2 +- intrinsics/probes/div_rp_f64.ll | 2 +- intrinsics/probes/div_rp_ftz_f32.ll | 2 +- intrinsics/probes/div_rz_f32.ll | 2 +- intrinsics/probes/div_rz_f64.ll | 2 +- intrinsics/probes/div_rz_ftz_f32.ll | 2 +- intrinsics/probes/dp2a_s32.ll | 2 +- intrinsics/probes/dp2a_u32.ll | 2 +- intrinsics/probes/dp4a_s32.ll | 2 +- intrinsics/probes/dp4a_u32.ll | 2 +- intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- intrinsics/probes/elect_sync.llvm_nvptx.ll | 2 +- intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- intrinsics/probes/ex2_approx_f16.ll | 2 +- intrinsics/probes/ex2_approx_f32.ll | 2 +- intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- intrinsics/probes/fma_bf16x2.ll | 2 +- intrinsics/probes/fma_f16x2.ll | 2 +- intrinsics/probes/fma_f32x2.ll | 2 +- intrinsics/probes/fma_ftz_f16x2.ll | 2 +- intrinsics/probes/fma_ftz_f32x2.ll | 2 +- intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- intrinsics/probes/fma_relu_bf16x2.ll | 2 +- intrinsics/probes/fma_relu_f16x2.ll | 2 +- intrinsics/probes/fma_rm_f32.ll | 2 +- intrinsics/probes/fma_rm_f64.ll | 2 +- intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rm_sat_f32.ll | 2 +- intrinsics/probes/fma_rn_f32.ll | 2 +- intrinsics/probes/fma_rn_f64.ll | 2 +- intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rn_sat_f32.ll | 2 +- intrinsics/probes/fma_rp_f32.ll | 2 +- intrinsics/probes/fma_rp_f64.ll | 2 +- intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rp_sat_f32.ll | 2 +- intrinsics/probes/fma_rz_f32.ll | 2 +- intrinsics/probes/fma_rz_f64.ll | 2 +- intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rz_sat_f32.ll | 2 +- intrinsics/probes/fma_sat_f16x2.ll | 2 +- intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- intrinsics/probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- intrinsics/probes/grid_dependency_wait.ll | 2 +- intrinsics/probes/grid_dim_x.ll | 2 +- intrinsics/probes/grid_dim_y.ll | 2 +- intrinsics/probes/grid_dim_z.ll | 2 +- intrinsics/probes/gridid.lib_nvvm.ll | 2 +- intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- intrinsics/probes/lane_id.ll | 2 +- intrinsics/probes/lanemask_eq.ll | 2 +- intrinsics/probes/lanemask_ge.ll | 2 +- intrinsics/probes/lanemask_gt.ll | 2 +- intrinsics/probes/lanemask_le.ll | 2 +- intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- intrinsics/probes/lg2_approx_f32.ll | 2 +- intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- intrinsics/probes/map_shared_rank.ll | 2 +- intrinsics/probes/match_all_i64_sync.ll | 2 +- intrinsics/probes/match_all_sync.ll | 2 +- intrinsics/probes/match_any_i64_sync.ll | 2 +- intrinsics/probes/match_any_sync.ll | 2 +- intrinsics/probes/max_bf16.ll | 2 +- intrinsics/probes/max_bf16x2.ll | 2 +- intrinsics/probes/max_f16.ll | 2 +- intrinsics/probes/max_f16x2.ll | 2 +- intrinsics/probes/max_ftz_f16.ll | 2 +- intrinsics/probes/max_ftz_f16x2.ll | 2 +- intrinsics/probes/max_ftz_nan_f16.ll | 2 +- intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_ftz_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_nan_bf16.ll | 2 +- intrinsics/probes/max_nan_bf16x2.ll | 2 +- intrinsics/probes/max_nan_f16.ll | 2 +- intrinsics/probes/max_nan_f16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_relu_s16x2.ll | 2 +- intrinsics/probes/max_relu_s32.ll | 2 +- intrinsics/probes/max_s16x2.ll | 2 +- intrinsics/probes/max_u16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- intrinsics/probes/max_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- intrinsics/probes/max_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- intrinsics/probes/mbarrier_arrive.ll | 2 +- intrinsics/probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- intrinsics/probes/mbarrier_init.ll | 2 +- intrinsics/probes/mbarrier_inval.ll | 2 +- intrinsics/probes/mbarrier_test_wait.ll | 2 +- intrinsics/probes/mbarrier_try_wait.ll | 2 +- intrinsics/probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- intrinsics/probes/min_bf16.ll | 2 +- intrinsics/probes/min_bf16x2.ll | 2 +- intrinsics/probes/min_f16.ll | 2 +- intrinsics/probes/min_f16x2.ll | 2 +- intrinsics/probes/min_ftz_f16.ll | 2 +- intrinsics/probes/min_ftz_f16x2.ll | 2 +- intrinsics/probes/min_ftz_nan_f16.ll | 2 +- intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_ftz_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_nan_bf16.ll | 2 +- intrinsics/probes/min_nan_bf16x2.ll | 2 +- intrinsics/probes/min_nan_f16.ll | 2 +- intrinsics/probes/min_nan_f16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_relu_s16x2.ll | 2 +- intrinsics/probes/min_relu_s32.ll | 2 +- intrinsics/probes/min_s16x2.ll | 2 +- intrinsics/probes/min_u16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- intrinsics/probes/min_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- intrinsics/probes/min_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- intrinsics/probes/mma_m16n8k16_f32_bf16.ll | 2 +- intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- intrinsics/probes/movmatrix_trans_b16.ll | 2 +- intrinsics/probes/mul_bf16x2.ll | 2 +- intrinsics/probes/mul_f16x2.ll | 2 +- intrinsics/probes/mul_f32x2.ll | 2 +- intrinsics/probes/mul_ftz_f32x2.ll | 2 +- intrinsics/probes/mul_rm_f32.ll | 2 +- intrinsics/probes/mul_rm_f64.ll | 2 +- intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- intrinsics/probes/mul_rn_f32.ll | 2 +- intrinsics/probes/mul_rn_f64.ll | 2 +- intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- intrinsics/probes/mul_rp_f32.ll | 2 +- intrinsics/probes/mul_rp_f64.ll | 2 +- intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- intrinsics/probes/mul_rz_f32.ll | 2 +- intrinsics/probes/mul_rz_f64.ll | 2 +- intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- intrinsics/probes/nanosleep.ll | 2 +- intrinsics/probes/neg_bf16x2.ll | 2 +- intrinsics/probes/neg_f16x2.ll | 2 +- intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- intrinsics/probes/packed_atomic_add_bf16x2.ll | 2 +- intrinsics/probes/packed_atomic_add_f16x2.ll | 2 +- intrinsics/probes/pmevent.ll | 2 +- intrinsics/probes/prefetch_tma_descriptor.ll | 2 +- intrinsics/probes/prmt.ll | 2 +- intrinsics/probes/prmt_b4e.ll | 2 +- intrinsics/probes/prmt_ecl.ll | 2 +- intrinsics/probes/prmt_ecr.ll | 2 +- intrinsics/probes/prmt_f4e.ll | 2 +- intrinsics/probes/prmt_rc16.ll | 2 +- intrinsics/probes/prmt_rc8.ll | 2 +- intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- intrinsics/probes/rcp_rm_f32.ll | 2 +- intrinsics/probes/rcp_rm_f64.ll | 2 +- intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rn_f32.ll | 2 +- intrinsics/probes/rcp_rn_f64.ll | 2 +- intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rp_f32.ll | 2 +- intrinsics/probes/rcp_rp_f64.ll | 2 +- intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rz_f32.ll | 2 +- intrinsics/probes/rcp_rz_f64.ll | 2 +- intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- intrinsics/probes/redux_sync_add.ll | 2 +- intrinsics/probes/redux_sync_and.ll | 2 +- intrinsics/probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_max_f32.ll | 2 +- intrinsics/probes/redux_sync_max_i32.ll | 2 +- intrinsics/probes/redux_sync_max_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_max_u32.ll | 2 +- intrinsics/probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_min_f32.ll | 2 +- intrinsics/probes/redux_sync_min_i32.ll | 2 +- intrinsics/probes/redux_sync_min_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_min_u32.ll | 2 +- intrinsics/probes/redux_sync_or.ll | 2 +- intrinsics/probes/redux_sync_xor.ll | 2 +- intrinsics/probes/rsqrt_approx_f32.ll | 2 +- intrinsics/probes/rsqrt_approx_f64.ll | 2 +- intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- intrinsics/probes/setmaxnreg_dec.ll | 2 +- intrinsics/probes/setmaxnreg_inc.ll | 2 +- intrinsics/probes/shuffle_down_f32_sync.ll | 2 +- intrinsics/probes/shuffle_down_sync.ll | 2 +- intrinsics/probes/shuffle_down_u64_sync.ll | 2 +- intrinsics/probes/shuffle_f32_sync.ll | 2 +- intrinsics/probes/shuffle_sync.ll | 2 +- intrinsics/probes/shuffle_u64_sync.ll | 2 +- intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- intrinsics/probes/shuffle_up_sync.ll | 2 +- intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- intrinsics/probes/shuffle_xor_sync.ll | 2 +- intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- intrinsics/probes/sin_approx_f32.ll | 2 +- intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- intrinsics/probes/smid.lib_nvvm.ll | 2 +- intrinsics/probes/smid.llvm_nvptx.ll | 2 +- intrinsics/probes/sqrt_approx_f32.ll | 2 +- intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rm_f32.ll | 2 +- intrinsics/probes/sqrt_rm_f64.ll | 2 +- intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rn_f32.ll | 2 +- intrinsics/probes/sqrt_rn_f64.ll | 2 +- intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rp_f32.ll | 2 +- intrinsics/probes/sqrt_rp_f64.ll | 2 +- intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rz_f32.ll | 2 +- intrinsics/probes/sqrt_rz_f64.ll | 2 +- intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- intrinsics/probes/sub_bf16x2.ll | 2 +- intrinsics/probes/sub_f16x2.ll | 2 +- intrinsics/probes/sub_f32x2.ll | 2 +- intrinsics/probes/sub_ftz_f32x2.ll | 2 +- intrinsics/probes/sync_mask.ll | 2 +- intrinsics/probes/sync_threads.ll | 2 +- intrinsics/probes/tanh_approx_f32.ll | 2 +- intrinsics/probes/tcgen05_alloc.ll | 2 +- intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- intrinsics/probes/tcgen05_commit.ll | 2 +- intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- intrinsics/probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- intrinsics/probes/tcgen05_dealloc.ll | 2 +- intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- intrinsics/probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- intrinsics/probes/tcgen05_load_wait.ll | 2 +- intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- intrinsics/probes/tcgen05_mma_f16.ll | 2 +- intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- intrinsics/probes/tcgen05_mma_shared.ll | 2 +- intrinsics/probes/tcgen05_mma_sp_shared.ll | 2 +- intrinsics/probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- intrinsics/probes/tcgen05_shift_down.ll | 2 +- intrinsics/probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- intrinsics/probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- intrinsics/probes/thread_idx_x.ll | 2 +- intrinsics/probes/thread_idx_y.ll | 2 +- intrinsics/probes/thread_idx_z.ll | 2 +- intrinsics/probes/threadfence.ll | 2 +- intrinsics/probes/threadfence_block.ll | 2 +- intrinsics/probes/threadfence_system.ll | 2 +- intrinsics/probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- intrinsics/probes/trap.ll | 2 +- intrinsics/probes/uni_sync.ll | 2 +- intrinsics/probes/warpid.lib_nvvm.ll | 2 +- intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- intrinsics/probes/wgmma_commit_group.ll | 2 +- intrinsics/probes/wgmma_fence.ll | 2 +- intrinsics/probes/wgmma_wait_group.ll | 2 +- 1353 files changed, 5770 insertions(+), 1869 deletions(-) create mode 100755 crates/rustc-codegen-cuda/examples/tma_copy/verify-code-shape.sh create mode 100644 intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-tma-cta.json create mode 100644 intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json create mode 100644 intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll create mode 100644 intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll create mode 100644 intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll create mode 100644 intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll create mode 100644 intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll diff --git a/crates/cuda-device/src/generated/async_copy.rs b/crates/cuda-device/src/generated/async_copy.rs index e44c9a333c..49d645697e 100644 --- a/crates/cuda-device/src/generated/async_copy.rs +++ b/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/atomic.rs b/crates/cuda-device/src/generated/atomic.rs index 47a5053d24..21cd96ef16 100644 --- a/crates/cuda-device/src/generated/atomic.rs +++ b/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/bf16.rs b/crates/cuda-device/src/generated/bf16.rs index fb8215c889..cf2e860f6d 100644 --- a/crates/cuda-device/src/generated/bf16.rs +++ b/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/bf16x2.rs b/crates/cuda-device/src/generated/bf16x2.rs index b9f4be0b75..8a660332c7 100644 --- a/crates/cuda-device/src/generated/bf16x2.rs +++ b/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/clc.rs b/crates/cuda-device/src/generated/clc.rs index 94c9cae176..188917564e 100644 --- a/crates/cuda-device/src/generated/clc.rs +++ b/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/cluster_barrier.rs b/crates/cuda-device/src/generated/cluster_barrier.rs index 9115aec1b5..ab3444e3eb 100644 --- a/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/cluster_memory.rs b/crates/cuda-device/src/generated/cluster_memory.rs index d665c6fc59..c5c84b4586 100644 --- a/crates/cuda-device/src/generated/cluster_memory.rs +++ b/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/cluster_sreg.rs b/crates/cuda-device/src/generated/cluster_sreg.rs index 4e2ce888c0..930f5197f5 100644 --- a/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included inside `cuda_device::cluster`. diff --git a/crates/cuda-device/src/generated/convert.rs b/crates/cuda-device/src/generated/convert.rs index f84862afdd..25bfbe04af 100644 --- a/crates/cuda-device/src/generated/convert.rs +++ b/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::convert`. diff --git a/crates/cuda-device/src/generated/counted_barrier.rs b/crates/cuda-device/src/generated/counted_barrier.rs index 877e00f15f..355580f20b 100644 --- a/crates/cuda-device/src/generated/counted_barrier.rs +++ b/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/debug_control.rs b/crates/cuda-device/src/generated/debug_control.rs index 6f52412e9a..d9959561b4 100644 --- a/crates/cuda-device/src/generated/debug_control.rs +++ b/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/debug_sreg.rs b/crates/cuda-device/src/generated/debug_sreg.rs index 2028a15d31..ff9115278b 100644 --- a/crates/cuda-device/src/generated/debug_sreg.rs +++ b/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/dotprod.rs b/crates/cuda-device/src/generated/dotprod.rs index 2e1e16f829..cf302824e2 100644 --- a/crates/cuda-device/src/generated/dotprod.rs +++ b/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f16.rs b/crates/cuda-device/src/generated/f16.rs index e0308d047a..03aed15536 100644 --- a/crates/cuda-device/src/generated/f16.rs +++ b/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f16x2.rs b/crates/cuda-device/src/generated/f16x2.rs index 00085fa0b6..f08208f707 100644 --- a/crates/cuda-device/src/generated/f16x2.rs +++ b/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f32x2.rs b/crates/cuda-device/src/generated/f32x2.rs index c5de9a8539..f6c4999526 100644 --- a/crates/cuda-device/src/generated/f32x2.rs +++ b/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/fence.rs b/crates/cuda-device/src/generated/fence.rs index 685f7e4c38..add71d3a89 100644 --- a/crates/cuda-device/src/generated/fence.rs +++ b/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/crates/cuda-device/src/generated/float.rs b/crates/cuda-device/src/generated/float.rs index fbe53bd878..f02e22fee6 100644 --- a/crates/cuda-device/src/generated/float.rs +++ b/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::float`. diff --git a/crates/cuda-device/src/generated/grid_dependency.rs b/crates/cuda-device/src/generated/grid_dependency.rs index 94d7848ee7..85e1c56267 100644 --- a/crates/cuda-device/src/generated/grid_dependency.rs +++ b/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/grid_sreg.rs b/crates/cuda-device/src/generated/grid_sreg.rs index ed3fa06aeb..7f8153d5e7 100644 --- a/crates/cuda-device/src/generated/grid_sreg.rs +++ b/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/i16x2.rs b/crates/cuda-device/src/generated/i16x2.rs index 054da50cd7..3b16729348 100644 --- a/crates/cuda-device/src/generated/i16x2.rs +++ b/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/int.rs b/crates/cuda-device/src/generated/int.rs index 87a798b163..4a6f80ef9a 100644 --- a/crates/cuda-device/src/generated/int.rs +++ b/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/ldmatrix.rs b/crates/cuda-device/src/generated/ldmatrix.rs index b288951019..46ca1cd191 100644 --- a/crates/cuda-device/src/generated/ldmatrix.rs +++ b/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/mbarrier_basic.rs b/crates/cuda-device/src/generated/mbarrier_basic.rs index eb377051e6..36a591b048 100644 --- a/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/mbarrier_extended.rs b/crates/cuda-device/src/generated/mbarrier_extended.rs index 7787d693cf..333f788622 100644 --- a/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/movmatrix.rs b/crates/cuda-device/src/generated/movmatrix.rs index 09ccf2ea4d..b6b8b3f30e 100644 --- a/crates/cuda-device/src/generated/movmatrix.rs +++ b/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/prmt.rs b/crates/cuda-device/src/generated/prmt.rs index 30fb07902d..7164d9b9d7 100644 --- a/crates/cuda-device/src/generated/prmt.rs +++ b/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/crates/cuda-device/src/generated/register_control.rs b/crates/cuda-device/src/generated/register_control.rs index 91d3e7f239..772e86460f 100644 --- a/crates/cuda-device/src/generated/register_control.rs +++ b/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/register_mma.rs b/crates/cuda-device/src/generated/register_mma.rs index f623acbe07..3bcd6f1c6e 100644 --- a/crates/cuda-device/src/generated/register_mma.rs +++ b/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/shared_sreg.rs b/crates/cuda-device/src/generated/shared_sreg.rs index 35198aaa6f..20c97c3167 100644 --- a/crates/cuda-device/src/generated/shared_sreg.rs +++ b/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/sparse_mma.rs b/crates/cuda-device/src/generated/sparse_mma.rs index a1e8bd0f63..9c94fdc74f 100644 --- a/crates/cuda-device/src/generated/sparse_mma.rs +++ b/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/sreg.rs b/crates/cuda-device/src/generated/sreg.rs index f972a0a046..f77c0b5921 100644 --- a/crates/cuda-device/src/generated/sreg.rs +++ b/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/crates/cuda-device/src/generated/stmatrix.rs b/crates/cuda-device/src/generated/stmatrix.rs index 73f22a70d0..a1f7f1a721 100644 --- a/crates/cuda-device/src/generated/stmatrix.rs +++ b/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/tcgen05.rs b/crates/cuda-device/src/generated/tcgen05.rs index cb1712552e..6673d2d64e 100644 --- a/crates/cuda-device/src/generated/tcgen05.rs +++ b/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/tma.rs b/crates/cuda-device/src/generated/tma.rs index 36156f0b8d..036dab6abe 100644 --- a/crates/cuda-device/src/generated/tma.rs +++ b/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tma` to keep its public API stable. @@ -233,6 +233,23 @@ pub unsafe fn cp_async_bulk_tensor_1d_g2s( unreachable!("cp_async_bulk_tensor_1d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_1d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, barrier); + unreachable!("cp_async_bulk_tensor_1d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -263,6 +280,24 @@ pub unsafe fn cp_async_bulk_tensor_2d_g2s( unreachable!("cp_async_bulk_tensor_2d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_2d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, barrier); + unreachable!("cp_async_bulk_tensor_2d_g2s_cta called outside CUDA kernel context") +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// # Safety @@ -329,6 +364,25 @@ pub unsafe fn cp_async_bulk_tensor_3d_g2s( unreachable!("cp_async_bulk_tensor_3d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_3d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, barrier); + unreachable!("cp_async_bulk_tensor_3d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -363,6 +417,26 @@ pub unsafe fn cp_async_bulk_tensor_4d_g2s( unreachable!("cp_async_bulk_tensor_4d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_4d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, coord3, barrier); + unreachable!("cp_async_bulk_tensor_4d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -402,6 +476,30 @@ pub unsafe fn cp_async_bulk_tensor_5d_g2s( unreachable!("cp_async_bulk_tensor_5d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +#[allow(clippy::too_many_arguments)] +pub unsafe fn cp_async_bulk_tensor_5d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + coord4: i32, + barrier: *mut Barrier, +) { + let _ = ( + dst, tensor_map, coord0, coord1, coord2, coord3, coord4, barrier, + ); + unreachable!("cp_async_bulk_tensor_5d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety diff --git a/crates/cuda-device/src/generated/warp_sreg.rs b/crates/cuda-device/src/generated/warp_sreg.rs index 8019adef7b..e5b2ce9885 100644 --- a/crates/cuda-device/src/generated/warp_sreg.rs +++ b/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/wgmma_control.rs b/crates/cuda-device/src/generated/wgmma_control.rs index fd1ea8426e..6973c7cf97 100644 --- a/crates/cuda-device/src/generated/wgmma_control.rs +++ b/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/crates/cuda-device/src/tma.rs b/crates/cuda-device/src/tma.rs index 3532cab7e4..fdbc2fb0be 100644 --- a/crates/cuda-device/src/tma.rs +++ b/crates/cuda-device/src/tma.rs @@ -113,6 +113,7 @@ //! - **sm_120 (Blackwell)**: Latest TMA capabilities use crate::barrier::Barrier; +use crate::shared::SharedPtr; // ============================================================================= // TMA Descriptor Type diff --git a/crates/cuda-intrinsics-gen/src/model/contracts/tma.rs b/crates/cuda-intrinsics-gen/src/model/contracts/tma.rs index d4690333ec..5270e609c7 100644 --- a/crates/cuda-intrinsics-gen/src/model/contracts/tma.rs +++ b/crates/cuda-intrinsics-gen/src/model/contracts/tma.rs @@ -30,12 +30,17 @@ impl Tma { #[serde(rename_all = "snake_case")] pub enum TmaOperation { G2sTile1d, + G2sCtaTile1d, G2sTile2d, + G2sCtaTile2d, G2sTile2dMulticast, G2sTile2dMulticastCg2, G2sTile3d, + G2sCtaTile3d, G2sTile4d, + G2sCtaTile4d, G2sTile5d, + G2sCtaTile5d, S2gTile1d, S2gTile2d, S2gTile3d, @@ -82,16 +87,28 @@ pub enum TmaOperation { } impl TmaOperation { + pub const fn is_cta_g2s(self) -> bool { + matches!( + self, + Self::G2sCtaTile1d + | Self::G2sCtaTile2d + | Self::G2sCtaTile3d + | Self::G2sCtaTile4d + | Self::G2sCtaTile5d + ) + } + pub const fn dimensions(self) -> Option { match self { - Self::G2sTile1d | Self::S2gTile1d => Some(1), + Self::G2sTile1d | Self::G2sCtaTile1d | Self::S2gTile1d => Some(1), Self::G2sTile2d + | Self::G2sCtaTile2d | Self::G2sTile2dMulticast | Self::G2sTile2dMulticastCg2 | Self::S2gTile2d => Some(2), - Self::G2sTile3d | Self::S2gTile3d => Some(3), - Self::G2sTile4d | Self::S2gTile4d => Some(4), - Self::G2sTile5d | Self::S2gTile5d => Some(5), + Self::G2sTile3d | Self::G2sCtaTile3d | Self::S2gTile3d => Some(3), + Self::G2sTile4d | Self::G2sCtaTile4d | Self::S2gTile4d => Some(4), + Self::G2sTile5d | Self::G2sCtaTile5d | Self::S2gTile5d => Some(5), Self::Reduce | Self::CommitGroup | Self::WaitGroup diff --git a/crates/cuda-intrinsics-gen/src/model/overlay.rs b/crates/cuda-intrinsics-gen/src/model/overlay.rs index 081f4b33b7..fbe3bf8c3d 100644 --- a/crates/cuda-intrinsics-gen/src/model/overlay.rs +++ b/crates/cuda-intrinsics-gen/src/model/overlay.rs @@ -322,6 +322,8 @@ pub struct ClcAdmissionVariant { pub struct TmaAdmission { pub llvm_evidence_profile: String, pub libnvvm_evidence_profile: String, + pub cta_llvm_evidence_profile: String, + pub cta_libnvvm_evidence_profile: String, #[serde(default)] pub reduce_llvm_evidence_profile: Option, #[serde(default)] diff --git a/crates/cuda-intrinsics-gen/src/render/compat.rs b/crates/cuda-intrinsics-gen/src/render/compat.rs index ecc48feca7..82e8f673cd 100644 --- a/crates/cuda-intrinsics-gen/src/render/compat.rs +++ b/crates/cuda-intrinsics-gen/src/render/compat.rs @@ -913,12 +913,17 @@ pub(super) fn render_compat_tma(catalog: &CatalogFile, hash: &str) -> String { let is_g2s = matches!( operation, TmaOperation::G2sTile1d + | TmaOperation::G2sCtaTile1d | TmaOperation::G2sTile2d + | TmaOperation::G2sCtaTile2d | TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 | TmaOperation::G2sTile3d + | TmaOperation::G2sCtaTile3d | TmaOperation::G2sTile4d + | TmaOperation::G2sCtaTile4d | TmaOperation::G2sTile5d + | TmaOperation::G2sCtaTile5d ); let is_s2g = matches!( operation, @@ -935,6 +940,11 @@ pub(super) fn render_compat_tma(catalog: &CatalogFile, hash: &str) -> String { output.push_str( "/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes.\n", ); + if operation.is_cta_g2s() { + output.push_str( + "/// `barrier` must point into the issuing CTA's shared memory, as for the other\n/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type.\n", + ); + } } else if is_s2g { output.push_str( "/// `src` and `tensor_map` must remain valid until the committed copy group completes.\n", @@ -961,8 +971,15 @@ pub(super) fn render_compat_tma(catalog: &CatalogFile, hash: &str) -> String { output.push_str("#[inline(never)]\n"); if is_g2s { let dimensions = dimensions.unwrap(); + // The CTA-local forms take a typed CTA-shared destination; the + // raw ABI keeps `*mut u8`, as `*mut Barrier` keeps `*mut u64`. + let destination = if operation.is_cta_g2s() { + "dst: SharedPtr" + } else { + "dst: *mut u8" + }; let mut arguments = vec![ - "dst: *mut u8".to_owned(), + destination.to_owned(), "tensor_map: *const TmaDescriptor".to_owned(), ]; let mut values = vec!["dst".to_owned(), "tensor_map".to_owned()]; diff --git a/crates/cuda-intrinsics-gen/src/render/importer.rs b/crates/cuda-intrinsics-gen/src/render/importer.rs index 78478f8c59..ee796b9e19 100644 --- a/crates/cuda-intrinsics-gen/src/render/importer.rs +++ b/crates/cuda-intrinsics-gen/src/render/importer.rs @@ -2402,6 +2402,18 @@ fn tma_arms(catalog: &CatalogFile) -> String { ) .unwrap(); } + TmaOperation::G2sCtaTile1d + | TmaOperation::G2sCtaTile2d + | TmaOperation::G2sCtaTile3d + | TmaOperation::G2sCtaTile4d + | TmaOperation::G2sCtaTile5d => { + writeln!( + output, + " Ok(Some(super::super::tma::emit_tma_g2s_cta(\n ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, {}, {marker:?},\n )?))", + operation.dimensions().unwrap() + ) + .unwrap(); + } TmaOperation::G2sTile2dMulticast => { writeln!( output, diff --git a/crates/cuda-intrinsics-gen/src/render/lowering.rs b/crates/cuda-intrinsics-gen/src/render/lowering.rs index 36b73243f1..d44f8f42e2 100644 --- a/crates/cuda-intrinsics-gen/src/render/lowering.rs +++ b/crates/cuda-intrinsics-gen/src/render/lowering.rs @@ -1519,6 +1519,18 @@ fn tma_impls(catalog: &CatalogFile) -> String { ) .unwrap(); } + TmaOperation::G2sCtaTile1d + | TmaOperation::G2sCtaTile2d + | TmaOperation::G2sCtaTile3d + | TmaOperation::G2sCtaTile4d + | TmaOperation::G2sCtaTile5d => { + writeln!( + output, + " convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, {})", + operation.dimensions().unwrap() + ) + .unwrap(); + } TmaOperation::G2sTile2dMulticast => output.push_str( " convert_g2s(ctx, rewriter, self.get_operation(), operands_info, 2, true)\n", ), @@ -2099,6 +2111,7 @@ const LOWERING_INTRINSIC_HELPERS: &[(&str, &str)] = &[ ("ReduceConfig", "tma::ReduceConfig"), ("convert_control", "tma::convert_control"), ("convert_g2s", "tma::convert_g2s"), + ("convert_g2s_cta", "tma::convert_g2s_cta"), ( "convert_g2s_multicast_cg2", "tma::convert_g2s_multicast_cg2", diff --git a/crates/cuda-intrinsics-gen/src/render/probes.rs b/crates/cuda-intrinsics-gen/src/render/probes.rs index 3c7858866c..980c034896 100644 --- a/crates/cuda-intrinsics-gen/src/render/probes.rs +++ b/crates/cuda-intrinsics-gen/src/render/probes.rs @@ -84,12 +84,17 @@ fn render_tma_probe(catalog: &CatalogFile, record: &CatalogIntrinsic, hash: &str let is_g2s = matches!( operation, TmaOperation::G2sTile1d + | TmaOperation::G2sCtaTile1d | TmaOperation::G2sTile2d + | TmaOperation::G2sCtaTile2d | TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 | TmaOperation::G2sTile3d + | TmaOperation::G2sCtaTile3d | TmaOperation::G2sTile4d + | TmaOperation::G2sCtaTile4d | TmaOperation::G2sTile5d + | TmaOperation::G2sCtaTile5d ); let coordinates = std::iter::repeat_n("i32", dimensions) .collect::>() @@ -103,6 +108,14 @@ fn render_tma_probe(catalog: &CatalogFile, record: &CatalogIntrinsic, hash: &str .collect::>() .join(", "); if is_g2s { + let cta = matches!( + operation, + TmaOperation::G2sCtaTile1d + | TmaOperation::G2sCtaTile2d + | TmaOperation::G2sCtaTile3d + | TmaOperation::G2sCtaTile4d + | TmaOperation::G2sCtaTile5d + ); let multicast = matches!( operation, TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 @@ -112,11 +125,19 @@ fn render_tma_probe(catalog: &CatalogFile, record: &CatalogIntrinsic, hash: &str } else { format!(", {coordinates}") }; - writeln!( - output, - "declare void @{symbol}(ptr addrspace(7), ptr addrspace(3), ptr{declaration_coordinates}, i16, i64, i1, i1, i32) #0\n" - ) - .unwrap(); + if cta { + writeln!( + output, + "declare void @{symbol}(ptr addrspace(3), ptr addrspace(3), ptr{declaration_coordinates}, i64, i1) #0\n" + ) + .unwrap(); + } else { + writeln!( + output, + "declare void @{symbol}(ptr addrspace(7), ptr addrspace(3), ptr{declaration_coordinates}, i16, i64, i1, i1, i32) #0\n" + ) + .unwrap(); + } let parameters = if coordinate_parameters.is_empty() { String::new() } else { @@ -129,9 +150,15 @@ fn render_tma_probe(catalog: &CatalogFile, record: &CatalogIntrinsic, hash: &str record.id ) .unwrap(); - output.push_str( - " %dst = addrspacecast ptr %dst_generic to ptr addrspace(7)\n %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3)\n", - ); + if cta { + output.push_str( + " %dst = addrspacecast ptr %dst_generic to ptr addrspace(3)\n %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3)\n", + ); + } else { + output.push_str( + " %dst = addrspacecast ptr %dst_generic to ptr addrspace(7)\n %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3)\n", + ); + } let arguments = if coordinate_arguments.is_empty() { String::new() } else { @@ -143,11 +170,19 @@ fn render_tma_probe(catalog: &CatalogFile, record: &CatalogIntrinsic, hash: &str } else { 0 }; - writeln!( - output, - " call void @{symbol}(ptr addrspace(7) %dst, ptr addrspace(3) %barrier, ptr %tensor_map{arguments}, i16 {mask}, i64 0, i1 {multicast}, i1 false, i32 {group}) #0" - ) - .unwrap(); + if cta { + writeln!( + output, + " call void @{symbol}(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map{arguments}, i64 0, i1 false) #0" + ) + .unwrap(); + } else { + writeln!( + output, + " call void @{symbol}(ptr addrspace(7) %dst, ptr addrspace(3) %barrier, ptr %tensor_map{arguments}, i16 {mask}, i64 0, i1 {multicast}, i1 false, i32 {group}) #0" + ) + .unwrap(); + } } else { let declaration_coordinates = if coordinates.is_empty() { String::new() diff --git a/crates/cuda-intrinsics-gen/src/render/tests/mma.rs b/crates/cuda-intrinsics-gen/src/render/tests/mma.rs index 67123ff4e6..0fa3bc9d96 100644 --- a/crates/cuda-intrinsics-gen/src/render/tests/mma.rs +++ b/crates/cuda-intrinsics-gen/src/render/tests/mma.rs @@ -345,7 +345,7 @@ fn register_mma_rendering_preserves_apis_order_convergence_and_variants() { let repo_root = Path::new(env!("CARGO_MANIFEST_DIR")).join("../.."); let catalog = crate::resolve::resolve(&repo_root).unwrap(); validate_renderable(&catalog).unwrap(); - assert_eq!(catalog.intrinsics.len(), 1029); + assert_eq!(catalog.intrinsics.len(), 1034); let records: Vec<_> = register_mmas(&catalog).collect(); assert_eq!(records.len(), 154); let generated_records = records diff --git a/crates/cuda-intrinsics-gen/src/render/tests/mod.rs b/crates/cuda-intrinsics-gen/src/render/tests/mod.rs index 729db95fa8..9afc2769f3 100644 --- a/crates/cuda-intrinsics-gen/src/render/tests/mod.rs +++ b/crates/cuda-intrinsics-gen/src/render/tests/mod.rs @@ -42,7 +42,7 @@ fn catalog_with_clc() -> CatalogFile { fn catalog_with_tma() -> CatalogFile { let repo_root = Path::new(env!("CARGO_MANIFEST_DIR")).join("../.."); let catalog = crate::resolve::test_catalog_with_tma(&repo_root).unwrap(); - assert_eq!(tma_intrinsics(&catalog).count(), 111); + assert_eq!(tma_intrinsics(&catalog).count(), 116); catalog } diff --git a/crates/cuda-intrinsics-gen/src/render/tests/tensor.rs b/crates/cuda-intrinsics-gen/src/render/tests/tensor.rs index d431282c51..dec130798d 100644 --- a/crates/cuda-intrinsics-gen/src/render/tests/tensor.rs +++ b/crates/cuda-intrinsics-gen/src/render/tests/tensor.rs @@ -419,7 +419,13 @@ fn tma_rendering_preserves_api_and_injects_backend_defaults() { assert!(compatibility.contains( "pub unsafe fn cp_async_bulk_tensor_reduce_xor_im2col_5d(src: *const u8, tensor_map: *const TmaDescriptor, coord0: i32, coord1: i32, coord2: i32, coord3: i32, coord4: i32)" )); + assert!(compatibility.contains( + "pub unsafe fn cp_async_bulk_tensor_2d_g2s_cta(dst: SharedPtr, tensor_map: *const TmaDescriptor, coord0: i32, coord1: i32, barrier: *mut Barrier)" + )); for dimensions in 1..=5 { + assert!(compatibility.contains(&format!( + "pub unsafe fn cp_async_bulk_tensor_{dimensions}d_g2s_cta(dst: SharedPtr," + ))); assert!(compatibility.contains(&format!( "pub unsafe fn cp_async_bulk_prefetch_tensor_{dimensions}d_l2(" ))); @@ -437,8 +443,8 @@ fn tma_rendering_preserves_api_and_injects_backend_defaults() { assert!(compatibility.contains("pub fn fence_proxy_tensormap_generic_release_system()")); let dialect = render_dialect_tma(&catalog, "test-hash"); - assert_eq!(dialect.matches("pub struct ").count(), 111); - assert_eq!(dialect.matches("NResultsInterface<0>").count(), 111); + assert_eq!(dialect.matches("pub struct ").count(), 116); + assert_eq!(dialect.matches("NResultsInterface<0>").count(), 116); assert!(dialect.contains("NOpdsInterface<10>")); assert!(dialect.contains("CpAsyncBulkWaitGroupReadOp::register(ctx)")); assert!(dialect.contains("CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op::register(ctx)")); diff --git a/crates/cuda-intrinsics-gen/src/render/validate.rs b/crates/cuda-intrinsics-gen/src/render/validate.rs index f78c55a7b2..defa726c97 100644 --- a/crates/cuda-intrinsics-gen/src/render/validate.rs +++ b/crates/cuda-intrinsics-gen/src/render/validate.rs @@ -733,10 +733,15 @@ pub(super) fn validate_renderable(catalog: &CatalogFile) -> Result<()> { && match (tma.operation, tma.adapter) { ( TmaOperation::G2sTile1d + | TmaOperation::G2sCtaTile1d | TmaOperation::G2sTile2d + | TmaOperation::G2sCtaTile2d | TmaOperation::G2sTile3d + | TmaOperation::G2sCtaTile3d | TmaOperation::G2sTile4d - | TmaOperation::G2sTile5d, + | TmaOperation::G2sCtaTile4d + | TmaOperation::G2sTile5d + | TmaOperation::G2sCtaTile5d, TmaAdapter::G2sPointersCoordinatesBarrierInjectDefaults, ) => { !record.rust.safe diff --git a/crates/cuda-intrinsics-gen/src/resolve/driver.rs b/crates/cuda-intrinsics-gen/src/resolve/driver.rs index 41a224c375..0273760be0 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/driver.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/driver.rs @@ -243,6 +243,8 @@ pub(crate) fn test_catalog_with_tma(repo_root: &Path) -> Result { let admission = TmaAdmission { llvm_evidence_profile: "llvm-tma-test".into(), libnvvm_evidence_profile: "libnvvm-tma-test".into(), + cta_llvm_evidence_profile: "llvm-tma-test".into(), + cta_libnvvm_evidence_profile: "libnvvm-tma-test".into(), reduce_llvm_evidence_profile: Some("llvm-tma-reduce-test".into()), reduce_libnvvm_evidence_profile: Some("libnvvm-tma-reduce-test".into()), runtime_validation: RuntimeValidation::Unexecuted, diff --git a/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs b/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs index 99c0ca90a0..95acbb28fd 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs @@ -19,14 +19,19 @@ use crate::resolve::guards::*; pub(in crate::resolve) const TMA_BLACKWELL_TARGETS: &str = "sm_100a|sm_101a|sm_103a|sm_110a"; pub(in crate::resolve) const TENSOR_MAP_REPLACE_TARGETS: &str = "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f|sm_90a"; -pub(in crate::resolve) const TMA_OPERATIONS: [TmaOperation; 47] = [ +pub(in crate::resolve) const TMA_OPERATIONS: [TmaOperation; 52] = [ TmaOperation::G2sTile1d, + TmaOperation::G2sCtaTile1d, TmaOperation::G2sTile2d, + TmaOperation::G2sCtaTile2d, TmaOperation::G2sTile2dMulticast, TmaOperation::G2sTile2dMulticastCg2, TmaOperation::G2sTile3d, + TmaOperation::G2sCtaTile3d, TmaOperation::G2sTile4d, + TmaOperation::G2sCtaTile4d, TmaOperation::G2sTile5d, + TmaOperation::G2sCtaTile5d, TmaOperation::S2gTile1d, TmaOperation::S2gTile2d, TmaOperation::S2gTile3d, @@ -409,6 +414,15 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "CpAsyncBulkTensorG2sTile1dOp", "nvvm.cp_async_bulk_tensor_g2s_tile_1d", ), + TmaOperation::G2sCtaTile1d => ( + "i1030", + "cp_async_bulk_tensor_1d_g2s_cta", + "memory.copy.async.bulk.tensor.g2s.cta.tile.1d", + "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", + "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "CpAsyncBulkTensorG2sCtaTile1dOp", + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + ), TmaOperation::G2sTile2d => ( "i0329", "cp_async_bulk_tensor_2d_g2s", @@ -418,6 +432,15 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "CpAsyncBulkTensorG2sTile2dOp", "nvvm.cp_async_bulk_tensor_g2s_tile_2d", ), + TmaOperation::G2sCtaTile2d => ( + "i1031", + "cp_async_bulk_tensor_2d_g2s_cta", + "memory.copy.async.bulk.tensor.g2s.cta.tile.2d", + "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", + "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "CpAsyncBulkTensorG2sCtaTile2dOp", + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + ), TmaOperation::G2sTile2dMulticast => ( "i0330", "cp_async_bulk_tensor_2d_g2s_multicast", @@ -445,6 +468,15 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "CpAsyncBulkTensorG2sTile3dOp", "nvvm.cp_async_bulk_tensor_g2s_tile_3d", ), + TmaOperation::G2sCtaTile3d => ( + "i1032", + "cp_async_bulk_tensor_3d_g2s_cta", + "memory.copy.async.bulk.tensor.g2s.cta.tile.3d", + "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", + "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", + "CpAsyncBulkTensorG2sCtaTile3dOp", + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + ), TmaOperation::G2sTile4d => ( "i0333", "cp_async_bulk_tensor_4d_g2s", @@ -454,6 +486,15 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "CpAsyncBulkTensorG2sTile4dOp", "nvvm.cp_async_bulk_tensor_g2s_tile_4d", ), + TmaOperation::G2sCtaTile4d => ( + "i1033", + "cp_async_bulk_tensor_4d_g2s_cta", + "memory.copy.async.bulk.tensor.g2s.cta.tile.4d", + "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", + "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", + "CpAsyncBulkTensorG2sCtaTile4dOp", + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + ), TmaOperation::G2sTile5d => ( "i0334", "cp_async_bulk_tensor_5d_g2s", @@ -463,6 +504,15 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "CpAsyncBulkTensorG2sTile5dOp", "nvvm.cp_async_bulk_tensor_g2s_tile_5d", ), + TmaOperation::G2sCtaTile5d => ( + "i1034", + "cp_async_bulk_tensor_5d_g2s_cta", + "memory.copy.async.bulk.tensor.g2s.cta.tile.5d", + "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", + "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", + "CpAsyncBulkTensorG2sCtaTile5dOp", + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + ), TmaOperation::S2gTile1d => ( "i0335", "cp_async_bulk_tensor_1d_s2g", @@ -829,12 +879,17 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { let is_g2s = matches!( operation, TmaOperation::G2sTile1d + | TmaOperation::G2sCtaTile1d | TmaOperation::G2sTile2d + | TmaOperation::G2sCtaTile2d | TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 | TmaOperation::G2sTile3d + | TmaOperation::G2sCtaTile3d | TmaOperation::G2sTile4d + | TmaOperation::G2sCtaTile4d | TmaOperation::G2sTile5d + | TmaOperation::G2sCtaTile5d ); let is_s2g = matches!( operation, @@ -848,6 +903,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { operation, TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 ); + let cta_g2s = operation.is_cta_g2s(); let cg2 = operation == TmaOperation::G2sTile2dMulticastCg2; let prefetch_coordinates = operation.prefetch_coordinate_count(); let is_release_fence = matches!( @@ -877,10 +933,19 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { } dialect_operands.extend(["ptr", "ptr", "ptr"]); dialect_operands.extend(std::iter::repeat_n("i32", dimensions.unwrap())); - dialect_operands.extend(["i16", "i64"]); - llvm_arguments.extend(["shared_cluster_ptr", "shared_ptr", "ptr"]); - llvm_arguments.extend(std::iter::repeat_n("i32", dimensions.unwrap())); - llvm_arguments.extend(["i16", "i64", "i1", "i1", "i32"]); + if !cta_g2s { + dialect_operands.push("i16"); + } + dialect_operands.push("i64"); + if cta_g2s { + llvm_arguments.extend(["shared_ptr", "shared_ptr", "ptr"]); + llvm_arguments.extend(std::iter::repeat_n("i32", dimensions.unwrap())); + llvm_arguments.extend(["i64", "i1"]); + } else { + llvm_arguments.extend(["shared_cluster_ptr", "shared_ptr", "ptr"]); + llvm_arguments.extend(std::iter::repeat_n("i32", dimensions.unwrap())); + llvm_arguments.extend(["i16", "i64", "i1", "i1", "i32"]); + } ( if multicast { TmaAdapter::G2sPointersCoordinatesBarrierMaskInjectDefaults @@ -892,6 +957,8 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { true, if multicast { "Starts a multicast TMA tile copy from global to cluster shared memory." + } else if cta_g2s { + "Starts a TMA tile copy from global to the issuing CTA's shared memory." } else { "Starts a TMA tile copy from global to cluster shared memory." }, @@ -1094,7 +1161,11 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "bulk".into(), "tensor".into(), format!("{}d", dimensions.unwrap()), - "shared::cluster".into(), + if cta_g2s { + "shared::cta".into() + } else { + "shared::cluster".into() + }, "global".into(), "tile".into(), "mbarrier::complete_tx::bytes".into(), @@ -1308,13 +1379,14 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { unreachable!("TMA operation category was matched") }; - let minimum_ptx = if blackwell_tma || operation == TmaOperation::ReplaceSwizzleAtomicity { - "8.6" - } else if is_replace || is_fence { - "8.3" - } else { - "8.0" - }; + let minimum_ptx = + if blackwell_tma || cta_g2s || operation == TmaOperation::ReplaceSwizzleAtomicity { + "8.6" + } else if is_replace || is_fence { + "8.3" + } else { + "8.0" + }; let (minimum_sm, targets) = if blackwell_tma { (None, TMA_BLACKWELL_TARGETS) } else if operation == TmaOperation::ReplaceSwizzleAtomicity { @@ -1386,7 +1458,9 @@ pub(in crate::resolve) fn expand_tma_admission( ); ensure!( !admission.llvm_evidence_profile.trim().is_empty() - && !admission.libnvvm_evidence_profile.trim().is_empty(), + && !admission.libnvvm_evidence_profile.trim().is_empty() + && !admission.cta_llvm_evidence_profile.trim().is_empty() + && !admission.cta_libnvvm_evidence_profile.trim().is_empty(), "compact TMA admission requires both backend evidence profiles" ); ensure!( @@ -1410,6 +1484,16 @@ pub(in crate::resolve) fn expand_tma_admission( recipe.id, recipe.abi_id ); + let llvm_evidence_profile = if recipe.operation.is_cta_g2s() { + &admission.cta_llvm_evidence_profile + } else { + &admission.llvm_evidence_profile + }; + let libnvvm_evidence_profile = if recipe.operation.is_cta_g2s() { + &admission.cta_libnvvm_evidence_profile + } else { + &admission.libnvvm_evidence_profile + }; Ok(OverlayIntrinsic { id: recipe.id.into(), abi_id: variant.abi_id.clone(), @@ -1462,7 +1546,7 @@ pub(in crate::resolve) fn expand_tma_admission( OverlayBackendLowering { backend: IntrinsicBackend::LlvmNvptx, mechanism: recipe.llvm_mechanism, - evidence_profile: admission.llvm_evidence_profile.clone(), + evidence_profile: llvm_evidence_profile.clone(), targets: None, minimum_ptx: Some(recipe.minimum_ptx.into()), minimum_sm: recipe.minimum_sm.map(Into::into), @@ -1470,7 +1554,7 @@ pub(in crate::resolve) fn expand_tma_admission( OverlayBackendLowering { backend: IntrinsicBackend::LibNvvm, mechanism: BackendLoweringMechanism::InlinePtx, - evidence_profile: admission.libnvvm_evidence_profile.clone(), + evidence_profile: libnvvm_evidence_profile.clone(), targets: None, minimum_ptx: Some(recipe.minimum_ptx.into()), minimum_sm: recipe.minimum_sm.map(Into::into), @@ -1781,6 +1865,14 @@ pub(in crate::resolve) fn validate_tma_reduction_policy( pub(in crate::resolve) fn tma_imported_properties(operation: TmaOperation) -> Vec { assert_ne!(operation, TmaOperation::Reduce); let dimensions = operation.dimensions(); + if operation.is_cta_g2s() { + return vec![ + format!("ImmArg", dimensions.unwrap() + 4), + "IntrConvergent".into(), + "ReadOnly".into(), + "WriteOnly".into(), + ]; + } if matches!( operation, TmaOperation::G2sTile1d diff --git a/crates/cuda-intrinsics-gen/src/resolve/guards.rs b/crates/cuda-intrinsics-gen/src/resolve/guards.rs index 4575d19c80..ac42753c76 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/guards.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/guards.rs @@ -307,10 +307,18 @@ pub(super) fn selection_matches_tma_policy( "TMA_G2S_TILE_CG0_1D", "cp.async.bulk.tensor.1d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0}}], [$mbar];", ), + TmaOperation::G2sCtaTile1d => ( + "TMA_G2S_CTA_TILE_1D", + "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + ), TmaOperation::G2sTile2d => ( "TMA_G2S_TILE_CG0_2D", "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", ), + TmaOperation::G2sCtaTile2d => ( + "TMA_G2S_CTA_TILE_2D", + "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + ), TmaOperation::G2sTile2dMulticast => ( "TMA_G2S_TILE_CG0_2D_MC", "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", @@ -323,14 +331,26 @@ pub(super) fn selection_matches_tma_policy( "TMA_G2S_TILE_CG0_3D", "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", ), + TmaOperation::G2sCtaTile3d => ( + "TMA_G2S_CTA_TILE_3D", + "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + ), TmaOperation::G2sTile4d => ( "TMA_G2S_TILE_CG0_4D", "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", ), + TmaOperation::G2sCtaTile4d => ( + "TMA_G2S_CTA_TILE_4D", + "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + ), TmaOperation::G2sTile5d => ( "TMA_G2S_TILE_CG0_5D", "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", ), + TmaOperation::G2sCtaTile5d => ( + "TMA_G2S_CTA_TILE_5D", + "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + ), TmaOperation::S2gTile1d => ( "TMA_TENSOR_S2G_TILE_1D", "cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [$tmap, {{$d0}}], [$src];", @@ -456,29 +476,48 @@ pub(super) fn selection_matches_tma_policy( if matches!( operation, TmaOperation::G2sTile1d + | TmaOperation::G2sCtaTile1d | TmaOperation::G2sTile2d + | TmaOperation::G2sCtaTile2d | TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 | TmaOperation::G2sTile3d + | TmaOperation::G2sCtaTile3d | TmaOperation::G2sTile4d + | TmaOperation::G2sCtaTile4d | TmaOperation::G2sTile5d + | TmaOperation::G2sCtaTile5d ) { let dimensions = operation.dimensions().unwrap(); - immediate_bindings.push(crate::model::ImportedImmediateBinding { - argument_index: dimensions + 5, - value: if matches!( - operation, - TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 - ) { - -1 - } else { - 0 - }, - }); - immediate_bindings.push(crate::model::ImportedImmediateBinding { - argument_index: dimensions + 6, - value: 0, - }); + if matches!( + operation, + TmaOperation::G2sCtaTile1d + | TmaOperation::G2sCtaTile2d + | TmaOperation::G2sCtaTile3d + | TmaOperation::G2sCtaTile4d + | TmaOperation::G2sCtaTile5d + ) { + immediate_bindings.push(crate::model::ImportedImmediateBinding { + argument_index: dimensions + 4, + value: 0, + }); + } else { + immediate_bindings.push(crate::model::ImportedImmediateBinding { + argument_index: dimensions + 5, + value: if matches!( + operation, + TmaOperation::G2sTile2dMulticast | TmaOperation::G2sTile2dMulticastCg2 + ) { + -1 + } else { + 0 + }, + }); + immediate_bindings.push(crate::model::ImportedImmediateBinding { + argument_index: dimensions + 6, + value: 0, + }); + } } else if matches!( operation, TmaOperation::S2gTile1d diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs index 082bfbf15e..a50dac5bf1 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs @@ -1520,6 +1520,8 @@ pub(super) fn test_tma_admission() -> TmaAdmission { TmaAdmission { llvm_evidence_profile: "llvm-tma-test".into(), libnvvm_evidence_profile: "libnvvm-tma-test".into(), + cta_llvm_evidence_profile: "llvm-tma-test".into(), + cta_libnvvm_evidence_profile: "libnvvm-tma-test".into(), reduce_llvm_evidence_profile: Some("llvm-tma-reduce-test".into()), reduce_libnvvm_evidence_profile: Some("libnvvm-tma-reduce-test".into()), runtime_validation: RuntimeValidation::Unexecuted, diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs index 44707537ae..c09b698ac8 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs @@ -49,7 +49,7 @@ fn overlay_manifest_loads_sorted_family_shards() { read_overlay(&repo_root, &repo_root.join("intrinsics/overlay.toml")).unwrap(); assert_eq!(overlay.schema, OVERLAY_SCHEMA); assert_eq!(overlay.shards.len(), 66); - assert_eq!(overlay.intrinsics.len(), 1029); + assert_eq!(overlay.intrinsics.len(), 1034); assert_eq!( overlay .intrinsics diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs index ceb416cb46..651d3962bb 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs @@ -190,12 +190,17 @@ fn compact_tma_admission_matches_llvm_and_fails_closed() { .collect::>(), [ ("i0328", "cp_async_bulk_tensor_1d_g2s"), + ("i1030", "cp_async_bulk_tensor_1d_g2s_cta"), ("i0329", "cp_async_bulk_tensor_2d_g2s"), + ("i1031", "cp_async_bulk_tensor_2d_g2s_cta"), ("i0330", "cp_async_bulk_tensor_2d_g2s_multicast"), ("i0331", "cp_async_bulk_tensor_2d_g2s_multicast_cg2"), ("i0332", "cp_async_bulk_tensor_3d_g2s"), + ("i1032", "cp_async_bulk_tensor_3d_g2s_cta"), ("i0333", "cp_async_bulk_tensor_4d_g2s"), + ("i1033", "cp_async_bulk_tensor_4d_g2s_cta"), ("i0334", "cp_async_bulk_tensor_5d_g2s"), + ("i1034", "cp_async_bulk_tensor_5d_g2s_cta"), ("i0335", "cp_async_bulk_tensor_1d_s2g"), ("i0336", "cp_async_bulk_tensor_2d_s2g"), ("i0337", "cp_async_bulk_tensor_3d_s2g"), @@ -277,7 +282,7 @@ fn compact_tma_admission_matches_llvm_and_fails_closed() { } assert_eq!( - parse_hardware_target(&records[3]).unwrap(), + parse_hardware_target(&records[5]).unwrap(), CatalogHardwareTarget::AnyOf { alternatives: vec![ CatalogHardwareAlternative::ExactArchitecture { sm: 100 }, diff --git a/crates/cuda-intrinsics/src/generated/abi_v1.rs b/crates/cuda-intrinsics/src/generated/abi_v1.rs index ce0289c0dd..b7bc07f2ee 100644 --- a/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Raw ABI functions recognized by cuda-oxide. @@ -1181,6 +1181,21 @@ pub unsafe fn i0328(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u6 ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_1d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must be valid shared-memory objects, and the tensor map must be a live descriptor for this dimensionality. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1030(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u64) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_1d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1217,6 +1232,27 @@ pub unsafe fn i0329( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_2d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must be valid shared-memory objects, and the tensor map must be a live descriptor for this dimensionality. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1031( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` executed outside device compilation" + ) +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// Catalog ID: `cp_async_bulk_tensor_2d_g2s_multicast`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;`. @@ -1298,6 +1334,28 @@ pub unsafe fn i0332( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_3d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must be valid shared-memory objects, and the tensor map must be a live descriptor for this dimensionality. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1032( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_3d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1336,6 +1394,29 @@ pub unsafe fn i0333( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_4d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must be valid shared-memory objects, and the tensor map must be a live descriptor for this dimensionality. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1033( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_4d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1383,6 +1464,31 @@ pub unsafe fn i0334( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_5d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must be valid shared-memory objects, and the tensor map must be a live descriptor for this dimensionality. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[allow(clippy::too_many_arguments)] +#[inline(never)] +pub unsafe fn i1034( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: i32, + _arg7: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_5d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];`. diff --git a/crates/cuda-intrinsics/src/generated/mod.rs b/crates/cuda-intrinsics/src/generated/mod.rs index 3ab1989e8d..bfdc08c4e1 100644 --- a/crates/cuda-intrinsics/src/generated/mod.rs +++ b/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. #[doc(hidden)] @@ -1094,6 +1094,11 @@ pub mod tma { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0984 as cp_async_bulk_tensor_reduce_xor_im2col_3d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0985 as cp_async_bulk_tensor_reduce_xor_im2col_4d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0986 as cp_async_bulk_tensor_reduce_xor_im2col_5d; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1030 as cp_async_bulk_tensor_1d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1031 as cp_async_bulk_tensor_2d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1032 as cp_async_bulk_tensor_3d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1033 as cp_async_bulk_tensor_4d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1034 as cp_async_bulk_tensor_5d_g2s_cta; } /// Generated `warp` intrinsic source API. @@ -1690,6 +1695,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_1d_g2s_cta_reexports_abi_i1030() { + let public: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_1d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1030; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_1d_s2g_reexports_abi_i0335() { let public: unsafe fn(*const u8, *const u8, i32) -> () = super::tma::cp_async_bulk_tensor_1d_s2g; @@ -1706,6 +1719,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_2d_g2s_cta_reexports_abi_i1031() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_2d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1031; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_2d_g2s_multicast_reexports_abi_i0330() { let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64, u16) -> () = super::tma::cp_async_bulk_tensor_2d_g2s_multicast; @@ -1738,6 +1759,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_3d_g2s_cta_reexports_abi_i1032() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_3d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1032; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_3d_s2g_reexports_abi_i0337() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_3d_s2g; @@ -1754,6 +1783,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_4d_g2s_cta_reexports_abi_i1033() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_4d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1033; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_4d_s2g_reexports_abi_i0338() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_4d_s2g; @@ -1770,6 +1807,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_5d_g2s_cta_reexports_abi_i1034() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_5d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1034; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_5d_s2g_reexports_abi_i0339() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_5d_s2g; diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index f0cc44758a..92c535a6ae 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index 9a62dd9586..c94b65b72d 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index aa284eb54e..7cdd4cde95 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index a3e2c44a3b..787e1b5f2a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index 8955f8cc8c..9a2476a285 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index b50a23afae..dbf216f79d 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index f73fef865f..139515cfde 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 0ada366dc9..8092f5f622 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index 52c9ffdd8d..6f9e690a6a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index 735115d8a9..4d65268fb4 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index dde0b8cefd..6bdfa514f5 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index 4f3de70be4..cfd1d794c5 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index 31cc8f6be6..820b5bd170 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index 6c9be5ac8a..25b7a6c436 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index c6aa47c031..5a85b8fc53 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index ea446d4825..0ee2d37777 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 74aa24dacd..d5998f459a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index c7e5547922..b5a58dd7b7 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index 1fd06948b5..88f46f460f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 4c4a2f8287..94782b9125 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index bdceaba496..b44da147a2 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index 2ccb0c0db1..ebb1798a67 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index 102a59e264..51b1e3938b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 66bef39204..10dd608d68 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index c0b78a50bf..79e8bb384a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 712fbc5a18..5b4b3c038e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index 8ac059d9cc..a0a516ff60 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index 556a7e788e..be8d19a661 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index c58ca91aab..1ba3cdda58 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index e667da42b1..bb6dc98790 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index f9a55b392c..24596ec499 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index ab96fe16aa..00c61f64fb 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index edd04c7507..f340ab55cc 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index fa7086dcc6..bac06a967a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index a00308f306..55b5cf317f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index 2fd40c18ed..5fadc0377c 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index 6076281fa9..2c48e81325 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index cc0926521e..12979bee80 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index 4635787a76..6dc7816224 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index 4c2a20c02a..4a198336ee 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index 123f15d932..dc8bfbadb0 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index 8081db240d..9abd913e65 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index 169baef8cf..fdc9e61fed 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 78ab2245bb..975e29f9c2 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index faf89ff5ba..03eeff8dff 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 053f68fc88..42b8720302 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index 675e1ca2cf..bec10805b8 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 5e4809b884..b1b31d0e33 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index b25414b0a5..bed1b2de10 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index f34109e6e6..3ef0d3d663 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests, grouped per family shard. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index 66b12afa5e..a6f5aeb91b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 75ff1a863c..48abe141bd 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index d8ceed3973..cb1795d511 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index 6d7fad75c4..55173d97ab 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index 214502fbda..c2b6eff263 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index c4857e8d56..01d0ab6567 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 48b805f004..8dc943da7d 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index 5216dc659d..ca4700d197 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index ca62c3c7de..6e5b1b1751 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index 98af462f88..195159fef8 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index 570b8b0232..e0a207a091 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index bad07267a2..d61792cbad 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index b800b7b44c..f7e1543d97 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index aa3027531b..f2051e9e09 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index fc5ddff329..ec34094eb5 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index 3dbb77ebd1..2a3dae516e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index b01dcd7de4..f9d0d6db1e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 90092d1fb8..9c6fe1ca6c 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index 589e374de8..6ba784374b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index e2ea3a9ddc..eee0567353 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index a324705f5e..8e8349b121 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index cf0d62e8d4..b71acfaecd 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tma` intrinsics. @@ -6643,3 +6643,348 @@ fn cp_async_bulk_tensor_reduce_xor_im2col_5d_target_matches_the_catalog() { ] as &[&str] ); } + +#[test] +fn cp_async_bulk_tensor_1d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1030").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_1d_g2s_cta"); + assert_eq!(target.abi_id, "i1030"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_2d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1031").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_2d_g2s_cta"); + assert_eq!(target.abi_id, "i1031"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_3d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1032").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_3d_g2s_cta"); + assert_eq!(target.abi_id, "i1032"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_4d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1033").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_4d_g2s_cta"); + assert_eq!(target.abi_id, "i1033"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_5d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1034").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_5d_g2s_cta"); + assert_eq!(target.abi_id, "i1034"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 34ab9c2e50..13c09c0ec1 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index e0c0274b88..80384ab41c 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index 0a7b559a3b..b582151fac 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index 17608cd3ac..f7473a273a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 2fe8dd720f..1fb76115cc 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index c0323beb8d..8123ae66ba 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tma` intrinsics. @@ -5670,4 +5670,294 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ result_range: None, }), }, + GeneratedIntrinsicTarget { + marker: "v1:i1030", + id: "cp_async_bulk_tensor_1d_g2s_cta", + abi_id: "i1030", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1031", + id: "cp_async_bulk_tensor_2d_g2s_cta", + abi_id: "i1031", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1032", + id: "cp_async_bulk_tensor_3d_g2s_cta", + abi_id: "i1032", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1033", + id: "cp_async_bulk_tensor_4d_g2s_cta", + abi_id: "i1033", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1034", + id: "cp_async_bulk_tensor_5d_g2s_cta", + abi_id: "i1034", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, ]; diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 73d122c8b5..f312adf643 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index b6829d7a5a..5877724c1b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index 8517fd196b..377904b30b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index 20cb4a2df6..6fadc3c38b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index e06d161614..e5d8424d21 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/target/detect.rs b/crates/cuda-oxide-codegen/src/target/detect.rs index 4e067ece0b..20ce82e87a 100644 --- a/crates/cuda-oxide-codegen/src/target/detect.rs +++ b/crates/cuda-oxide-codegen/src/target/detect.rs @@ -574,15 +574,17 @@ pub(super) fn contains_tma_cta_group_features(contents: &str) -> bool { }) } -/// Checks TMA copies whose destination is CTA-local shared memory. +/// Checks TMA copies whose destination is CTA-local shared memory, as inline +/// PTX or as the typed `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.*` intrinsics. /// /// `.shared::cta` already existed as a source state space for shared-to-global /// copies, so the following `.global` source qualifier is part of the match. /// The destination form was introduced in PTX 8.6 but is valid on sm_90. pub(super) fn contains_tma_shared_cta_destination(contents: &str) -> bool { - contents.split(';').any(|statement| { - statement.contains("cp.async.bulk.") && statement.contains(".shared::cta.global") - }) + contents.contains("cp.async.bulk.tensor.g2s.cta.") + || contents.split(';').any(|statement| { + statement.contains("cp.async.bulk.") && statement.contains(".shared::cta.global") + }) } /// Checks PTX 8.6 TMA modifiers with a generic sm_100 architecture floor. diff --git a/crates/cuda-oxide-codegen/src/target/tests/detection_async_memory.rs b/crates/cuda-oxide-codegen/src/target/tests/detection_async_memory.rs index a81d615885..9c644810bb 100644 --- a/crates/cuda-oxide-codegen/src/target/tests/detection_async_memory.rs +++ b/crates/cuda-oxide-codegen/src/target/tests/detection_async_memory.rs @@ -58,6 +58,15 @@ fn tma_and_wgmma_raise_their_independent_ptx_floors() { assert_eq!(shared_cta_requirements.features, DetectedFeatures::Tma); assert_eq!(shared_cta_requirements.ptx_isa, PtxIsaRequirement::new(86)); + let typed_shared_cta = "call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d(ptr addrspace(3) %0, ptr addrspace(3) %1, ptr %2, i32 %3, i32 %4, i64 0, i1 false)"; + assert!(contains_tma_shared_cta_destination(typed_shared_cta)); + let typed_requirements = detect_module_requirements_in_llvm_text(typed_shared_cta); + assert_eq!(typed_requirements.features, DetectedFeatures::Tma); + assert_eq!(typed_requirements.ptx_isa, PtxIsaRequirement::new(86)); + assert!(!contains_tma_shared_cta_destination( + "call void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d(ptr addrspace(7) %0)" + )); + let shared_source = "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group;"; assert!(!contains_tma_shared_cta_destination(shared_source)); assert_eq!( diff --git a/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 449ea2c5e7..08e0ebca6d 100644 --- a/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for the generated active warp mask. diff --git a/crates/dialect-nvvm/src/ops/generated/clc.rs b/crates/dialect-nvvm/src/ops/generated/clc.rs index 1c52191590..b30ae54fc3 100644 --- a/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Cluster Launch Control operations. diff --git a/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index fe423c1171..bd84325796 100644 --- a/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated operations for the closed cluster-barrier family. diff --git a/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index 488275d4d4..80a95d46fe 100644 --- a/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/crates/dialect-nvvm/src/ops/generated/cp_async.rs index bfb3bed78c..db12d3b8cf 100644 --- a/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for classic `cp.async` instructions. diff --git a/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/crates/dialect-nvvm/src/ops/generated/debug_control.rs index d82e88c6aa..40da7799fe 100644 --- a/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated PTX debug controls. diff --git a/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/crates/dialect-nvvm/src/ops/generated/dotprod.rs index ae2ab10cf8..7459823c6e 100644 --- a/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated packed integer dot products. diff --git a/crates/dialect-nvvm/src/ops/generated/elect.rs b/crates/dialect-nvvm/src/ops/generated/elect.rs index 666b1b6c3f..f0f382fce6 100644 --- a/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated warp leader-election operation. diff --git a/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/crates/dialect-nvvm/src/ops/generated/execution_control.rs index ec29e13a32..c0e5172bc8 100644 --- a/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index 693b61f9c3..747140667a 100644 --- a/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for extended floating-point min/max. diff --git a/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 93e4b44632..003f8c9f91 100644 --- a/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated extended integer min/max. diff --git a/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index fac1180302..e06aacf2b3 100644 --- a/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index 38578d4237..65902416ff 100644 --- a/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index ec97c54ac3..c10428d2c6 100644 --- a/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/crates/dialect-nvvm/src/ops/generated/mod.rs b/crates/dialect-nvvm/src/ops/generated/mod.rs index 3c2463da32..c92b28ee04 100644 --- a/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. mod active_mask; diff --git a/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index e7a3136723..bd53a4acb8 100644 --- a/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated in-register matrix transpose. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index 30c46117b3..cd27e666da 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated packed floating-point arithmetic. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index b7ac0ce129..c631b0cfbf 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for the closed generated packed-atomic family. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index d992cb6310..f0420490c0 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated packed conversion. diff --git a/crates/dialect-nvvm/src/ops/generated/prmt.rs b/crates/dialect-nvvm/src/ops/generated/prmt.rs index ea9a88d3ac..bda0ef7215 100644 --- a/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for the closed generated `prmt` family. diff --git a/crates/dialect-nvvm/src/ops/generated/redux.rs b/crates/dialect-nvvm/src/ops/generated/redux.rs index 2236ae8a42..2217feab2f 100644 --- a/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 552a0cb3c3..1d2a557071 100644 --- a/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index 726a2256e7..5939b123cc 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated scalar arithmetic. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index 1d2c361ae5..85490e3703 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index 424fd246dc..ff680a23f4 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated scalar math. diff --git a/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 578e3a3e4e..54ebf232f5 100644 --- a/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated sparse MMA variants. diff --git a/crates/dialect-nvvm/src/ops/generated/sreg.rs b/crates/dialect-nvvm/src/ops/generated/sreg.rs index ed4f1c62d1..2c872e4d56 100644 --- a/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural NVVM operations for generated special-register reads. diff --git a/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index c3e587573f..db316ed2f0 100644 --- a/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the four generated `stmatrix` stores. diff --git a/crates/dialect-nvvm/src/ops/generated/sync.rs b/crates/dialect-nvvm/src/ops/generated/sync.rs index 02d2db9500..f681b13952 100644 --- a/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated CTA synchronization. diff --git a/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index 990baf92b6..72144096fc 100644 --- a/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Tensor Core Generation 5 operations. diff --git a/crates/dialect-nvvm/src/ops/generated/tma.rs b/crates/dialect-nvvm/src/ops/generated/tma.rs index b6457435ee..372c345fa5 100644 --- a/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Tensor Memory Accelerator operations. @@ -227,6 +227,21 @@ impl CpAsyncBulkTensorG2sTile1dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<5>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile1dOp; + +impl CpAsyncBulkTensorG2sCtaTile1dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_1d", @@ -257,6 +272,21 @@ impl CpAsyncBulkTensorG2sTile2dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<6>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile2dOp; + +impl CpAsyncBulkTensorG2sCtaTile2dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast", @@ -317,6 +347,21 @@ impl CpAsyncBulkTensorG2sTile3dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<7>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile3dOp; + +impl CpAsyncBulkTensorG2sCtaTile3dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_3d", @@ -347,6 +392,21 @@ impl CpAsyncBulkTensorG2sTile4dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<8>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile4dOp; + +impl CpAsyncBulkTensorG2sCtaTile4dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_4d", @@ -377,6 +437,21 @@ impl CpAsyncBulkTensorG2sTile5dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<9>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile5dOp; + +impl CpAsyncBulkTensorG2sCtaTile5dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_5d", @@ -1697,16 +1772,21 @@ pub(super) fn register(ctx: &mut Context) { CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op::register(ctx); CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp::register(ctx); CpAsyncBulkTensorG2sTile1dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile1dOp::register(ctx); CpAsyncBulkTensorS2gTile1dOp::register(ctx); CpAsyncBulkTensorG2sTile2dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastCg2Op::register(ctx); CpAsyncBulkTensorS2gTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile3dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile3dOp::register(ctx); CpAsyncBulkTensorS2gTile3dOp::register(ctx); CpAsyncBulkTensorG2sTile4dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile4dOp::register(ctx); CpAsyncBulkTensorS2gTile4dOp::register(ctx); CpAsyncBulkTensorG2sTile5dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile5dOp::register(ctx); CpAsyncBulkTensorS2gTile5dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col3dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col4dOp::register(ctx); diff --git a/crates/dialect-nvvm/src/ops/generated/vote.rs b/crates/dialect-nvvm/src/ops/generated/vote.rs index def9f11d20..5e413073ef 100644 --- a/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `vote.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 32aab1739b..a16cc7b5f6 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated warp synchronization. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/crates/dialect-nvvm/src/ops/generated/warp_match.rs index c0c01f5ee1..9541e45db8 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `match.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 767391efc9..65b7bfe332 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `shfl.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index 9872b0e4b9..6e139f646c 100644 --- a/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated WGMMA controls. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index 7363584ce2..adc8fa92d8 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index 54643ce57c..3c74699fbd 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index ea162f2edc..2f87e7d439 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index 695d3c8f6c..6f46ddaa3a 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index 017ed0c92e..33bae53935 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index 2bd7ca2bae..c154e0373c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 14363fb24f..36a1fd7b1a 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 7d585a7a52..845ec43d7c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 697372cc9e..f41d8459e6 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 07e0c4f542..21f01640ea 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index 7e76c72bb5..ca89cd7257 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 62a9d968a9..702746d4a4 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index cc0df48deb..eecf1013db 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index 6a719bb8a1..d650d53b20 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index 48657a9535..e140ccff05 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index c1370c9621..080ee43dad 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -313,10 +313,14 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -325,14 +329,20 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" @@ -2301,16 +2311,21 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3431,10 +3446,14 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" => Some("v1:i0328"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => Some("v1:i1030"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => Some("v1:i0335"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" => Some("v1:i0329"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => Some("v1:i1031"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => Some("v1:i0330"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -3443,14 +3462,20 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" => Some("v1:i0336"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" => Some("v1:i0332"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => Some("v1:i1032"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => Some("v1:i0337"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" => Some("v1:i0333"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => Some("v1:i1033"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => Some("v1:i0338"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" => Some("v1:i0334"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => Some("v1:i1034"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => Some("v1:i0339"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 271dc99c59..0b7c01a1d8 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 5f22170a73..5526786859 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index 95ae3a7713..8402b701f5 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index 811e9681b3..be8047dec5 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index 47f3018ae3..ca492213cb 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index acb9d2d2be..12b2fee552 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index 2d333957b1..33a5f07684 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 2593523055..beafa7e9dc 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index e91e9a70f9..ac88eb739c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index e33d6cec27..c18a1b689e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 9ad19f8a62..76ac1073e7 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 0817a6c2da..13b89cc9c2 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index 6789f53f39..5b1183b001 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index e0d6949b1a..8c56b9985b 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 1f4f183f86..f6df91c0a7 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index ee83464604..30593333be 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 3c4cff1501..91b9b4e194 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index 02cf1c9317..10f43e6807 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index e7123a1cbb..8a5dfe3d4f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index 299cdaee53..623c72549a 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index 04e44ae65b..83f94c0380 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index 522e70ad78..40407ce75e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index b1525803a9..dbf92a9f93 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index fab4c9c642..26451839ff 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index 365e539146..2818ba57de 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index 57bca785f5..3a09f6a347 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index 63a533fa45..6c655964dc 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index 7b3ca56b96..65a8454547 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index 01a5ccca63..9e39d91460 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index 9e25c77525..093de4fa41 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index 2cdb7b9a67..f6aebe33df 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index 569197d5fa..9ae61158ff 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 0670447d85..8cf74264d0 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index cc78b6207b..7c12a3b202 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index d52350e08f..233708d11f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 453444a7f5..5722fa327e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index c6c75b16f3..8c65b92000 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 04c8091720..eb3910bd94 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index df75e02029..4bf683c292 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index 3963fd8c5e..6449d6a860 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index 2d65aaa95e..7582378600 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index a228c4b86d..caf0a2ede6 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index d0ad70c41c..9e3270644c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index 90726a69e7..510bff0951 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 6cf699ea72..821f7b152c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 7ba1abc20c..d106386150 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 8c2e34bd8f..fe4ef86cae 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 0e2925c790..f4fab119ef 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 71a2fcd9bd..b2a51fef71 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 023347d56c..390fe6d6f3 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `tma` intrinsics. @@ -529,6 +529,42 @@ fn cp_async_bulk_tensor_1d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_1d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030"), + Some("v1:i1030") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta"), + Some("v1:i1030") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1030" + )); +} +#[test] fn cp_async_bulk_tensor_1d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" @@ -601,6 +637,42 @@ fn cp_async_bulk_tensor_2d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_2d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031"), + Some("v1:i1031") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta"), + Some("v1:i1031") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1031" + )); +} +#[test] fn cp_async_bulk_tensor_2d_g2s_multicast_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" @@ -745,6 +817,42 @@ fn cp_async_bulk_tensor_3d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_3d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032"), + Some("v1:i1032") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta"), + Some("v1:i1032") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1032" + )); +} +#[test] fn cp_async_bulk_tensor_3d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" @@ -817,6 +925,42 @@ fn cp_async_bulk_tensor_4d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_4d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033"), + Some("v1:i1033") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta"), + Some("v1:i1033") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1033" + )); +} +#[test] fn cp_async_bulk_tensor_4d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" @@ -889,6 +1033,42 @@ fn cp_async_bulk_tensor_5d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_5d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034"), + Some("v1:i1034") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta"), + Some("v1:i1034") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1034" + )); +} +#[test] fn cp_async_bulk_tensor_5d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index bf76ce6cd9..e0adb03554 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 481095a476..6763d61808 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index ed61420ea6..09352aea8f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index 6698b2e851..93d41ffb90 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index f587d63acd..b300e7c5cb 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 7d87b7e96a..83254bdb84 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. @@ -587,6 +587,13 @@ pub(super) fn try_dispatch( "v1:i0328", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 1, + "v1:i1030", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -601,6 +608,13 @@ pub(super) fn try_dispatch( "v1:i0329", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 2, + "v1:i1031", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => { Ok(Some(super::super::tma::emit_tma_g2s_multicast( @@ -627,6 +641,13 @@ pub(super) fn try_dispatch( "v1:i0332", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 3, + "v1:i1032", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -641,6 +662,13 @@ pub(super) fn try_dispatch( "v1:i0333", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 4, + "v1:i1033", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -655,6 +683,13 @@ pub(super) fn try_dispatch( "v1:i0334", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 5, + "v1:i1034", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 016a3ae61b..a99ef6f7b0 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index ea2efec42c..a5b667e712 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 7da0eb23dd..6c936ff379 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index f445adee0d..a3c9bf0b5e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 966b12094d..e163ac51a1 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs index 054f33755c..5393f5287c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs @@ -13,7 +13,9 @@ use crate::translator::rvalue; use crate::translator::values::ValueMap; use dialect_mir::ops::MirConstantOp; use dialect_nvvm::ops::{ - CpAsyncBulkCommitGroupOp, CpAsyncBulkTensorG2sTile1dOp, + CpAsyncBulkCommitGroupOp, CpAsyncBulkTensorG2sCtaTile1dOp, CpAsyncBulkTensorG2sCtaTile2dOp, + CpAsyncBulkTensorG2sCtaTile3dOp, CpAsyncBulkTensorG2sCtaTile4dOp, + CpAsyncBulkTensorG2sCtaTile5dOp, CpAsyncBulkTensorG2sTile1dOp, CpAsyncBulkTensorG2sTile2dMulticastCg2Op, CpAsyncBulkTensorG2sTile2dMulticastOp, CpAsyncBulkTensorG2sTile2dOp, CpAsyncBulkTensorG2sTile3dOp, CpAsyncBulkTensorG2sTile4dOp, CpAsyncBulkTensorG2sTile5dOp, CpAsyncBulkTensorS2gTile1dOp, CpAsyncBulkTensorS2gTile2dOp, @@ -61,6 +63,44 @@ pub fn emit_tma_g2s( loc: Location, dims: usize, marker: &str, +) -> TranslationResult> { + emit_tma_g2s_impl( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, dims, marker, false, + ) +} + +/// Emits a TMA G2S form whose destination belongs to the issuing CTA. +pub fn emit_tma_g2s_cta( + ctx: &mut Context, + body: &mir::Body, + args: &[mir::Operand], + target: &Option, + block_ptr: Ptr, + prev_op: Option>, + value_map: &mut ValueMap, + block_map: &[Ptr], + loc: Location, + dims: usize, + marker: &str, +) -> TranslationResult> { + emit_tma_g2s_impl( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, dims, marker, true, + ) +} + +fn emit_tma_g2s_impl( + ctx: &mut Context, + body: &mir::Body, + args: &[mir::Operand], + target: &Option, + block_ptr: Ptr, + prev_op: Option>, + value_map: &mut ValueMap, + block_map: &[Ptr], + loc: Location, + dims: usize, + marker: &str, + cta: bool, ) -> TranslationResult> { // Expected args: dst, tensor_map, coord0, [coord1, ...], barrier let expected_args = 3 + dims; // dst + tensor_map + coords + barrier @@ -135,38 +175,38 @@ pub fn emit_tma_g2s( last_op = last_op_after; } - // Add default cta_mask (i16 = 0) and cache_hint (i64 = 0) + // Cluster G2S carries a default CTA mask. CTA-local G2S has no mask. + // Both forms carry the default cache hint. use pliron::utils::apint::APInt; use std::num::NonZeroUsize; - let i16_type = IntegerType::get(ctx, 16, Signedness::Signed); let i64_type = IntegerType::get(ctx, 64, Signedness::Unsigned); - // Create constant for cta_mask = 0 - let cta_mask_apint = APInt::from_i64(0, NonZeroUsize::new(16).unwrap()); - let cta_mask_attr = pliron::builtin::attributes::IntegerAttr::new(i16_type, cta_mask_apint); - - let cta_mask_raw_op = Operation::new( - ctx, - MirConstantOp::get_concrete_op_info(), - vec![i16_type.to_handle()], - vec![], - vec![], - 0, - ); - cta_mask_raw_op.deref_mut(ctx).set_loc(loc.clone()); - let cta_mask_const = MirConstantOp::new(cta_mask_raw_op); - cta_mask_const.set_attr_value(ctx, cta_mask_attr); - - if let Some(prev) = last_op { - cta_mask_const.get_operation().insert_after(ctx, prev); - } else { - cta_mask_const - .get_operation() - .insert_at_front(block_ptr, ctx); + if !cta { + let i16_type = IntegerType::get(ctx, 16, Signedness::Signed); + let cta_mask_apint = APInt::from_i64(0, NonZeroUsize::new(16).unwrap()); + let cta_mask_attr = pliron::builtin::attributes::IntegerAttr::new(i16_type, cta_mask_apint); + let cta_mask_raw_op = Operation::new( + ctx, + MirConstantOp::get_concrete_op_info(), + vec![i16_type.to_handle()], + vec![], + vec![], + 0, + ); + cta_mask_raw_op.deref_mut(ctx).set_loc(loc.clone()); + let cta_mask_const = MirConstantOp::new(cta_mask_raw_op); + cta_mask_const.set_attr_value(ctx, cta_mask_attr); + if let Some(prev) = last_op { + cta_mask_const.get_operation().insert_after(ctx, prev); + } else { + cta_mask_const + .get_operation() + .insert_at_front(block_ptr, ctx); + } + operands.push(cta_mask_const.get_operation().deref(ctx).get_result(0)); + last_op = Some(cta_mask_const.get_operation()); } - let cta_mask = cta_mask_const.get_operation().deref(ctx).get_result(0); - operands.push(cta_mask); // Create constant for cache_hint = 0 let cache_hint_apint = APInt::from_i64(0, NonZeroUsize::new(64).unwrap()); @@ -183,19 +223,28 @@ pub fn emit_tma_g2s( cache_hint_raw_op.deref_mut(ctx).set_loc(loc.clone()); let cache_hint_const = MirConstantOp::new(cache_hint_raw_op); cache_hint_const.set_attr_value(ctx, cache_hint_attr); - cache_hint_const - .get_operation() - .insert_after(ctx, cta_mask_const.get_operation()); + if let Some(prev) = last_op { + cache_hint_const.get_operation().insert_after(ctx, prev); + } else { + cache_hint_const + .get_operation() + .insert_at_front(block_ptr, ctx); + } let cache_hint = cache_hint_const.get_operation().deref(ctx).get_result(0); operands.push(cache_hint); // Select the appropriate NVVM op based on dimensions let op_id = match dims { + 1 if cta => CpAsyncBulkTensorG2sCtaTile1dOp::get_concrete_op_info(), 1 => CpAsyncBulkTensorG2sTile1dOp::get_concrete_op_info(), + 2 if cta => CpAsyncBulkTensorG2sCtaTile2dOp::get_concrete_op_info(), 2 => CpAsyncBulkTensorG2sTile2dOp::get_concrete_op_info(), + 3 if cta => CpAsyncBulkTensorG2sCtaTile3dOp::get_concrete_op_info(), 3 => CpAsyncBulkTensorG2sTile3dOp::get_concrete_op_info(), + 4 if cta => CpAsyncBulkTensorG2sCtaTile4dOp::get_concrete_op_info(), 4 => CpAsyncBulkTensorG2sTile4dOp::get_concrete_op_info(), + 5 if cta => CpAsyncBulkTensorG2sCtaTile5dOp::get_concrete_op_info(), 5 => CpAsyncBulkTensorG2sTile5dOp::get_concrete_op_info(), _ => { return input_err!( @@ -213,7 +262,7 @@ pub fn emit_tma_g2s( ctx, op_id, vec![], // No results - operands, // dst, barrier, tensor_map, coords..., cta_mask, cache_hint + operands, // dst, barrier, tensor_map, coords..., optional cta_mask, cache_hint vec![], 0, ); diff --git a/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 95b3f0daf5..5641f9ac5e 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index fc7b13775f..0b468c1c9e 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index 79e59a186a..dbe8e9e185 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 809bdd2fe6..e1fa2d9c78 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index 30ec31ca6c..5ed884bae1 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index cc37b0bcac..302e0d4bc4 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 5f1c71aa9e..b3490bd297 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index 49d0485ba0..8edbdc0d9e 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index eb33f917ed..4da24c7413 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index 64976d1c34..e379ce282b 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index dbfe47cc8c..382e1a9ef2 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index 3c029721fe..dc9fbde433 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index 760b526608..0800bbf77b 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 6a2e00d2da..f461597b04 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index fdaa4bf21c..ddca17e5db 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index ce5d980b17..036e6d7335 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index 3d9d39e5ca..2c0404efb1 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index 8b10729fce..2fd1dc5722 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index 7f9e1b49fb..3f101fe544 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_conversion` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 7fcacbd0b9..a1f0158879 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index a776f60f9d..2acb080498 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index 3258cf9bd4..d7a77c9863 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index eb2f26f21f..3f20f7a6f4 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index 0d5779dab4..1694655b26 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 329b2242f7..3d0093a72a 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index 6d6b104715..8432184915 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index e8c341fbba..e3e2dacf70 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 639c29ce43..1dcd0db242 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 0385565067..d3e034febd 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index 6819daab9d..d25e0d6c94 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index 5ba6cbc944..e62e15bdb9 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `tma` intrinsics. @@ -12,9 +12,9 @@ use crate::conversion_interface::MirToLlvmConversion; use crate::convert::intrinsics::{ tma::PrefetchTileConfig, tma::ReduceConfig, tma::convert_control, tma::convert_g2s, - tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, tma::convert_prefetch_tile, - tma::convert_reduce_s2g, tma::convert_s2g, tma::convert_tensormap_fence, - tma::convert_tensormap_replace, + tma::convert_g2s_cta, tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, + tma::convert_prefetch_tile, tma::convert_reduce_s2g, tma::convert_s2g, + tma::convert_tensormap_fence, tma::convert_tensormap_replace, }; use dialect_nvvm::ops::{ CpAsyncBulkCommitGroupOp, CpAsyncBulkPrefetchTensor1dL2CacheHintOp, @@ -23,7 +23,9 @@ use dialect_nvvm::ops::{ CpAsyncBulkPrefetchTensor3dL2Op, CpAsyncBulkPrefetchTensor4dL2CacheHintOp, CpAsyncBulkPrefetchTensor4dL2Op, CpAsyncBulkPrefetchTensor5dL2CacheHintOp, CpAsyncBulkPrefetchTensor5dL2Op, CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp, - CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sTile1dOp, + CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sCtaTile1dOp, + CpAsyncBulkTensorG2sCtaTile2dOp, CpAsyncBulkTensorG2sCtaTile3dOp, + CpAsyncBulkTensorG2sCtaTile4dOp, CpAsyncBulkTensorG2sCtaTile5dOp, CpAsyncBulkTensorG2sTile1dOp, CpAsyncBulkTensorG2sTile2dMulticastCg2Op, CpAsyncBulkTensorG2sTile2dMulticastOp, CpAsyncBulkTensorG2sTile2dOp, CpAsyncBulkTensorG2sTile3dOp, CpAsyncBulkTensorG2sTile4dOp, CpAsyncBulkTensorG2sTile5dOp, CpAsyncBulkTensorReduceAddIm2col3dOp, @@ -383,6 +385,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile1dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile1dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 1) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile1dOp { fn convert( @@ -407,6 +421,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile2dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 2) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dMulticastOp { fn convert( @@ -455,6 +481,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile3dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile3dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 3) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile3dOp { fn convert( @@ -479,6 +517,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile4dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile4dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 4) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile4dOp { fn convert( @@ -503,6 +553,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile5dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile5dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 5) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile5dOp { fn convert( diff --git a/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index f385fe8109..e6ff753560 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index 6d49318480..bb46d3cbd0 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index fd4aa60aa9..7021adfefb 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index d6822754f5..5d69c69c78 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index c60e551e5a..2811d45cc9 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/crates/mir-lower/src/convert/intrinsics/tma.rs b/crates/mir-lower/src/convert/intrinsics/tma.rs index 9c5fcae464..cf1af03320 100644 --- a/crates/mir-lower/src/convert/intrinsics/tma.rs +++ b/crates/mir-lower/src/convert/intrinsics/tma.rs @@ -17,6 +17,7 @@ use pliron::context::{Context, Ptr}; use pliron::irbuild::dialect_conversion::{DialectConversionRewriter, OperandsInfo}; use pliron::irbuild::inserter::Inserter; use pliron::irbuild::rewriter::Rewriter; +use pliron::location::Located; use pliron::op::Op; use pliron::operation::Operation; use pliron::result::Result; @@ -58,6 +59,160 @@ fn g2s_inline_asm(dims: usize, multicast: bool, cta_group: i32) -> (String, Stri (template, constraints.join(",")) } +fn g2s_cta_inline_asm(dims: usize) -> (String, String) { + let coordinates = (0..dims) + .map(|index| format!("${}", 3 + index)) + .collect::>() + .join(", "); + let template = format!( + "cp.async.bulk.tensor.{dims}d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$0], [$2, {{{coordinates}}}], [$1];" + ); + let mut constraints = vec!["r", "r", "l"]; + constraints.extend(std::iter::repeat_n("r", dims)); + constraints.push("~{memory}"); + (template, constraints.join(",")) +} + +/// Address space of a lowered pointer operand, or `None` if it is not a +/// pointer. +fn pointer_address_space(ctx: &Context, value: pliron::value::Value) -> Option { + value + .get_type(ctx) + .deref(ctx) + .downcast_ref::() + .map(|pointer| pointer.address_space()) +} + +/// Convert a TMA G2S operation whose destination and barrier belong to the +/// issuing CTA rather than to cluster shared memory. +/// +/// The caller promises CTA locality, so generic and shared (`addrspace(3)`) +/// pointers are accepted and converted to the shared window. A pointer known +/// to be cluster shared (`addrspace(7)`, e.g. from `map_shared_rank`) may name +/// another CTA's shared memory, so it is rejected rather than reinterpreted as +/// a local offset; such copies use the cluster form. +pub(crate) fn convert_g2s_cta( + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + op: Ptr, + _operands_info: &OperandsInfo, + dims: usize, +) -> Result<()> { + let operands: Vec<_> = op.deref(ctx).operands().collect(); + let expected_operands = 3 + dims + 1; + if operands.len() != expected_operands { + return pliron::input_err_noloc!( + "TMA CTA G2S {}D requires {} operands, got {}", + dims, + expected_operands, + operands.len() + ); + } + for (index, role) in [(0, "destination"), (1, "barrier")] { + match pointer_address_space(ctx, operands[index]) { + Some(0 | 3) => {} + Some(7) => { + return pliron::input_err!( + op.deref(ctx).loc(), + "cp_async_bulk_tensor_{}d_g2s_cta requires a CTA-local shared {}, \ + but it is a cluster shared memory pointer (addrspace 7), which may \ + name another CTA's shared memory; use cp_async_bulk_tensor_{}d_g2s \ + for cluster destinations", + dims, + role, + dims + ); + } + Some(space) => { + return pliron::input_err!( + op.deref(ctx).loc(), + "cp_async_bulk_tensor_{}d_g2s_cta requires a shared memory {}, \ + but it is in address space {}", + dims, + role, + space + ); + } + None => { + return pliron::input_err!( + op.deref(ctx).loc(), + "cp_async_bulk_tensor_{}d_g2s_cta {} must be a pointer", + dims, + role + ); + } + } + } + + let dst_shared = cast_to_shared_addrspace(ctx, rewriter, operands[0]); + let barrier_shared = cast_to_shared_addrspace(ctx, rewriter, operands[1]); + let void_ty = llvm_types::VoidType::get(ctx); + + if context::lowering_options(ctx).intrinsic_backend == IntrinsicBackend::LibNvvm { + // Legacy NVVM IR has no typed CTA-destination intrinsic. Pass the + // 32-bit shared-window addresses to the exact PTX instruction. + let i32_ty = IntegerType::get(ctx, 32, Signedness::Signless); + let dst_address = llvm::PtrToIntOp::new(ctx, dst_shared, i32_ty.into()); + rewriter.insert_operation(ctx, dst_address.get_operation()); + let barrier_address = llvm::PtrToIntOp::new(ctx, barrier_shared, i32_ty.into()); + rewriter.insert_operation(ctx, barrier_address.get_operation()); + let mut inputs = vec![ + dst_address.get_operation().deref(ctx).get_result(0), + barrier_address.get_operation().deref(ctx).get_result(0), + operands[2], + ]; + inputs.extend(operands[3..3 + dims].iter().copied()); + let (template, constraints) = g2s_cta_inline_asm(dims); + inline_asm_convergent( + ctx, + rewriter, + op, + void_ty.into(), + inputs, + &template, + &constraints, + ); + rewriter.erase_operation(ctx, op); + return Ok(()); + } + + // LLVM NVPTX: the typed intrinsic the catalog probes record. + let i32_ty = IntegerType::get(ctx, 32, Signedness::Signless); + let i64_ty = IntegerType::get(ctx, 64, Signedness::Signless); + let i1_ty = IntegerType::get(ctx, 1, Signedness::Signless); + let smem_ptr_ty = llvm_types::PointerType::get(ctx, 3); + let generic_ptr_ty = llvm_types::PointerType::get(ctx, 0); + let mut arg_types: Vec = vec![ + smem_ptr_ty.into(), + smem_ptr_ty.into(), + generic_ptr_ty.into(), + ]; + for _ in 0..dims { + arg_types.push(i32_ty.into()); + } + arg_types.push(i64_ty.into()); // cache_hint + arg_types.push(i1_ty.into()); // use_cache_hint + + let intrinsic_name = format!("llvm_nvvm_cp_async_bulk_tensor_g2s_cta_tile_{}d", dims); + let func_ty = llvm_types::FuncType::get(ctx, void_ty.into(), arg_types, false); + let parent_block = op.deref(ctx).get_parent_block().unwrap(); + helpers::ensure_intrinsic_declared(ctx, parent_block, &intrinsic_name, func_ty) + .map_err(|e| pliron::input_error_noloc!("{}", e))?; + + let mut call_args = vec![dst_shared, barrier_shared]; + call_args.extend(operands[2..].iter().copied()); + let use_cache_hint = create_i1_const(ctx, rewriter, false); + call_args.push(use_cache_hint); + + let sym_name: pliron::identifier::Identifier = intrinsic_name.as_str().try_into().unwrap(); + let callee = CallOpCallable::Direct(sym_name); + let llvm_call = llvm::CallOp::new(ctx, callee, func_ty, call_args); + crate::convert::preserve_location(ctx, op, llvm_call.get_operation()); + rewriter.insert_operation(ctx, llvm_call.get_operation()); + rewriter.erase_operation(ctx, op); + Ok(()) +} + fn convert_g2s_impl( ctx: &mut Context, rewriter: &mut DialectConversionRewriter, @@ -746,7 +901,7 @@ pub(crate) fn convert_control( #[cfg(test)] mod tests { - use super::{g2s_inline_asm, reduce_inline_asm, s2g_inline_asm}; + use super::{g2s_cta_inline_asm, g2s_inline_asm, reduce_inline_asm, s2g_inline_asm}; #[test] fn inline_tma_templates_keep_exact_ptx_shapes() { @@ -764,6 +919,13 @@ mod tests { "l,l,l,r,r,h,~{memory}".into(), ) ); + assert_eq!( + g2s_cta_inline_asm(2), + ( + "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$0], [$2, {$3, $4}], [$1];".into(), + "r,r,l,r,r,~{memory}".into(), + ) + ); assert_eq!( s2g_inline_asm(5), ( diff --git a/crates/mir-lower/tests/lowering_test/tma.rs b/crates/mir-lower/tests/lowering_test/tma.rs index 7cdf3f6d81..1215c19cdb 100644 --- a/crates/mir-lower/tests/lowering_test/tma.rs +++ b/crates/mir-lower/tests/lowering_test/tma.rs @@ -171,3 +171,147 @@ fn tma_modern_libnvvm_preserves_existing_cluster_pointer_semantics() -> Result<( assert!(!ir.contains("cvta.to.shared::cta"), "{ir}"); Ok(()) } + +/// Lowers the five CTA-local G2S forms with the given destination and +/// barrier address spaces. +fn lower_g2s_cta_forms( + destination_space: u32, + barrier_space: u32, + backend: mir_lower::IntrinsicBackend, +) -> Result<(Context, Ptr), anyhow::Error> { + let mut ctx = make_test_ctx(); + let byte = IntegerType::get(&ctx, 8, Signedness::Signless).into(); + let i32_ty = IntegerType::get(&ctx, 32, Signedness::Signless).into(); + let i64_ty = IntegerType::get(&ctx, 64, Signedness::Signless).into(); + let destination = MirPtrType::get(&mut ctx, byte, true, destination_space).into(); + let barrier = MirPtrType::get(&mut ctx, byte, true, barrier_space).into(); + let pointer = MirPtrType::get_generic(&mut ctx, byte, true).into(); + let (module, entry) = build_test_kernel( + &mut ctx, + vec![ + destination, + barrier, + pointer, + i32_ty, + i32_ty, + i32_ty, + i32_ty, + i32_ty, + i64_ty, + ], + ); + let args: Vec<_> = (0..9) + .map(|index| entry.deref(&ctx).get_argument(index)) + .collect(); + for (info, dimensions) in [ + ( + nvvm::CpAsyncBulkTensorG2sCtaTile1dOp::get_concrete_op_info(), + 1, + ), + ( + nvvm::CpAsyncBulkTensorG2sCtaTile2dOp::get_concrete_op_info(), + 2, + ), + ( + nvvm::CpAsyncBulkTensorG2sCtaTile3dOp::get_concrete_op_info(), + 3, + ), + ( + nvvm::CpAsyncBulkTensorG2sCtaTile4dOp::get_concrete_op_info(), + 4, + ), + ( + nvvm::CpAsyncBulkTensorG2sCtaTile5dOp::get_concrete_op_info(), + 5, + ), + ] { + let mut operands = args[..3 + dimensions].to_vec(); + operands.push(args[8]); + Operation::new(&mut ctx, info, vec![], operands, vec![], 0).insert_at_back(entry, &ctx); + } + append_return(&mut ctx, entry); + mir_lower::lower_mir_to_llvm_with_options( + &mut ctx, + module, + mir_lower::LoweringOptions { + intrinsic_backend: backend, + ..Default::default() + }, + ) + .map_err(|error| anyhow::anyhow!("{error}"))?; + Ok((ctx, module)) +} + +#[test] +fn tma_cta_llvm_nvptx_uses_typed_shared_cta_intrinsics() -> Result<(), anyhow::Error> { + for (destination_space, barrier_space) in [(0, 0), (3, 3), (0, 3), (3, 0)] { + let (ctx, module) = lower_g2s_cta_forms( + destination_space, + barrier_space, + mir_lower::IntrinsicBackend::LlvmNvptx, + )?; + let module = Operation::get_op::(module, &ctx).unwrap(); + let ir = llvm_export::export::export_module_to_string(&ctx, &module).unwrap(); + for dimensions in 1..=5 { + let call = format!( + "call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.{dimensions}d(ptr addrspace(3)" + ); + assert_eq!(ir.matches(&call).count(), 1, "{ir}"); + } + assert!(!ir.contains("addrspace(7)"), "{ir}"); + assert!(!ir.contains("g2s.tile."), "{ir}"); + assert!(!ir.contains("asm sideeffect"), "{ir}"); + } + Ok(()) +} + +#[test] +fn tma_cta_libnvvm_emits_exact_shared_cta_ptx() -> Result<(), anyhow::Error> { + for (destination_space, barrier_space) in [(0, 0), (3, 3)] { + let (ctx, module) = lower_g2s_cta_forms( + destination_space, + barrier_space, + mir_lower::IntrinsicBackend::LibNvvm, + )?; + let module = Operation::get_op::(module, &ctx).unwrap(); + for dialect in [NvvmIrDialect::LegacyLlvm7, NvvmIrDialect::Modern] { + let ir = + export_module_to_string_with_config(&ctx, &module, &NvvmExportConfig::new(dialect)) + .unwrap(); + for dimensions in 1..=5 { + let instruction = format!( + "cp.async.bulk.tensor.{dimensions}d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$0], [$2," + ); + assert_eq!(ir.matches(&instruction).count(), 1, "{ir}"); + } + assert!(!ir.contains("addrspace(7)"), "{ir}"); + assert!(!ir.contains("shared::cluster"), "{ir}"); + } + } + Ok(()) +} + +#[test] +fn tma_cta_rejects_cluster_shared_operands() { + for backend in [ + mir_lower::IntrinsicBackend::LlvmNvptx, + mir_lower::IntrinsicBackend::LibNvvm, + ] { + for (destination_space, barrier_space, role) in [ + (7, 3, "destination"), + (3, 7, "barrier"), + (7, 0, "destination"), + ] { + let error = lower_g2s_cta_forms(destination_space, barrier_space, backend) + .err() + .unwrap_or_else(|| panic!("an addrspace(7) {role} lowered with {backend:?}")) + .to_string(); + assert!( + error.contains("g2s_cta requires a CTA-local shared") + && error.contains(role) + && error.contains("addrspace 7"), + "{error}" + ); + } + } +} diff --git a/crates/rustc-codegen-cuda/examples/tma_copy/README.md b/crates/rustc-codegen-cuda/examples/tma_copy/README.md index b89b9503b0..0f8b32dba3 100644 --- a/crates/rustc-codegen-cuda/examples/tma_copy/README.md +++ b/crates/rustc-codegen-cuda/examples/tma_copy/README.md @@ -10,9 +10,17 @@ For TMA multicast (also sm_90+), see the [`tma_multicast`](../tma_multicast/) ex ## What This Example Does 1. **tma_copy_2d_test**: Async 2D tile copy from global to shared memory -2. **tma_pipeline_test**: TMA with mbarrier for completion tracking +2. **tma_copy_2d_cta_test**: The same copy with `cp_async_bulk_tensor_2d_g2s_cta`, + whose destination is a `SharedPtr`, the issuing CTA's own shared memory by type +3. **tma_pipeline_test**: TMA with mbarrier for completion tracking -Both kernels use `#[grid_constant] tensor_map: &TmaDescriptor`. The generated +`cp_async_bulk_tensor_2d_g2s` addresses cluster shared memory, so its +destination may belong to another CTA of the cluster. The `_cta` form takes a +`SharedPtr` destination, here `SharedArray::shared_ptr(&raw mut TILE)`, and +emits `cp.async.bulk.tensor.2d.shared::cta.global` without converting any +address; `verify-code-shape.sh` checks both forms in the generated PTX. + +All kernels use `#[grid_constant] tensor_map: &TmaDescriptor`. The generated host launcher takes the descriptor by value and places its 128 bytes directly in the kernel's parameter storage. Every thread can address that same read-only value; no separate device descriptor allocation or upload is needed. The tensor diff --git a/crates/rustc-codegen-cuda/examples/tma_copy/src/main.rs b/crates/rustc-codegen-cuda/examples/tma_copy/src/main.rs index 04ba8415de..aa21c06446 100644 --- a/crates/rustc-codegen-cuda/examples/tma_copy/src/main.rs +++ b/crates/rustc-codegen-cuda/examples/tma_copy/src/main.rs @@ -9,6 +9,9 @@ //! //! Demonstrates TMA (Tensor Memory Accelerator) usage: //! - cp_async_bulk_tensor_2d_g2s: Async 2D tensor copy global → shared +//! - cp_async_bulk_tensor_2d_g2s_cta: the same copy into the issuing CTA's +//! own shared memory, whose destination is a `SharedPtr`, a CTA-shared +//! pointer by type //! - mbarrier: Barrier-based completion tracking //! - `#[grid_constant]`: the descriptor is passed by value in the launch packet, //! without allocating or uploading a separate descriptor buffer. @@ -34,7 +37,9 @@ use cuda_device::barrier::{ Barrier, fence_proxy_async_shared_cta, mbarrier_arrive, mbarrier_arrive_expect_tx, mbarrier_init, mbarrier_try_wait, }; -use cuda_device::tma::{TmaDescriptor, cp_async_bulk_tensor_2d_g2s}; +use cuda_device::tma::{ + TmaDescriptor, cp_async_bulk_tensor_2d_g2s, cp_async_bulk_tensor_2d_g2s_cta, +}; use cuda_device::{DisjointSlice, SharedArray, kernel, thread}; use cuda_host::cuda_module; use std::mem::MaybeUninit; @@ -125,6 +130,65 @@ mod kernels { } } + /// `tma_copy_2d_test` with the CTA-local copy: the tile and the barrier + /// are this CTA's own shared statics, so the `.shared::cta` form applies. + #[kernel] + pub fn tma_copy_2d_cta_test( + #[grid_constant] tensor_map: &TmaDescriptor, + mut out: DisjointSlice, + tile_x: i32, + tile_y: i32, + ) { + const TILE_SIZE: usize = 64 * 64; + const TILE_BYTES: u32 = (TILE_SIZE * 4) as u32; + static mut TILE: SharedArray = SharedArray::UNINIT; + static mut BAR: Barrier = Barrier::UNINIT; + + let tid = thread::threadIdx_x(); + let block_size = thread::blockDim_x(); + let gid = thread::index_1d(); + + if tid == 0 { + unsafe { + mbarrier_init(&raw mut BAR, block_size); + fence_proxy_async_shared_cta(); + } + } + thread::sync_threads(); + + if tid == 0 { + // SAFETY: TILE and BAR outlive the copy, which completes on BAR + // before any thread reads TILE. + unsafe { + cp_async_bulk_tensor_2d_g2s_cta( + SharedArray::shared_ptr(&raw mut TILE).cast::(), + tensor_map, + tile_x, + tile_y, + &raw mut BAR, + ); + } + } + + let token = unsafe { + if tid == 0 { + mbarrier_arrive_expect_tx(&raw const BAR, 1, TILE_BYTES) + } else { + mbarrier_arrive(&raw const BAR) + } + }; + unsafe { while !mbarrier_try_wait(&raw const BAR, token) {} } + thread::sync_threads(); + + let idx = gid.get(); + if idx < TILE_SIZE { + let val = unsafe { TILE[idx] }; + if let Some(out_elem) = out.get_mut(gid) { + *out_elem = val; + } + } + } + /// Simple TMA pipeline test - ALL threads participate in barrier. #[kernel] pub fn tma_pipeline_test( @@ -227,7 +291,8 @@ fn main() -> Result<(), Box> { println!("✓ Module loaded successfully\n"); // Run tests - run_tma_copy_test(&stream, &module)?; + run_tma_copy_test(&stream, &module, false)?; + run_tma_copy_test(&stream, &module, true)?; run_tma_pipeline_test(&stream, &module)?; println!("\n=== TMA Copy Test Complete ==="); @@ -265,8 +330,14 @@ fn verify_ptx_only(ctx: &Arc) -> Result<(), Box, module: &kernels::LoadedModule, + cta: bool, ) -> Result<(), Box> { - println!("--- Test 1: TMA Copy (tma_copy_2d_test) ---\n"); + let kernel = if cta { + "tma_copy_2d_cta_test" + } else { + "tma_copy_2d_test" + }; + println!("--- Test: TMA Copy ({kernel}) ---\n"); println!( "1. Allocating {}x{} tensor ({} floats, {} KB)", @@ -294,7 +365,7 @@ fn run_tma_copy_test( TILE_HEIGHT, )?; - println!("3. Launching tma_copy_2d_test kernel..."); + println!("3. Launching {kernel} kernel..."); let tile_x: i32 = 0; let tile_y: i32 = 0; @@ -309,14 +380,25 @@ fn run_tma_copy_test( // SAFETY: launch shape/resources match the kernel; buffers cover its accesses. unsafe { - module.tma_copy_2d_test( - (stream).as_ref(), - cfg, - tensor_map, - &mut dev_output, - tile_x, - tile_y, - ) + if cta { + module.tma_copy_2d_cta_test( + (stream).as_ref(), + cfg, + tensor_map, + &mut dev_output, + tile_x, + tile_y, + ) + } else { + module.tma_copy_2d_test( + (stream).as_ref(), + cfg, + tensor_map, + &mut dev_output, + tile_x, + tile_y, + ) + } }?; stream.synchronize()?; @@ -359,7 +441,7 @@ fn run_tma_pipeline_test( stream: &Arc, module: &kernels::LoadedModule, ) -> Result<(), Box> { - println!("\n--- Test 2: TMA Pipeline (tma_pipeline_test) ---\n"); + println!("\n--- Test: TMA Pipeline (tma_pipeline_test) ---\n"); const PIPELINE_TILE_WIDTH: u32 = 32; const PIPELINE_TILE_HEIGHT: u32 = 32; diff --git a/crates/rustc-codegen-cuda/examples/tma_copy/verify-code-shape.sh b/crates/rustc-codegen-cuda/examples/tma_copy/verify-code-shape.sh new file mode 100755 index 0000000000..818506f8f9 --- /dev/null +++ b/crates/rustc-codegen-cuda/examples/tma_copy/verify-code-shape.sh @@ -0,0 +1,75 @@ +#!/usr/bin/env bash +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 + +# cp_async_bulk_tensor_2d_g2s_cta must copy into the issuing CTA's shared +# memory: its kernel emits the `.shared::cta` destination form and, since its +# `SharedPtr` destination is shared by type, converts no address to the shared +# or cluster window. The cluster API must keep its +# `.shared::cluster` form, since its destination may belong to another CTA. + +set -euo pipefail + +root="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +ll="${root}/tma_copy.ll" +ptx="${root}/tma_copy.ptx" + +test -s "${ll}" +test -s "${ptx}" + +# Print the body of one PTX entry. Waits for the header's `{` so a forward +# declaration (which ends in `;`) is skipped. +entry_body() { + local symbol="$1" + awk -v marker="${symbol}(" ' + !emit && !candidate && index($0, marker) && index($0, ".entry") { + candidate = 1 + } + candidate && $0 ~ /^[[:space:]]*;[[:space:]]*$/ { + candidate = 0 + next + } + candidate && $0 ~ /^[[:space:]]*\{[[:space:]]*$/ { + emit = 1 + candidate = 0 + } + emit { print } + emit && index($0, "End function") != 0 { exit } + ' "${ptx}" +} + +require_entry_shape() { + local symbol="$1" + local description="$2" + local pattern="$3" + if ! entry_body "${symbol}" | grep -E "${pattern}" >/dev/null; then + echo "error: missing ${description} in ${ptx}:${symbol}" >&2 + exit 1 + fi +} + +forbid_entry_shape() { + local symbol="$1" + local description="$2" + local pattern="$3" + if entry_body "${symbol}" | grep -E "${pattern}" >/dev/null; then + echo "error: unexpected ${description} in ${ptx}:${symbol}" >&2 + exit 1 + fi +} + +if ! grep -E 'call void @llvm\.nvvm\.cp\.async\.bulk\.tensor\.g2s\.cta\.tile\.2d\(ptr addrspace\(3\)' \ + "${ll}" >/dev/null; then + echo "error: missing typed CTA-local G2S call in ${ll}" >&2 + exit 1 +fi + +require_entry_shape tma_copy_2d_cta_test "CTA-local TMA load" \ + 'cp\.async\.bulk\.tensor\.2d\.shared::cta\.global\.tile\.mbarrier::complete_tx::bytes' +forbid_entry_shape tma_copy_2d_cta_test "cluster-window destination" 'shared::cluster' +forbid_entry_shape tma_copy_2d_cta_test "generic-to-shared address conversion" 'cvta\.to\.shared' + +require_entry_shape tma_copy_2d_test "cluster TMA load" \ + 'cp\.async\.bulk\.tensor\.2d\.shared::cluster\.global\.tile\.mbarrier::complete_tx::bytes' + +echo "tma_copy code shape: PASS" diff --git a/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index 2c0b79d2e7..95429d0fed 100644 --- a/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated collector predicates for intrinsic placeholder functions. @@ -99,16 +99,21 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929", @@ -1131,16 +1136,21 @@ pub(crate) const GENERATED_INTRINSIC_PUBLIC_PATHS: &[&str] = &[ "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2", "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d", @@ -2169,16 +2179,21 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3208,16 +3223,21 @@ pub(crate) fn is_generated_intrinsic_compatibility_path(path: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2" | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_3d" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_4d" diff --git a/cuda-oxide-book/advanced/tensor-memory-accelerator.md b/cuda-oxide-book/advanced/tensor-memory-accelerator.md index a4b9da986a..4417b9ff3f 100644 --- a/cuda-oxide-book/advanced/tensor-memory-accelerator.md +++ b/cuda-oxide-book/advanced/tensor-memory-accelerator.md @@ -119,6 +119,27 @@ The coordinates (`tile_x`, `tile_y`) specify which tile to copy, in units of the box dimensions encoded in the descriptor. The barrier pointer tells the TMA engine where to signal completion. +These functions address cluster shared memory: the destination may belong to +any CTA of the cluster, so the compiler converts it to the cluster window. +When the destination is the issuing CTA's own shared memory, +`cp_async_bulk_tensor_*_g2s_cta` emits the `.shared::cta` destination form +(PTX 8.6, sm_90+) without that conversion. Its destination is a +[`SharedPtr`](shared-memory-and-synchronization.md), +a CTA-shared pointer by type, so the locality holds wherever the pointer came +from, including through helper functions: + +```rust +cp_async_bulk_tensor_2d_g2s_cta( + SharedArray::shared_ptr(&raw mut TILE).cast::(), + tensor_map, tile_x, tile_y, + &raw mut BAR, +); +``` + +On SM12x this form matters for performance: ptxas implements a non-multicast +`.shared::cluster` TMA load it cannot prove CTA-local with a runtime check and +a system call, which also makes it ignore `setmaxnreg`. + ### Only one thread issues the copy TMA copies are **not** collective operations across the block. Exactly diff --git a/cuda-oxide-book/appendix/api-quick-reference.md b/cuda-oxide-book/appendix/api-quick-reference.md index 64ab9ee8ff..5285ef7c64 100644 --- a/cuda-oxide-book/appendix/api-quick-reference.md +++ b/cuda-oxide-book/appendix/api-quick-reference.md @@ -449,6 +449,7 @@ cp_async_bulk_commit_group(); | Function | Description | |:----------------------------------------------|:-------------------------------------| | `cp_async_bulk_tensor_{1..5}d_g2s(...)` | Global → shared async bulk copy | +| `cp_async_bulk_tensor_{1..5}d_g2s_cta(...)` | Global → this CTA's `SharedPtr` | | `cp_async_bulk_tensor_{1..5}d_s2g(...)` | Shared → global async bulk copy | | `cp_async_bulk_tensor_2d_g2s_multicast(...)` | Multicast to all CTAs in cluster | | `cp_async_bulk_commit_group()` | Commit outstanding copies | diff --git a/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide-book/compiler/mlir-dialects.md index b47c20d6cb..022d4670d4 100644 --- a/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `016c9d9b` (the stamp in every `ops/generated/` file -header), the dialect holds 576 operations across 42 modules, and they come +At catalog SHA-256 `07180a97` (the stamp in every `ops/generated/` file +header), the dialect holds 581 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `016c9d9b`, the counts on this page predate the +stamp no longer starts with `07180a97`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `crates/dialect-nvvm/src/ops/`. These are the @@ -327,13 +327,13 @@ describe. There are seven modules and 26 operations: **Generated**, under `ops/generated/`, from `intrinsics/catalog.json` by `cuda-intrinsics-gen`. Every file there opens with `// @generated ... DO NOT EDIT.`, and editing one by hand is undone by the next generator run. This is -the large majority -- 35 modules and 550 operations, resolved from 1029 catalog +the large majority -- 35 modules and 555 operations, resolved from 1034 catalog entries, since several intrinsics can share one structural op: | Area | Modules | Ops | | :-------------------------- | :---------------------------------------------------------------------------- | --: | | Tensor Core Gen 5 + TMEM | `tcgen05` | 210 | -| Tensor Memory Accelerator | `tma` | 111 | +| Tensor Memory Accelerator | `tma` | 116 | | Special registers | `sreg` | 44 | | Packed (SIMD-in-register) | `packed_alu`, `packed_conversion`, `packed_atomic` | 51 | | Async copy and barriers | `cp_async`, `mbarrier_extended`, `mbarrier_basic`, `sync` | 35 | diff --git a/intrinsics/abi-v1.toml b/intrinsics/abi-v1.toml index 04cefc916a..3a06cd74fe 100644 --- a/intrinsics/abi-v1.toml +++ b/intrinsics/abi-v1.toml @@ -7173,3 +7173,38 @@ status = "active" catalog_id = "mma_sp_m16n8k64_f32_e5m2_e5m2_f32" operation_key = "matrix.mma.sp.m16n8k64.row.col.f32.e5m2.e5m2.f32.not_applicable.standard_metadata" raw_rust_signature = { safe = false, arguments = ["[f32; 4]", "[u32; 4]", "[u32; 4]", "u32", "u32"], result = "[f32; 4]" } + +[[entry]] +abi_id = "i1030" +status = "active" +catalog_id = "cp_async_bulk_tensor_1d_g2s_cta" +operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.1d" +raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "*mut u64"], result = "()" } + +[[entry]] +abi_id = "i1031" +status = "active" +catalog_id = "cp_async_bulk_tensor_2d_g2s_cta" +operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.2d" +raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "*mut u64"], result = "()" } + +[[entry]] +abi_id = "i1032" +status = "active" +catalog_id = "cp_async_bulk_tensor_3d_g2s_cta" +operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.3d" +raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "i32", "*mut u64"], result = "()" } + +[[entry]] +abi_id = "i1033" +status = "active" +catalog_id = "cp_async_bulk_tensor_4d_g2s_cta" +operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.4d" +raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "i32", "i32", "*mut u64"], result = "()" } + +[[entry]] +abi_id = "i1034" +status = "active" +catalog_id = "cp_async_bulk_tensor_5d_g2s_cta" +operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.5d" +raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "i32", "i32", "i32", "*mut u64"], result = "()" } diff --git a/intrinsics/catalog.json b/intrinsics/catalog.json index 1066e57eaf..0febab8315 100644 --- a/intrinsics/catalog.json +++ b/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "3f54245b231a5b6e19f3225710c3865b7e1a106242cce873250ef5fdc32e532e", - "overlay_sha256": "ecc7d0d7d70db737303abb6271666c67e3757a2727ba2191946a6d4e66e9f717", - "abi_ledger_sha256": "fc378e74ef6da50df31de62c3fafcb3264dfb414034e5852147359fcad4184eb", + "overlay_sha256": "ae96d212d1181075568d7e1817d05812e422ccb9ac12290f0ac7f5a80b1aff2c", + "abi_ledger_sha256": "ce40abc882afa0a313cc657758fd07aef21dd98338d8015f9cccf4edf85eb83f", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -70,6 +70,7 @@ "cuda-13.3-libnvvm-13.3.73-execution-control.json:f66216bc7aaccca9ea66fc557e19fdeeaac8c277781f669ee969e727e203aac5", "cuda-13.3-libnvvm-13.3.73-register-mma-mxf8f6f4-f32.json:e437bf84570f4df0a28a743a132e0b93f359f9431d1af03b443f324f62b715ba", "cuda-13.3-libnvvm-13.3.73-sparse-mma-fp8.json:8a57d207ffd510b8c1bb40643bdb828d3d22e5d1c5ab24e1bc1b697538e9bc21", + "cuda-13.3-libnvvm-13.3.73-tma-cta.json:352f3e82f0f1c190d99e12e717ef5357a05729521a1fbab100b40964bc9526e7", "cuda-13.3-libnvvm-13.3.73-tma-descriptor-control.json:ac44f485122404f9a900cfb94bff658160a6c23d9e2f8ef74bb751f044b848fc", "cuda-13.3-libnvvm-13.3.73-tma-reduce.json:f5df5bb51068966594660cb01b1b28c3b72cb004c456fcd8dff377ff0bee4fc0", "cuda-13.3-libnvvm-13.3.73-tma.json:8607f93605f215858a98075a0ea1394f2718d090538bcaf0a795e11b4cb66d47", @@ -127,6 +128,7 @@ "rust-llvm-23.1.0-16696adc-tcgen05-st.json:3ac8a0870333ea8a4b93a1f3ea95b27fb76477b3c8e2597b3df549125a7de010", "rust-llvm-23.1.0-16696adc-tcgen05.json:9a7fda406f9f8204ffcb99f5d055550b64fa34e939b64c2ae65a9249f7f3773b", "rust-llvm-23.1.0-16696adc-threadfence.json:35e804e8bb9c2bd2358046df50c4ccee779fbe3c0290742af0c213fbb16c2aea", + "rust-llvm-23.1.0-16696adc-tma-cta.json:a2cd6cd0e4680a53713141c8ae94feccea55865265d27ea8c4c09e754a31993f", "rust-llvm-23.1.0-16696adc-tma-descriptor-control.json:6b9255dd80bb30b3b25fea0bf00a9abdfdfa8d18926a51dee38e5133e6ee3a3f", "rust-llvm-23.1.0-16696adc-tma-reduce.json:e58a0d70e1678bb1a95b0b455b7fc424c8a17e6ec7f2bc2c4c8f1c51c034cb16", "rust-llvm-23.1.0-16696adc-tma.json:2a28f49b6965d82324b82f510e2b9a39df775aa1bb05a0c7d8dd6ba25d4719da", @@ -18550,25 +18552,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_1d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.1d", + "id": "cp_async_bulk_tensor_1d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.1d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_1d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_1D", - "asm": "cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [$tmap, {{$d0}}], [$src];", + "source_record": "TMA_G2S_CTA_TILE_1D", + "asm": "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 4, + "argument_index": 5, "value": 0 } ] @@ -18576,37 +18578,41 @@ } ], "rust": { - "abi_id": "i0335", + "abi_id": "i1030", "module": "tma", - "name": "cp_async_bulk_tensor_1d_s2g", + "name": "cp_async_bulk_tensor_1d_g2s_cta", "arguments": [ + "*mut u8", "*const u8", - "*const u8", - "i32" + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile1dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_1d", + "op_type": "CpAsyncBulkTensorG2sCtaTile1dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", "operands": [ "ptr", "ptr", - "i32" + "ptr", + "i32", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", "arguments": [ + "shared_ptr", "shared_ptr", "ptr", "i32", @@ -18615,10 +18621,10 @@ ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -18632,7 +18638,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -18649,21 +18655,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.0-16696adc-tma", + "profile": "rust-llvm-23.1.0-16696adc-tma-cta", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -18684,11 +18690,11 @@ "targets": [ "sm_90" ], - "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -18697,13 +18703,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -18712,26 +18718,26 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" } ] }, { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -18753,11 +18759,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -18768,11 +18774,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -18782,8 +18788,8 @@ } ], "tma": { - "operation": "s2g_tile1d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_cta_tile1d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -18794,34 +18800,37 @@ "bulk", "tensor", "1d", - "global", "shared::cta", + "global", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d", + "id": "cp_async_bulk_tensor_1d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.1d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_1d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_2D", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + "source_record": "TMA_TENSOR_S2G_TILE_1D", + "asm": "cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [$tmap, {{$d0}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -18829,11 +18838,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 7, - "value": 0 - }, - { - "argument_index": 8, + "argument_index": 4, "value": 0 } ] @@ -18841,63 +18846,49 @@ } ], "rust": { - "abi_id": "i0329", + "abi_id": "i0335", "module": "tma", - "name": "cp_async_bulk_tensor_2d_g2s", + "name": "cp_async_bulk_tensor_1d_s2g", "arguments": [ - "*mut u8", "*const u8", - "i32", - "i32", - "*mut u64" + "*const u8", + "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile2dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d", + "op_type": "CpAsyncBulkTensorS2gTile1dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_1d", "operands": [ "ptr", "ptr", - "ptr", - "i32", - "i32", - "i16", - "i64" + "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d", "arguments": [ - "shared_cluster_ptr", "shared_ptr", "ptr", "i32", - "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", - "ReadOnly", - "WriteOnly" + "ReadOnly", + "ReadOnly" ], "result_facts": { "no_undef": false, @@ -19061,8 +19052,8 @@ } ], "tma": { - "operation": "g2s_tile2d", - "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "operation": "s2g_tile1d", + "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19072,29 +19063,26 @@ "async", "bulk", "tensor", - "2d", - "shared::cluster", + "1d", "global", + "shared::cta", "tile", - "mbarrier::complete_tx::bytes" + "bulk_group" ], "operands": [ { "kind": "address" }, - { - "kind": "address" - }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "id": "cp_async_bulk_tensor_2d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d", "family": "tma", "source": { "kind": "llvm_imported", @@ -19102,8 +19090,8 @@ }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_TILE_CG0_2D", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19112,7 +19100,7 @@ "immediate_bindings": [ { "argument_index": 7, - "value": -1 + "value": 0 }, { "argument_index": 8, @@ -19123,30 +19111,29 @@ } ], "rust": { - "abi_id": "i0330", + "abi_id": "i0329", "module": "tma", - "name": "cp_async_bulk_tensor_2d_g2s_multicast", + "name": "cp_async_bulk_tensor_2d_g2s", "arguments": [ "*mut u8", "*const u8", "i32", "i32", - "*mut u64", - "u16" + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile2dMulticastOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast", + "op_type": "CpAsyncBulkTensorG2sTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d", "operands": [ "ptr", "ptr", @@ -19344,8 +19331,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_tile2d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19359,8 +19346,1437 @@ "shared::cluster", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster" + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.2d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_CTA_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + "predicates": [ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 6, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i1031", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_cta", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sCtaTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "arguments": [ + "shared_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc-tma-cta", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "artifact_kind": "cubin", + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_cta_tile2d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cta", + "global", + "tile", + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0330", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_multicast", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64", + "u16" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile2dMulticastOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc-tma", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->hasTMABlackwellSupport()" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0331", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64", + "u16" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + }, + "ptx_result": "()", + "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc-tma", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100a", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_100a" + ], + "representation": "one LLVM module containing the CTA-pair TMA declaration and call", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 521-byte source had SHA-256 c8f988c762ae488faa6a7a31f1391a74b33918a088449f4ec2bfbdb1ca33a201; the 1,964-byte bitcode had SHA-256 dd23f8a1a678731e95b7e2283da19530bce758b7b55ded0c98d01e99c971b978.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "typed LLVM CTA-pair TMA call", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted the exact cta_group::2 instruction with no extern call. The 1,098-byte PTX hashes were 88d30f75f80d7603ec5852572942e3406db97d10a49bdb7b65c64fa1bcea2257 for sm_100a, ed24ee1e914179959d69b7b153d5d65bbead0c864bfeb624871f9c19f5cc17f3 for sm_101a, 451331c3b63021119d0e9ccdc7617e70b505943c69e1066d193ef1b74b54523e for sm_103a, and 13c7be98f933799fafa4055423395bc1d5d4c2f4303e70f27c38ded9a0e1ace8 for sm_110a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the CTA-pair TMA instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced cubins for sm_100a, sm_103a, and sm_110a with SHA-256 7c522a3420f9856b526e23c6e3dc4067d244b27095057b91ed1b4af27a8abbcf, b3e3934adc407c847b428d93936f7d8e170059f4004218774c8755f71d599419, and ef654bdc632b1887f8622e765f8f5a2d0571b97c28eb39f17fd355483d8b9d31. Separate probes rejected sm_120a and sm_121a because cta_group::2 is unsupported.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_100a", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx93" + ], + "representation": "linked cubin retaining the CTA-pair TMA instruction", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast_cg2", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster", + "cta_group::2" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_TENSOR_S2G_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 5, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0336", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_s2g", + "arguments": [ + "*const u8", + "*const u8", + "i32", + "i32" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorS2gTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "operands": [ + "ptr", + "ptr", + "i32", + "i32" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "arguments": [ + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "ReadOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc-tma", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "s2g_tile2d", + "adapter": "s2g_pointers_coordinates_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "global", + "shared::cta", + "tile", + "bulk_group" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from shared to global memory." + }, + { + "id": "cp_async_bulk_tensor_3d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 8, + "value": 0 + }, + { + "argument_index": 9, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0332", + "module": "tma", + "name": "cp_async_bulk_tensor_3d_g2s", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc-tma", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "3d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19371,37 +20787,31 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "id": "cp_async_bulk_tensor_3d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_CTA_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", "predicates": [ - "Subtarget->hasTMABlackwellSupport()" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { "argument_index": 7, - "value": -1 - }, - { - "argument_index": 8, "value": 0 } ] @@ -19409,62 +20819,57 @@ } ], "rust": { - "abi_id": "i0331", + "abi_id": "i1032", "module": "tma", - "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "name": "cp_async_bulk_tensor_3d_g2s_cta", "arguments": [ "*mut u8", "*const u8", "i32", "i32", - "*mut u64", - "u16" + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "op_type": "CpAsyncBulkTensorG2sCtaTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", "operands": [ "ptr", "ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ "ImmArg", - "ImmArg", - "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -19485,63 +20890,39 @@ "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] }, "ptx_result": "()", - "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "targets": "all", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.0-16696adc-tma", + "profile": "rust-llvm-23.1.0-16696adc-tma-cta", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100a", + "gpu_target": "sm_90", "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19554,13 +20935,13 @@ "stages": [ { "targets": [ - "sm_100a" + "sm_90" ], - "representation": "one LLVM module containing the CTA-pair TMA declaration and call", + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "llvm-as canonicalized the intrinsic attributes. The 521-byte source had SHA-256 c8f988c762ae488faa6a7a31f1391a74b33918a088449f4ec2bfbdb1ca33a201; the 1,964-byte bitcode had SHA-256 dd23f8a1a678731e95b7e2283da19530bce758b7b55ded0c98d01e99c971b978.", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19568,14 +20949,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "typed LLVM CTA-pair TMA call", + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "The pinned LLVM emitted the exact cta_group::2 instruction with no extern call. The 1,098-byte PTX hashes were 88d30f75f80d7603ec5852572942e3406db97d10a49bdb7b65c64fa1bcea2257 for sm_100a, ed24ee1e914179959d69b7b153d5d65bbead0c864bfeb624871f9c19f5cc17f3 for sm_101a, 451331c3b63021119d0e9ccdc7617e70b505943c69e1066d193ef1b74b54523e for sm_103a, and 13c7be98f933799fafa4055423395bc1d5d4c2f4303e70f27c38ded9a0e1ace8 for sm_110a.", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19583,45 +20964,33 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "LLVM-produced PTX retaining the CTA-pair TMA instruction", + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced cubins for sm_100a, sm_103a, and sm_110a with SHA-256 7c522a3420f9856b526e23c6e3dc4067d244b27095057b91ed1b4af27a8abbcf, b3e3934adc407c847b428d93936f7d8e170059f4004218774c8755f71d599419, and ef654bdc632b1887f8622e765f8f5a2d0571b97c28eb39f17fd355483d8b9d31. Separate probes rejected sm_120a and sm_121a because cta_group::2 is unsupported.", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" } ] }, { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19634,14 +21003,14 @@ "stages": [ { "targets": [ - "compute_100a", + "compute_90", "ptx93" ], - "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19649,14 +21018,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx93" ], - "representation": "linked cubin retaining the CTA-pair TMA instruction", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -19666,8 +21035,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast_cg2", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_cta_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19677,13 +21046,11 @@ "async", "bulk", "tensor", - "2d", - "shared::cluster", + "3d", + "shared::cta", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster", - "cta_group::2" + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19694,26 +21061,23 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_2d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "id": "cp_async_bulk_tensor_3d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_2D", - "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "source_record": "TMA_TENSOR_S2G_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19721,7 +21085,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 5, + "argument_index": 6, "value": 0 } ] @@ -19729,49 +21093,52 @@ } ], "rust": { - "abi_id": "i0336", + "abi_id": "i0337", "module": "tma", - "name": "cp_async_bulk_tensor_2d_s2g", + "name": "cp_async_bulk_tensor_3d_s2g", "arguments": [ "*const u8", "*const u8", "i32", + "i32", "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile2dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "op_type": "CpAsyncBulkTensorS2gTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", "operands": [ "ptr", "ptr", "i32", + "i32", "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", "arguments": [ "shared_ptr", "ptr", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", "ReadOnly", "ReadOnly" @@ -19938,7 +21305,7 @@ } ], "tma": { - "operation": "s2g_tile2d", + "operation": "s2g_tile3d", "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, @@ -19949,7 +21316,7 @@ "async", "bulk", "tensor", - "2d", + "3d", "global", "shared::cta", "tile", @@ -19967,17 +21334,17 @@ "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_3d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_3D", - "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "source_record": "TMA_G2S_TILE_CG0_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19985,11 +21352,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 8, + "argument_index": 9, "value": 0 }, { - "argument_index": 9, + "argument_index": 10, "value": 0 } ] @@ -19997,30 +21364,31 @@ } ], "rust": { - "abi_id": "i0332", + "abi_id": "i0333", "module": "tma", - "name": "cp_async_bulk_tensor_3d_g2s", + "name": "cp_async_bulk_tensor_4d_g2s", "arguments": [ "*mut u8", "*const u8", "i32", "i32", "i32", + "i32", "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "op_type": "CpAsyncBulkTensorG2sTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", "operands": [ "ptr", "ptr", @@ -20028,13 +21396,14 @@ "i32", "i32", "i32", + "i32", "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", "arguments": [ "shared_cluster_ptr", "shared_ptr", @@ -20042,6 +21411,7 @@ "i32", "i32", "i32", + "i32", "i16", "i64", "i1", @@ -20051,10 +21421,10 @@ "results": [], "properties": [ "ImmArg", - "ImmArg", + "ImmArg", "ImmArg", "IntrConvergent", - "Range", + "Range", "ReadOnly", "WriteOnly" ], @@ -20220,7 +21590,7 @@ } ], "tma": { - "operation": "g2s_tile3d", + "operation": "g2s_tile4d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -20231,7 +21601,7 @@ "async", "bulk", "tensor", - "3d", + "4d", "shared::cluster", "global", "tile", @@ -20252,25 +21622,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_3d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_3D", - "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", + "source_record": "TMA_G2S_CTA_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 6, + "argument_index": 8, "value": 0 } ] @@ -20278,55 +21648,62 @@ } ], "rust": { - "abi_id": "i0337", + "abi_id": "i1033", "module": "tma", - "name": "cp_async_bulk_tensor_3d_s2g", + "name": "cp_async_bulk_tensor_4d_g2s_cta", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", + "op_type": "CpAsyncBulkTensorG2sCtaTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", - "i32" + "i32", + "i32", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", "arguments": [ + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -20340,7 +21717,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20357,21 +21734,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.0-16696adc-tma", + "profile": "rust-llvm-23.1.0-16696adc-tma-cta", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20392,11 +21769,11 @@ "targets": [ "sm_90" ], - "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20405,13 +21782,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20420,26 +21797,26 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" } ] }, { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20461,11 +21838,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20476,11 +21853,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -20490,8 +21867,8 @@ } ], "tma": { - "operation": "s2g_tile3d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_cta_tile4d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20501,35 +21878,38 @@ "async", "bulk", "tensor", - "3d", - "global", + "4d", "shared::cta", + "global", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_4d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", + "id": "cp_async_bulk_tensor_4d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_4D", - "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + "source_record": "TMA_TENSOR_S2G_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20537,11 +21917,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 9, - "value": 0 - }, - { - "argument_index": 10, + "argument_index": 7, "value": 0 } ] @@ -20549,69 +21925,58 @@ } ], "rust": { - "abi_id": "i0333", + "abi_id": "i0338", "module": "tma", - "name": "cp_async_bulk_tensor_4d_g2s", + "name": "cp_async_bulk_tensor_4d_s2g", "arguments": [ - "*mut u8", + "*const u8", "*const u8", "i32", "i32", "i32", - "i32", - "*mut u64" + "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", + "op_type": "CpAsyncBulkTensorS2gTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", "operands": [ "ptr", "ptr", - "ptr", - "i32", "i32", "i32", "i32", - "i16", - "i64" + "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", "arguments": [ - "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", - "ReadOnly", - "WriteOnly" + "ReadOnly", + "ReadOnly" ], "result_facts": { "no_undef": false, @@ -20775,8 +22140,8 @@ } ], "tma": { - "operation": "g2s_tile4d", - "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "operation": "s2g_tile4d", + "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20787,37 +22152,34 @@ "bulk", "tensor", "4d", - "shared::cluster", "global", + "shared::cta", "tile", - "mbarrier::complete_tx::bytes" + "bulk_group" ], "operands": [ { "kind": "address" }, - { - "kind": "address" - }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_4d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", + "id": "cp_async_bulk_tensor_5d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_4D", - "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", + "source_record": "TMA_G2S_TILE_CG0_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20825,7 +22187,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 7, + "argument_index": 10, + "value": 0 + }, + { + "argument_index": 11, "value": 0 } ] @@ -20833,58 +22199,72 @@ } ], "rust": { - "abi_id": "i0338", + "abi_id": "i0334", "module": "tma", - "name": "cp_async_bulk_tensor_4d_s2g", + "name": "cp_async_bulk_tensor_5d_g2s", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", + "op_type": "CpAsyncBulkTensorG2sTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "i16", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", "arguments": [ + "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", + "i32", + "i16", "i64", - "i1" + "i1", + "i1", + "i32" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", + "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "Range", + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -21048,8 +22428,8 @@ } ], "tma": { - "operation": "s2g_tile4d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_tile5d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -21059,47 +22439,46 @@ "async", "bulk", "tensor", - "4d", + "5d", + "shared::cluster", "global", - "shared::cta", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_5d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", + "id": "cp_async_bulk_tensor_5d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_5D", - "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + "source_record": "TMA_G2S_CTA_TILE_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 10, - "value": 0 - }, - { - "argument_index": 11, + "argument_index": 9, "value": 0 } ] @@ -21107,9 +22486,9 @@ } ], "rust": { - "abi_id": "i0334", + "abi_id": "i1034", "module": "tma", - "name": "cp_async_bulk_tensor_5d_g2s", + "name": "cp_async_bulk_tensor_5d_g2s_cta", "arguments": [ "*mut u8", "*const u8", @@ -21124,15 +22503,15 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile5dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", + "op_type": "CpAsyncBulkTensorG2sCtaTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", "operands": [ "ptr", "ptr", @@ -21142,15 +22521,14 @@ "i32", "i32", "i32", - "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", @@ -21158,19 +22536,13 @@ "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -21186,7 +22558,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21203,21 +22575,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.0-16696adc-tma", + "profile": "rust-llvm-23.1.0-16696adc-tma-cta", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.0-16696adc-tma", + "evidence_profile": "rust-llvm-23.1.0-16696adc-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21238,11 +22610,11 @@ "targets": [ "sm_90" ], - "representation": "one LLVM module containing the 14 Hopper TMA declarations and calls", + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83.", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21251,13 +22623,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "typed LLVM calls for every non-CTA-pair TMA operation", + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "The pinned LLVM emitted all 14 exact instructions with no extern calls. The 2,495-byte PTX had SHA-256 451ccb2843bfc650cf19991aaed55bcb69468de996534cdfbdca5b3f7e83916a.", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21266,26 +22638,26 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "LLVM-produced PTX retaining all 14 Hopper TMA instructions", + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 ea137713143ecca0ef1b76ebb161354034537ac05c89a34172f7b6e9c8be9f6c.", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" } ] }, { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21307,11 +22679,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21322,11 +22694,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -21336,7 +22708,7 @@ } ], "tma": { - "operation": "g2s_tile5d", + "operation": "g2s_cta_tile5d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -21348,7 +22720,7 @@ "bulk", "tensor", "5d", - "shared::cluster", + "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes" @@ -21365,7 +22737,7 @@ } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { "id": "cp_async_bulk_tensor_5d_s2g", diff --git a/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-tma-cta.json b/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-tma-cta.json new file mode 100644 index 0000000000..21d38463ee --- /dev/null +++ b/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-tma-cta.json @@ -0,0 +1,109 @@ +{ + "schema": 6, + "backend_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "backend_kind": "lib_nvvm", + "llvm_revision": "cuda-13.3-libnvvm", + "backend_version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "backend_sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "build_id_prefix": "88b802", + "nvvm_ir_version": "2.0", + "debug_ir_version": "3.2", + "defaults": { + "llvm_results": [], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "compute_90", + "ptx_feature": "PTX ISA 9.3 emitted by CUDA 13.3", + "status": "validated", + "runtime_validation": "unexecuted" + }, + "fixtures": [ + { + "id": "tma_cta_inline_sm90", + "coverage_count": 5, + "stages": [ + { + "targets": ["compute_90", "ptx93"], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60." + }, + { + "targets": ["sm_90", "ptx93"], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "matrices": [ + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_1d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_inline_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "1d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_2d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_inline_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "2d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_3d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_inline_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "3d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_4d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_inline_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "4d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_5d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_inline_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "5d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + } + ] +} diff --git a/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json b/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json new file mode 100644 index 0000000000..90dc9dd32e --- /dev/null +++ b/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json @@ -0,0 +1,116 @@ +{ + "schema": 6, + "backend_profile": "rust-llvm-23.1.0-16696adc-tma-cta", + "backend_kind": "llvm_nvptx", + "llvm_revision": "16696adcd119e6ba9cc175207d984d7021211acb", + "backend_version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "backend_sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "defaults": { + "llvm_results": [], + "concrete_llvm_results": [], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86", + "status": "validated", + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted" + }, + "fixtures": [ + { + "id": "tma_cta_typed_sm90", + "coverage_count": 5, + "stages": [ + { + "targets": ["sm_90"], + "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "ptxas", + "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" + } + ] + } + ], + "matrices": [ + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_1d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "1d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_2d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "2d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_3d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "3d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_4d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "4d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_5d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "5d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + } + ] +} diff --git a/intrinsics/generated-reference.md b/intrinsics/generated-reference.md index 71c68648ee..3e38fbb4f1 100644 --- a/intrinsics/generated-reference.md +++ b/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. --> @@ -97,16 +97,21 @@ This table is generated from the resolved catalog. Evidence stages below disting | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4 [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4.L2::cache_hint [
], ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.1d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-cta`; expects `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-cta`; expects `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster.cta_group::2 [
], [
], [
], ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-cta`; expects `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-cta`; expects `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-cta`; expects `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma`; expects `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-reduce`; expects `cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.4d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.0-16696adc-tma-reduce`; expects `cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | @@ -1388,6 +1393,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s` +- `cp_async_bulk_tensor_1d_g2s_cta` (`i1030`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta` - `cp_async_bulk_tensor_1d_s2g` (`i0335`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` @@ -1396,6 +1405,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s` +- `cp_async_bulk_tensor_2d_g2s_cta` (`i1031`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta` - `cp_async_bulk_tensor_2d_g2s_multicast` (`i0330`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` @@ -1412,6 +1425,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s` +- `cp_async_bulk_tensor_3d_g2s_cta` (`i1032`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta` - `cp_async_bulk_tensor_3d_s2g` (`i0337`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` @@ -1420,6 +1437,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s` +- `cp_async_bulk_tensor_4d_g2s_cta` (`i1033`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta` - `cp_async_bulk_tensor_4d_s2g` (`i0338`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` @@ -1428,6 +1449,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s` +- `cp_async_bulk_tensor_5d_g2s_cta` (`i1034`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta` - `cp_async_bulk_tensor_5d_s2g` (`i0339`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` @@ -5803,16 +5828,21 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cp_async_bulk_prefetch_tensor_gather4_2d_l2`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D_CH`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_1d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_1d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_1d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_2d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D_MC`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast_cg2`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_2D_MC`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_2d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_3d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_4d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_5d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_reduce_add_im2col_3d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cp_async_bulk_tensor_reduce_add_im2col_4d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none @@ -7280,6 +7310,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_1d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `ptxas` reports `CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_1d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83. @@ -7296,6 +7334,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_2d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `ptxas` reports `CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_2d_g2s_multicast`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83. @@ -7328,6 +7374,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_3d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `ptxas` reports `CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_3d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83. @@ -7344,6 +7398,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_4d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `ptxas` reports `CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_4d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83. @@ -7360,6 +7422,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_5d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `ptxas` reports `CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_5d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.0-16696adc-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_90`: `succeeded` — llvm-as canonicalized the intrinsic attributes. The 3,580-byte source had SHA-256 c87e60df3991da94c6997bfd9bb14b177031d3cf77567bebb569542c06866929; the 3,116-byte bitcode had SHA-256 00ce86db99146a468482d5471cdbb30bf092703bdb8e73d975b473e3e8db4d83. diff --git a/intrinsics/overlay/tma.toml b/intrinsics/overlay/tma.toml index b9b182c98b..36a69ae00d 100644 --- a/intrinsics/overlay/tma.toml +++ b/intrinsics/overlay/tma.toml @@ -6,6 +6,8 @@ family = "tma" [tma] llvm_evidence_profile = "rust-llvm-23.1.0-16696adc-tma" libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.73-tma" +cta_llvm_evidence_profile = "rust-llvm-23.1.0-16696adc-tma-cta" +cta_libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.73-tma-cta" reduce_llvm_evidence_profile = "rust-llvm-23.1.0-16696adc-tma-reduce" reduce_libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.73-tma-reduce" runtime_validation = "unexecuted" @@ -14,10 +16,18 @@ runtime_validation = "unexecuted" abi_id = "i0328" operation = "g2s_tile1d" +[[tma.variant]] +abi_id = "i1030" +operation = "g2s_cta_tile1d" + [[tma.variant]] abi_id = "i0329" operation = "g2s_tile2d" +[[tma.variant]] +abi_id = "i1031" +operation = "g2s_cta_tile2d" + [[tma.variant]] abi_id = "i0330" operation = "g2s_tile2d_multicast" @@ -30,14 +40,26 @@ operation = "g2s_tile2d_multicast_cg2" abi_id = "i0332" operation = "g2s_tile3d" +[[tma.variant]] +abi_id = "i1032" +operation = "g2s_cta_tile3d" + [[tma.variant]] abi_id = "i0333" operation = "g2s_tile4d" +[[tma.variant]] +abi_id = "i1033" +operation = "g2s_cta_tile4d" + [[tma.variant]] abi_id = "i0334" operation = "g2s_tile5d" +[[tma.variant]] +abi_id = "i1034" +operation = "g2s_cta_tile5d" + [[tma.variant]] abi_id = "i0335" operation = "s2g_tile1d" diff --git a/intrinsics/probes/abs_bf16x2.ll b/intrinsics/probes/abs_bf16x2.ll index c759d6494a..73dc4ed055 100644 --- a/intrinsics/probes/abs_bf16x2.ll +++ b/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/abs_f16x2.ll b/intrinsics/probes/abs_f16x2.ll index f79af7cbb5..4172174c1c 100644 --- a/intrinsics/probes/abs_f16x2.ll +++ b/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/active_mask.ll b/intrinsics/probes/active_mask.ll index 3c522a8933..55f1e07c16 100644 --- a/intrinsics/probes/active_mask.ll +++ b/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_bf16x2.ll b/intrinsics/probes/add_bf16x2.ll index c3ded6e68a..5ae970c303 100644 --- a/intrinsics/probes/add_bf16x2.ll +++ b/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_f16x2.ll b/intrinsics/probes/add_f16x2.ll index 899682ab2f..98a885d90a 100644 --- a/intrinsics/probes/add_f16x2.ll +++ b/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_f32x2.ll b/intrinsics/probes/add_f32x2.ll index 5642fd65d9..42d2341f70 100644 --- a/intrinsics/probes/add_f32x2.ll +++ b/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_ftz_f32x2.ll b/intrinsics/probes/add_ftz_f32x2.ll index 2cf4f23b8c..5da4dfa95d 100644 --- a/intrinsics/probes/add_ftz_f32x2.ll +++ b/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_f32.ll b/intrinsics/probes/add_rm_f32.ll index 22090a5d47..9db9cb414e 100644 --- a/intrinsics/probes/add_rm_f32.ll +++ b/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_f64.ll b/intrinsics/probes/add_rm_f64.ll index c142dc8d38..5edb2f144c 100644 --- a/intrinsics/probes/add_rm_f64.ll +++ b/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_ftz_f32.ll b/intrinsics/probes/add_rm_ftz_f32.ll index 07a1b1b8a9..4e44865c03 100644 --- a/intrinsics/probes/add_rm_ftz_f32.ll +++ b/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_ftz_sat_f32.ll b/intrinsics/probes/add_rm_ftz_sat_f32.ll index f42e2f9a98..ad88d52a81 100644 --- a/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_sat_f32.ll b/intrinsics/probes/add_rm_sat_f32.ll index 25e4199acd..0f9dd4d143 100644 --- a/intrinsics/probes/add_rm_sat_f32.ll +++ b/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_f32.ll b/intrinsics/probes/add_rn_f32.ll index a462c088b9..8ea8dc9b27 100644 --- a/intrinsics/probes/add_rn_f32.ll +++ b/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_f64.ll b/intrinsics/probes/add_rn_f64.ll index 3c07852ba2..c7779c1885 100644 --- a/intrinsics/probes/add_rn_f64.ll +++ b/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_ftz_f32.ll b/intrinsics/probes/add_rn_ftz_f32.ll index 65b6b22de1..6142fd05de 100644 --- a/intrinsics/probes/add_rn_ftz_f32.ll +++ b/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_ftz_sat_f32.ll b/intrinsics/probes/add_rn_ftz_sat_f32.ll index bb5f6b1bfa..4bc9cbc85e 100644 --- a/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_sat_f32.ll b/intrinsics/probes/add_rn_sat_f32.ll index 0b0a8722d0..18485b6b82 100644 --- a/intrinsics/probes/add_rn_sat_f32.ll +++ b/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_f32.ll b/intrinsics/probes/add_rp_f32.ll index e9831836d4..ccfa6da280 100644 --- a/intrinsics/probes/add_rp_f32.ll +++ b/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_f64.ll b/intrinsics/probes/add_rp_f64.ll index a0e14b6d2d..eb0a8d5120 100644 --- a/intrinsics/probes/add_rp_f64.ll +++ b/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_ftz_f32.ll b/intrinsics/probes/add_rp_ftz_f32.ll index 4ef74410fc..de930aee4d 100644 --- a/intrinsics/probes/add_rp_ftz_f32.ll +++ b/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_ftz_sat_f32.ll b/intrinsics/probes/add_rp_ftz_sat_f32.ll index 37afea0e75..f930077c9c 100644 --- a/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_sat_f32.ll b/intrinsics/probes/add_rp_sat_f32.ll index ac90be34ac..5d36282480 100644 --- a/intrinsics/probes/add_rp_sat_f32.ll +++ b/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_f32.ll b/intrinsics/probes/add_rz_f32.ll index dcd877842b..11d2fc0264 100644 --- a/intrinsics/probes/add_rz_f32.ll +++ b/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_f64.ll b/intrinsics/probes/add_rz_f64.ll index 26ca838cfe..b5b462df31 100644 --- a/intrinsics/probes/add_rz_f64.ll +++ b/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_ftz_f32.ll b/intrinsics/probes/add_rz_ftz_f32.ll index e4b909f948..eec88ddb1c 100644 --- a/intrinsics/probes/add_rz_ftz_f32.ll +++ b/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_ftz_sat_f32.ll b/intrinsics/probes/add_rz_ftz_sat_f32.ll index 01e08bb065..e6fd105527 100644 --- a/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_sat_f32.ll b/intrinsics/probes/add_rz_sat_f32.ll index 350187eb9a..205d5f0afb 100644 --- a/intrinsics/probes/add_rz_sat_f32.ll +++ b/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/all_sync.ll b/intrinsics/probes/all_sync.ll index 54df3f0fc5..f1b544b7c2 100644 --- a/intrinsics/probes/all_sync.ll +++ b/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/any_sync.ll b/intrinsics/probes/any_sync.ll index c62e870fbb..c40f6980e3 100644 --- a/intrinsics/probes/any_sync.ll +++ b/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ballot_sync.ll b/intrinsics/probes/ballot_sync.ll index 1c0a90549d..27f06ec9b8 100644 --- a/intrinsics/probes/ballot_sync.ll +++ b/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive.ll b/intrinsics/probes/barrier_cluster_arrive.ll index 158357505b..6e8393cf1c 100644 --- a/intrinsics/probes/barrier_cluster_arrive.ll +++ b/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/intrinsics/probes/barrier_cluster_arrive_aligned.ll index 901fffaf32..b36c0543fd 100644 --- a/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index e85e359968..a0e296d3bf 100644 --- a/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 9d983a9cd9..edcc5490e6 100644 --- a/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_wait.ll b/intrinsics/probes/barrier_cluster_wait.ll index b7bfb2f9c2..d574963e01 100644 --- a/intrinsics/probes/barrier_cluster_wait.ll +++ b/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_wait_aligned.ll b/intrinsics/probes/barrier_cluster_wait_aligned.ll index 1ce1ef8a24..a5612f6c8d 100644 --- a/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_arrive.ll b/intrinsics/probes/barrier_cta_arrive.ll index 8c37c624cb..d019ffbc87 100644 --- a/intrinsics/probes/barrier_cta_arrive.ll +++ b/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_arrive_aligned.ll b/intrinsics/probes/barrier_cta_arrive_aligned.ll index 5e9555a1b5..cc61cbb9f3 100644 --- a/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_sync.ll b/intrinsics/probes/barrier_cta_sync.ll index 026e7984d5..2ed3f2fa83 100644 --- a/intrinsics/probes/barrier_cta_sync.ll +++ b/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_sync_aligned.ll b/intrinsics/probes/barrier_cta_sync_aligned.ll index 6309af77f5..c2c4aaf914 100644 --- a/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_x.ll b/intrinsics/probes/block_dim_x.ll index eec87d7cb2..196b23bb0a 100644 --- a/intrinsics/probes/block_dim_x.ll +++ b/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_y.ll b/intrinsics/probes/block_dim_y.ll index 78c06ddb31..c0be67c1b1 100644 --- a/intrinsics/probes/block_dim_y.ll +++ b/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_z.ll b/intrinsics/probes/block_dim_z.ll index fc09e16013..592eb625ae 100644 --- a/intrinsics/probes/block_dim_z.ll +++ b/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_x.ll b/intrinsics/probes/block_idx_x.ll index 443bbc24f2..72c97409cd 100644 --- a/intrinsics/probes/block_idx_x.ll +++ b/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_y.ll b/intrinsics/probes/block_idx_y.ll index a7099664d3..b60f69b9f0 100644 --- a/intrinsics/probes/block_idx_y.ll +++ b/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_z.ll b/intrinsics/probes/block_idx_z.ll index 9953c01142..87f0febcb4 100644 --- a/intrinsics/probes/block_idx_z.ll +++ b/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/breakpoint.ll b/intrinsics/probes/breakpoint.ll index 032084ab2f..d8a2308285 100644 --- a/intrinsics/probes/breakpoint.ll +++ b/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/intrinsics/probes/clc_query_get_first_ctaid_x.ll index 3dae251d11..a9f43fd718 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/intrinsics/probes/clc_query_get_first_ctaid_y.ll index 77a2459449..fea6695ae9 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/intrinsics/probes/clc_query_get_first_ctaid_z.ll index 36566c4490..c6c87b553b 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_is_canceled.ll b/intrinsics/probes/clc_query_is_canceled.ll index 9f0ada4b39..20f7a2d713 100644 --- a/intrinsics/probes/clc_query_is_canceled.ll +++ b/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_try_cancel.ll b/intrinsics/probes/clc_try_cancel.ll index f89bfecf25..e8276025a4 100644 --- a/intrinsics/probes/clc_try_cancel.ll +++ b/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_try_cancel_multicast.ll b/intrinsics/probes/clc_try_cancel_multicast.ll index fa96d77c86..b58fc2a5d8 100644 --- a/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock.lib_nvvm.ll b/intrinsics/probes/clock.lib_nvvm.ll index 8ac5ec9072..7eb0cdd3f4 100644 --- a/intrinsics/probes/clock.lib_nvvm.ll +++ b/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock.llvm_nvptx.ll b/intrinsics/probes/clock.llvm_nvptx.ll index 1673854a38..c9a58f92c3 100644 --- a/intrinsics/probes/clock.llvm_nvptx.ll +++ b/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock64.lib_nvvm.ll b/intrinsics/probes/clock64.lib_nvvm.ll index 726cf4829a..392601507f 100644 --- a/intrinsics/probes/clock64.lib_nvvm.ll +++ b/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock64.llvm_nvptx.ll b/intrinsics/probes/clock64.llvm_nvptx.ll index 77d94474c2..29f6fd7d19 100644 --- a/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_count.ll b/intrinsics/probes/cluster_block_count.ll index c77af0cb8b..5fe0659420 100644 --- a/intrinsics/probes/cluster_block_count.ll +++ b/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_x.ll b/intrinsics/probes/cluster_block_idx_x.ll index f3a4ddd821..d0b0cd8308 100644 --- a/intrinsics/probes/cluster_block_idx_x.ll +++ b/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_y.ll b/intrinsics/probes/cluster_block_idx_y.ll index 7f21d0a12e..f2baa2b1fd 100644 --- a/intrinsics/probes/cluster_block_idx_y.ll +++ b/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_z.ll b/intrinsics/probes/cluster_block_idx_z.ll index 91ad3fd6cd..f67ca93127 100644 --- a/intrinsics/probes/cluster_block_idx_z.ll +++ b/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_rank.ll b/intrinsics/probes/cluster_block_rank.ll index cfb6a0ea65..754e51ec87 100644 --- a/intrinsics/probes/cluster_block_rank.ll +++ b/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_x.ll b/intrinsics/probes/cluster_dim_x.ll index 38705818f8..4d1b13b7a5 100644 --- a/intrinsics/probes/cluster_dim_x.ll +++ b/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_y.ll b/intrinsics/probes/cluster_dim_y.ll index 45853409df..d26d3f889b 100644 --- a/intrinsics/probes/cluster_dim_y.ll +++ b/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_z.ll b/intrinsics/probes/cluster_dim_z.ll index 9b9e9ab058..da26f0a80f 100644 --- a/intrinsics/probes/cluster_dim_z.ll +++ b/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_x.ll b/intrinsics/probes/cluster_grid_dim_x.ll index 8b304b0807..1a99dd4224 100644 --- a/intrinsics/probes/cluster_grid_dim_x.ll +++ b/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_y.ll b/intrinsics/probes/cluster_grid_dim_y.ll index 5146b69270..990ba89a68 100644 --- a/intrinsics/probes/cluster_grid_dim_y.ll +++ b/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_z.ll b/intrinsics/probes/cluster_grid_dim_z.ll index ad3e521abf..dc2dd7570a 100644 --- a/intrinsics/probes/cluster_grid_dim_z.ll +++ b/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_x.ll b/intrinsics/probes/cluster_idx_x.ll index 28a37d350f..0520e2f736 100644 --- a/intrinsics/probes/cluster_idx_x.ll +++ b/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_y.ll b/intrinsics/probes/cluster_idx_y.ll index b7265dc311..6793ae195c 100644 --- a/intrinsics/probes/cluster_idx_y.ll +++ b/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_z.ll b/intrinsics/probes/cluster_idx_z.ll index 2432fc25b3..d6986021de 100644 --- a/intrinsics/probes/cluster_idx_z.ll +++ b/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cos_approx_f32.ll b/intrinsics/probes/cos_approx_f32.ll index 463c30ec0e..7a3dc64cde 100644 --- a/intrinsics/probes/cos_approx_f32.ll +++ b/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cos_approx_ftz_f32.ll b/intrinsics/probes/cos_approx_ftz_f32.ll index 0cc4f4c0fe..58f088c47c 100644 --- a/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_commit_group.ll b/intrinsics/probes/cp_async_bulk_commit_group.ll index 57e937b295..866d3e0a28 100644 --- a/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index 0e945a5908..43df7a48fd 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 80d5f1d28a..4252290e59 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index 2e92cb76af..427df6bf1d 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index c2c54d7f6b..d2fd7989a9 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 05aa3cf7ea..4f98a3f0ee 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index 78afd6cecf..d2ede41597 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index bf5b50969a..5e478dcd46 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index a6a48ff86d..f09c9934eb 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index b6d31f8453..f2c8a4d8ba 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index e9f51ee313..b982155c65 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index 979a66ff62..563a8f5a7d 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index 388124aedd..9dd1b2853a 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index cf4946e20f..468ccf5e3f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll new file mode 100644 index 0000000000..fd4a2694ff --- /dev/null +++ b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll @@ -0,0 +1,18 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d(ptr addrspace(3), ptr addrspace(3), ptr, i32, i64, i1) #0 + +define void @probe_cp_async_bulk_tensor_1d_g2s_cta(ptr %dst_generic, ptr %barrier_generic, ptr %tensor_map, i32 %coord0) #0 { + %dst = addrspacecast ptr %dst_generic to ptr addrspace(3) + %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3) + call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map, i32 %coord0, i64 0, i1 false) #0 + ret void +} + +attributes #0 = { convergent } diff --git a/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 1f759f4e24..a7b5bc4129 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 805fb69ca3..9a7028cc4d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll new file mode 100644 index 0000000000..560f306d9e --- /dev/null +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll @@ -0,0 +1,18 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d(ptr addrspace(3), ptr addrspace(3), ptr, i32, i32, i64, i1) #0 + +define void @probe_cp_async_bulk_tensor_2d_g2s_cta(ptr %dst_generic, ptr %barrier_generic, ptr %tensor_map, i32 %coord0, i32 %coord1) #0 { + %dst = addrspacecast ptr %dst_generic to ptr addrspace(3) + %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3) + call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map, i32 %coord0, i32 %coord1, i64 0, i1 false) #0 + ret void +} + +attributes #0 = { convergent } diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index 37ed46fd0d..dddec13f7e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index 3c74138fa2..ee04a8d503 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 3a3aaf72c5..1b6152b185 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 2bddcc5431..af46738ba2 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll new file mode 100644 index 0000000000..7b669f098c --- /dev/null +++ b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll @@ -0,0 +1,18 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d(ptr addrspace(3), ptr addrspace(3), ptr, i32, i32, i32, i64, i1) #0 + +define void @probe_cp_async_bulk_tensor_3d_g2s_cta(ptr %dst_generic, ptr %barrier_generic, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2) #0 { + %dst = addrspacecast ptr %dst_generic to ptr addrspace(3) + %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3) + call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2, i64 0, i1 false) #0 + ret void +} + +attributes #0 = { convergent } diff --git a/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index d49e54e46d..c15db54f14 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index 171dcd3f52..04a0d94622 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll new file mode 100644 index 0000000000..2d7bc8dd40 --- /dev/null +++ b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll @@ -0,0 +1,18 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d(ptr addrspace(3), ptr addrspace(3), ptr, i32, i32, i32, i32, i64, i1) #0 + +define void @probe_cp_async_bulk_tensor_4d_g2s_cta(ptr %dst_generic, ptr %barrier_generic, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2, i32 %coord3) #0 { + %dst = addrspacecast ptr %dst_generic to ptr addrspace(3) + %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3) + call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2, i32 %coord3, i64 0, i1 false) #0 + ret void +} + +attributes #0 = { convergent } diff --git a/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index 4a92de939e..2e54e79072 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index fdd9db39f5..c1f073ca09 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll new file mode 100644 index 0000000000..6af4b57c94 --- /dev/null +++ b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll @@ -0,0 +1,18 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +declare void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d(ptr addrspace(3), ptr addrspace(3), ptr, i32, i32, i32, i32, i32, i64, i1) #0 + +define void @probe_cp_async_bulk_tensor_5d_g2s_cta(ptr %dst_generic, ptr %barrier_generic, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2, i32 %coord3, i32 %coord4) #0 { + %dst = addrspacecast ptr %dst_generic to ptr addrspace(3) + %barrier = addrspacecast ptr %barrier_generic to ptr addrspace(3) + call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d(ptr addrspace(3) %dst, ptr addrspace(3) %barrier, ptr %tensor_map, i32 %coord0, i32 %coord1, i32 %coord2, i32 %coord3, i32 %coord4, i64 0, i1 false) #0 + ret void +} + +attributes #0 = { convergent } diff --git a/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 57ba0b709b..a9961cd5cd 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index 73c0167e00..07cf6cc436 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index c535f257e9..83631dcb0c 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 10a45a3fd0..5f382d4b0f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 3264bac5b8..340116a1f9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index c31a4e0abe..cc5bf3e0df 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index c670cf96c5..cbdbf7cd5b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index b9c9b2f995..6f6be0c805 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index f97b3728ad..04227b1cb2 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index 361cd7a4a9..b5d936a8d8 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index cfcb0f8958..aea165ea75 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index dce921c5da..099b03ca66 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 8411827e33..d1bf9cd21b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index 66a8bc4819..6799f4ed12 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index ac83efb656..2f35b0394d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index 2abbcf4bb5..1f9cc87190 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index b9e69dd89c..7d4479cf78 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index 7d459c971e..20cff6aa4b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index d721c627c3..5f8d5d37c0 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 289fd56aac..ec7f48e30b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index f716e7af93..bec4cbeb96 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 9434682d63..f864236d96 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index d9248a8d04..e4f807f594 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index a287e08da9..1474a4a287 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index a901151478..71ac9d7a40 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index 78f91df966..6c3d58246d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index 9ab6cbe102..b94712e780 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index 9a94956802..292d9c44c1 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index e88074227e..e8e749be37 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index 40c298537e..b5baa4080e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index 1817063a88..fb6008e4a3 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index d8a420acc2..f1214021cb 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 3c6f129025..c7ebc966b3 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index 0873d0284c..82ba15dd2d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index a56a952adf..ce6c2df252 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index b53d95e547..7c729a1d78 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index 0f8ece64b8..14de4c42de 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index eb2657563d..816dddd41c 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index 8fb7afa4f3..3d8664354b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 73391f8c73..5d51f636bc 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index f6786f5755..bef5c9a3c4 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index 136abe2ccd..fa766315dc 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index c022a43347..c103c1552d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index a5a9cef492..75b3378207 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index 2e2b3a86bc..9a659adcb3 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index b2800ffc98..66efe41031 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index 0a667ceaf5..5b220eaf71 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index 14711c751c..95c2caaafe 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index 0f28a56460..eeae82e8aa 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index b163c49caa..e70c04e291 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index 375b37477d..f740d0f431 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 64cd8a78b1..3302d3c9a6 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index d1471d795e..781f4e6e21 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index f5e8cc6daf..de00c91732 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index 0dfcb9d37c..1162b2921a 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index d703ae7903..7cbc532968 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 6f7c1d709b..8a3c368fed 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index dc8eb106e1..8da771ce32 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index 9c41388b19..827b953e90 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 60a9c2278a..d0c4aa9f35 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 1fa891fd77..8ccccde8da 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index 5474511ebd..1d10d1f0a6 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index b8b0fae3a1..b38ef2e3c9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 840ca4b4dd..0cb475ef18 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index 406c297013..679769c336 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_wait_group.ll b/intrinsics/probes/cp_async_bulk_wait_group.ll index 9b7fcc408a..346c764dec 100644 --- a/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/intrinsics/probes/cp_async_bulk_wait_group_read.ll index 003374d612..31982788e8 100644 --- a/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_16.ll b/intrinsics/probes/cp_async_ca_16.ll index 6b878036c1..553be55a5c 100644 --- a/intrinsics/probes/cp_async_ca_16.ll +++ b/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_4.ll b/intrinsics/probes/cp_async_ca_4.ll index d95bfc6ae6..633be090e3 100644 --- a/intrinsics/probes/cp_async_ca_4.ll +++ b/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_8.ll b/intrinsics/probes/cp_async_ca_8.ll index 74320807a5..718c7ca495 100644 --- a/intrinsics/probes/cp_async_ca_8.ll +++ b/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_16.ll b/intrinsics/probes/cp_async_ca_zfill_16.ll index 4d7bd1fe1a..0bdca43a87 100644 --- a/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_4.ll b/intrinsics/probes/cp_async_ca_zfill_4.ll index 3639ac6a73..f72a0d59a5 100644 --- a/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_8.ll b/intrinsics/probes/cp_async_ca_zfill_8.ll index 0bd93f1922..763ff0175a 100644 --- a/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_cg_16.ll b/intrinsics/probes/cp_async_cg_16.ll index 1e716fa3d8..0b93fd810f 100644 --- a/intrinsics/probes/cp_async_cg_16.ll +++ b/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_cg_zfill_16.ll b/intrinsics/probes/cp_async_cg_zfill_16.ll index 31a9f08401..2d2ca38a85 100644 --- a/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_commit_group.ll b/intrinsics/probes/cp_async_commit_group.ll index 170b62c9e1..a4dff96de2 100644 --- a/intrinsics/probes/cp_async_commit_group.ll +++ b/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive.ll b/intrinsics/probes/cp_async_mbarrier_arrive.ll index a2275af196..1c0af9abb6 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index 23b21f6087..5839ff74c5 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 0060317dd7..15ccf769bd 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index e2a4eca4a0..a753f1f8bd 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_wait_all.ll b/intrinsics/probes/cp_async_wait_all.ll index 9e65e90dd1..9f5e71e02e 100644 --- a/intrinsics/probes/cp_async_wait_all.ll +++ b/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_wait_group.ll b/intrinsics/probes/cp_async_wait_group.ll index d47231be2d..b2d422928d 100644 --- a/intrinsics/probes/cp_async_wait_group.ll +++ b/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_f16x2_f32.ll b/intrinsics/probes/cvt_f16x2_f32.ll index 0674c78705..4c260bef91 100644 --- a/intrinsics/probes/cvt_f16x2_f32.ll +++ b/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_f32x2_bf16x2.ll b/intrinsics/probes/cvt_f32x2_bf16x2.ll index 70276d8a43..b3f3af6ea8 100644 --- a/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 63e4f7b437..320fded207 100644 --- a/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index 30185b23b0..6a56e85b21 100644 --- a/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index bf5a1ce5c2..0f7e575060 100644 --- a/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 82db013f53..04330043d8 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index fed9d0a3f7..289957728f 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index 7807de0859..28b7209282 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index 03dde0e7eb..55b2c91525 100644 --- a/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index ec79a85d5b..cf56bc4ce2 100644 --- a/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index 001698b45e..29251e4424 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index d77da8e78a..17794753d8 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index a771cfdd27..7b1d0baf78 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index d40e5056b0..f5a030e216 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index 38d5553e4f..d3dd9ba529 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index f9078aba54..b11861568d 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index 65f3756f9c..c858e37e34 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index 0353b4ee4e..8e06375d2b 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index a345d26bf9..26f0e31468 100644 --- a/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_tf32_f32.ll b/intrinsics/probes/cvt_rn_tf32_f32.ll index 7298293841..e1c75b5ae0 100644 --- a/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index b14cf39b8c..a32e81b979 100644 --- a/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rna_tf32_f32.ll b/intrinsics/probes/cvt_rna_tf32_f32.ll index ce19a950e7..fbe6bda7e5 100644 --- a/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 685083190f..1bba7a54d7 100644 --- a/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_f16x2_f32.ll b/intrinsics/probes/cvt_rz_f16x2_f32.ll index a4fb07c467..bc4f5e1840 100644 --- a/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index 95213c1296..ff3933779c 100644 --- a/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 94080ccd34..0ac6f04011 100644 --- a/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 09cf6b5803..b1591a5e89 100644 --- a/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_tf32_f32.ll b/intrinsics/probes/cvt_rz_tf32_f32.ll index a3a8616f32..163001942c 100644 --- a/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_f32.ll b/intrinsics/probes/div_rm_f32.ll index 0f61c9350e..d904ab408a 100644 --- a/intrinsics/probes/div_rm_f32.ll +++ b/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_f64.ll b/intrinsics/probes/div_rm_f64.ll index a190589ea7..14e55cc137 100644 --- a/intrinsics/probes/div_rm_f64.ll +++ b/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_ftz_f32.ll b/intrinsics/probes/div_rm_ftz_f32.ll index 232079de93..5737423d5c 100644 --- a/intrinsics/probes/div_rm_ftz_f32.ll +++ b/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_f32.ll b/intrinsics/probes/div_rn_f32.ll index 714ec0e096..bb34766e55 100644 --- a/intrinsics/probes/div_rn_f32.ll +++ b/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_f64.ll b/intrinsics/probes/div_rn_f64.ll index 4d0717b270..cfd734dd71 100644 --- a/intrinsics/probes/div_rn_f64.ll +++ b/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_ftz_f32.ll b/intrinsics/probes/div_rn_ftz_f32.ll index b93da31ab7..5328b563d0 100644 --- a/intrinsics/probes/div_rn_ftz_f32.ll +++ b/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_f32.ll b/intrinsics/probes/div_rp_f32.ll index c02cc35d92..22cd237157 100644 --- a/intrinsics/probes/div_rp_f32.ll +++ b/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_f64.ll b/intrinsics/probes/div_rp_f64.ll index a886b936c6..3db53fde50 100644 --- a/intrinsics/probes/div_rp_f64.ll +++ b/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_ftz_f32.ll b/intrinsics/probes/div_rp_ftz_f32.ll index 2505944120..c7fd8eae3d 100644 --- a/intrinsics/probes/div_rp_ftz_f32.ll +++ b/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_f32.ll b/intrinsics/probes/div_rz_f32.ll index 586c0d30b9..c0a393cfc3 100644 --- a/intrinsics/probes/div_rz_f32.ll +++ b/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_f64.ll b/intrinsics/probes/div_rz_f64.ll index afa12610f9..82fd9e58d9 100644 --- a/intrinsics/probes/div_rz_f64.ll +++ b/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_ftz_f32.ll b/intrinsics/probes/div_rz_ftz_f32.ll index 6f4bb9b7c5..cba3446224 100644 --- a/intrinsics/probes/div_rz_ftz_f32.ll +++ b/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp2a_s32.ll b/intrinsics/probes/dp2a_s32.ll index 3bec0f096c..6bb67d7c7b 100644 --- a/intrinsics/probes/dp2a_s32.ll +++ b/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp2a_u32.ll b/intrinsics/probes/dp2a_u32.ll index 620f3cb82b..ade881ced3 100644 --- a/intrinsics/probes/dp2a_u32.ll +++ b/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp4a_s32.ll b/intrinsics/probes/dp4a_s32.ll index 59e3002b8a..42ba93620b 100644 --- a/intrinsics/probes/dp4a_s32.ll +++ b/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp4a_u32.ll b/intrinsics/probes/dp4a_u32.ll index 28e9ba0263..13831365ea 100644 --- a/intrinsics/probes/dp4a_u32.ll +++ b/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dsmem_read_u32.ll b/intrinsics/probes/dsmem_read_u32.ll index 1af51dc3da..707d63326d 100644 --- a/intrinsics/probes/dsmem_read_u32.ll +++ b/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index fd2a06a984..5ce0c61fd6 100644 --- a/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index dd1a391113..cdf7fca387 100644 --- a/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/elect_sync.lib_nvvm.ll b/intrinsics/probes/elect_sync.lib_nvvm.ll index 62e82e1c05..5493bdf416 100644 --- a/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/elect_sync.llvm_nvptx.ll b/intrinsics/probes/elect_sync.llvm_nvptx.ll index 159f7f611e..5e969528f2 100644 --- a/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg1.lib_nvvm.ll b/intrinsics/probes/envreg1.lib_nvvm.ll index ab2c0faa35..8ea4cd24b5 100644 --- a/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg1.llvm_nvptx.ll b/intrinsics/probes/envreg1.llvm_nvptx.ll index 5a34008805..fbac7b8d20 100644 --- a/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg2.lib_nvvm.ll b/intrinsics/probes/envreg2.lib_nvvm.ll index a51f98d74c..4745ded04e 100644 --- a/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg2.llvm_nvptx.ll b/intrinsics/probes/envreg2.llvm_nvptx.ll index 63942849b8..75c1f3eb57 100644 --- a/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_f16.ll b/intrinsics/probes/ex2_approx_f16.ll index e72dcc6a2c..5a0ab6946e 100644 --- a/intrinsics/probes/ex2_approx_f16.ll +++ b/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_f32.ll b/intrinsics/probes/ex2_approx_f32.ll index d9ca0c7e96..13b172d8b2 100644 --- a/intrinsics/probes/ex2_approx_f32.ll +++ b/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_ftz_f32.ll b/intrinsics/probes/ex2_approx_ftz_f32.ll index 5b3c4a6c53..d6de6f2752 100644 --- a/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index 709ea6786b..33a8b09548 100644 --- a/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index d37dcb85c4..b6c01e93dc 100644 --- a/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index 0420e795fb..6ffa4750b8 100644 --- a/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_shared_cta.ll b/intrinsics/probes/fence_proxy_async_shared_cta.ll index 756ab152ef..b91c618b36 100644 --- a/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 9795cc8a82..bee7ee4304 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index af28e9032a..1ee5156051 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index e01e367782..3f40e160bb 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index 99925c3ec9..d4be6cb83d 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index a01a1e6b9a..1765fb2ede 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index 435e4e97c8..6731e3ae51 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index 15ab9ca536..6adfd1ee11 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index 61f2340f44..d45ca1c3f3 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_bf16x2.ll b/intrinsics/probes/fma_bf16x2.ll index d03cb35b49..8ff01ee25e 100644 --- a/intrinsics/probes/fma_bf16x2.ll +++ b/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_f16x2.ll b/intrinsics/probes/fma_f16x2.ll index 11570cb7dc..2b15134189 100644 --- a/intrinsics/probes/fma_f16x2.ll +++ b/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_f32x2.ll b/intrinsics/probes/fma_f32x2.ll index 1441f8cb69..198490e232 100644 --- a/intrinsics/probes/fma_f32x2.ll +++ b/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_f16x2.ll b/intrinsics/probes/fma_ftz_f16x2.ll index 56787d210a..96e87d7a96 100644 --- a/intrinsics/probes/fma_ftz_f16x2.ll +++ b/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_f32x2.ll b/intrinsics/probes/fma_ftz_f32x2.ll index aea071a677..9a6b8dfbf8 100644 --- a/intrinsics/probes/fma_ftz_f32x2.ll +++ b/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_relu_f16x2.ll b/intrinsics/probes/fma_ftz_relu_f16x2.ll index f95f66a01f..f966112fdd 100644 --- a/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_sat_f16x2.ll b/intrinsics/probes/fma_ftz_sat_f16x2.ll index 3b91ebee55..11f0d540df 100644 --- a/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_relu_bf16x2.ll b/intrinsics/probes/fma_relu_bf16x2.ll index b9782e147a..fdcf88f4c3 100644 --- a/intrinsics/probes/fma_relu_bf16x2.ll +++ b/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_relu_f16x2.ll b/intrinsics/probes/fma_relu_f16x2.ll index 5ce2a6ce6a..da1bc87b26 100644 --- a/intrinsics/probes/fma_relu_f16x2.ll +++ b/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_f32.ll b/intrinsics/probes/fma_rm_f32.ll index f0c4c72614..d87026fa3c 100644 --- a/intrinsics/probes/fma_rm_f32.ll +++ b/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_f64.ll b/intrinsics/probes/fma_rm_f64.ll index cfc8075316..2df943381e 100644 --- a/intrinsics/probes/fma_rm_f64.ll +++ b/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_ftz_f32.ll b/intrinsics/probes/fma_rm_ftz_f32.ll index 6324e54129..5ff1db9381 100644 --- a/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/intrinsics/probes/fma_rm_ftz_sat_f32.ll index f4bf2b34c0..9556311b71 100644 --- a/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_sat_f32.ll b/intrinsics/probes/fma_rm_sat_f32.ll index 7adc765483..64bf86a555 100644 --- a/intrinsics/probes/fma_rm_sat_f32.ll +++ b/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_f32.ll b/intrinsics/probes/fma_rn_f32.ll index 880d1833af..bcaa0baeb3 100644 --- a/intrinsics/probes/fma_rn_f32.ll +++ b/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_f64.ll b/intrinsics/probes/fma_rn_f64.ll index 7a21f94476..7ef6e432b5 100644 --- a/intrinsics/probes/fma_rn_f64.ll +++ b/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_ftz_f32.ll b/intrinsics/probes/fma_rn_ftz_f32.ll index 5701f6a574..46b28f499e 100644 --- a/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/intrinsics/probes/fma_rn_ftz_sat_f32.ll index c4eb49e0fe..3c4add78e1 100644 --- a/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_sat_f32.ll b/intrinsics/probes/fma_rn_sat_f32.ll index 9ecae992d5..ee1b488f38 100644 --- a/intrinsics/probes/fma_rn_sat_f32.ll +++ b/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_f32.ll b/intrinsics/probes/fma_rp_f32.ll index 5c7e2e252b..1add06a238 100644 --- a/intrinsics/probes/fma_rp_f32.ll +++ b/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_f64.ll b/intrinsics/probes/fma_rp_f64.ll index a69a5ffb03..f3397e36e1 100644 --- a/intrinsics/probes/fma_rp_f64.ll +++ b/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_ftz_f32.ll b/intrinsics/probes/fma_rp_ftz_f32.ll index 59c486b424..df7d173918 100644 --- a/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/intrinsics/probes/fma_rp_ftz_sat_f32.ll index eb3e059038..874403e7f7 100644 --- a/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_sat_f32.ll b/intrinsics/probes/fma_rp_sat_f32.ll index cd7e49669a..70dd992440 100644 --- a/intrinsics/probes/fma_rp_sat_f32.ll +++ b/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_f32.ll b/intrinsics/probes/fma_rz_f32.ll index fad52e7945..2f73dda2f7 100644 --- a/intrinsics/probes/fma_rz_f32.ll +++ b/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_f64.ll b/intrinsics/probes/fma_rz_f64.ll index b26219dbb9..1af7f41b46 100644 --- a/intrinsics/probes/fma_rz_f64.ll +++ b/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_ftz_f32.ll b/intrinsics/probes/fma_rz_ftz_f32.ll index 0f7d190873..c1822cffd8 100644 --- a/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 774bdb33fc..aa07209bf8 100644 --- a/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_sat_f32.ll b/intrinsics/probes/fma_rz_sat_f32.ll index e8b2396b07..04f3de7ac8 100644 --- a/intrinsics/probes/fma_rz_sat_f32.ll +++ b/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_sat_f16x2.ll b/intrinsics/probes/fma_sat_f16x2.ll index c69f4d0b0d..b8ad2768b8 100644 --- a/intrinsics/probes/fma_sat_f16x2.ll +++ b/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/globaltimer.lib_nvvm.ll b/intrinsics/probes/globaltimer.lib_nvvm.ll index bd0606ec80..31ecb9b040 100644 --- a/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/globaltimer.llvm_nvptx.ll b/intrinsics/probes/globaltimer.llvm_nvptx.ll index f8833f42c6..e8a76d0fa4 100644 --- a/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dependency_launch_dependents.ll b/intrinsics/probes/grid_dependency_launch_dependents.ll index 8793515f3e..51e695d0df 100644 --- a/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dependency_wait.ll b/intrinsics/probes/grid_dependency_wait.ll index 3dcc697fda..b1300bcf66 100644 --- a/intrinsics/probes/grid_dependency_wait.ll +++ b/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_x.ll b/intrinsics/probes/grid_dim_x.ll index 6160b58479..a393ef90b2 100644 --- a/intrinsics/probes/grid_dim_x.ll +++ b/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_y.ll b/intrinsics/probes/grid_dim_y.ll index e568eef09c..16bd5bf3d3 100644 --- a/intrinsics/probes/grid_dim_y.ll +++ b/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_z.ll b/intrinsics/probes/grid_dim_z.ll index e3ab0c2c78..0c220d0eb8 100644 --- a/intrinsics/probes/grid_dim_z.ll +++ b/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/gridid.lib_nvvm.ll b/intrinsics/probes/gridid.lib_nvvm.ll index 91bf8bae59..25b0693ef4 100644 --- a/intrinsics/probes/gridid.lib_nvvm.ll +++ b/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/gridid.llvm_nvptx.ll b/intrinsics/probes/gridid.llvm_nvptx.ll index 4d3af9c030..e9b91d0e4d 100644 --- a/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lane_id.ll b/intrinsics/probes/lane_id.ll index 4375748889..a54a930e0b 100644 --- a/intrinsics/probes/lane_id.ll +++ b/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_eq.ll b/intrinsics/probes/lanemask_eq.ll index 3f0c5becda..76a3e52700 100644 --- a/intrinsics/probes/lanemask_eq.ll +++ b/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_ge.ll b/intrinsics/probes/lanemask_ge.ll index dcaf2c81f7..6c5f1587bd 100644 --- a/intrinsics/probes/lanemask_ge.ll +++ b/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_gt.ll b/intrinsics/probes/lanemask_gt.ll index 4d5ea2baab..92f7da0549 100644 --- a/intrinsics/probes/lanemask_gt.ll +++ b/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_le.ll b/intrinsics/probes/lanemask_le.ll index 42a538ad53..4af4efe667 100644 --- a/intrinsics/probes/lanemask_le.ll +++ b/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_lt.ll b/intrinsics/probes/lanemask_lt.ll index 29445102d1..bf28d611e5 100644 --- a/intrinsics/probes/lanemask_lt.ll +++ b/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index 48aae7f31f..dd660a9373 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 335ea7f964..d440bb3e25 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 46110462c5..14a1d8476b 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index e54bd8ddb6..10ee174a3d 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index 186e16a9f4..9220bdfc7a 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index 48a46dd253..d2c9c2a60a 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index a14091597c..9041768cd6 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index 300fe10d75..b31fe24464 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index 4d08f3a647..c57812ea15 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 04357788fb..971008883f 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index 47a24b0408..ab5b7fdcbc 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 16b6279264..2085148c57 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 6de5101df0..dc04d0df20 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 2e62e6f8bc..49e448fb7f 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index ff9314f5c5..2d99cc876a 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index c8a8192730..c8b9871103 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index ef545a1dec..ec2fe5958b 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index 8feaa576cf..380cd9887e 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lg2_approx_f32.ll b/intrinsics/probes/lg2_approx_f32.ll index f162fddc72..436326e2fb 100644 --- a/intrinsics/probes/lg2_approx_f32.ll +++ b/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lg2_approx_ftz_f32.ll b/intrinsics/probes/lg2_approx_ftz_f32.ll index fe37b75dd4..7beb6429c4 100644 --- a/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/map_shared_rank.ll b/intrinsics/probes/map_shared_rank.ll index 3e92103cf9..4d308a4734 100644 --- a/intrinsics/probes/map_shared_rank.ll +++ b/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_all_i64_sync.ll b/intrinsics/probes/match_all_i64_sync.ll index 5c4808bd9a..2ce11028c9 100644 --- a/intrinsics/probes/match_all_i64_sync.ll +++ b/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_all_sync.ll b/intrinsics/probes/match_all_sync.ll index 10c5ee442b..b606f0b4ea 100644 --- a/intrinsics/probes/match_all_sync.ll +++ b/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_any_i64_sync.ll b/intrinsics/probes/match_any_i64_sync.ll index 90b70fd545..9de76747a2 100644 --- a/intrinsics/probes/match_any_i64_sync.ll +++ b/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_any_sync.ll b/intrinsics/probes/match_any_sync.ll index 9288ae011d..579cf8cef0 100644 --- a/intrinsics/probes/match_any_sync.ll +++ b/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_bf16.ll b/intrinsics/probes/max_bf16.ll index 42e2be579f..0469b8a697 100644 --- a/intrinsics/probes/max_bf16.ll +++ b/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_bf16x2.ll b/intrinsics/probes/max_bf16x2.ll index ce7f8125b1..53f7967c91 100644 --- a/intrinsics/probes/max_bf16x2.ll +++ b/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_f16.ll b/intrinsics/probes/max_f16.ll index e208b199e1..85b588254d 100644 --- a/intrinsics/probes/max_f16.ll +++ b/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_f16x2.ll b/intrinsics/probes/max_f16x2.ll index 8789a5d3ec..86619e88b7 100644 --- a/intrinsics/probes/max_f16x2.ll +++ b/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_f16.ll b/intrinsics/probes/max_ftz_f16.ll index 458003c675..1c566a97c6 100644 --- a/intrinsics/probes/max_ftz_f16.ll +++ b/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_f16x2.ll b/intrinsics/probes/max_ftz_f16x2.ll index 1bd819eaa1..a9ca00be35 100644 --- a/intrinsics/probes/max_ftz_f16x2.ll +++ b/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_f16.ll b/intrinsics/probes/max_ftz_nan_f16.ll index f3ed7912a0..69d81a27c9 100644 --- a/intrinsics/probes/max_ftz_nan_f16.ll +++ b/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_f16x2.ll b/intrinsics/probes/max_ftz_nan_f16x2.ll index cfc8fec1dd..b6366cf4f8 100644 --- a/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index dcc8bf5ce5..b25d68c6e6 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index e1a1ac6d9c..09fc4939bb 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 5aab79c28f..67392130e4 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 06a882d0cd..cfe0fad8ae 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 5cee3d6a52..be852377d3 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index f6fed3626a..2e1d790803 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_bf16.ll b/intrinsics/probes/max_nan_bf16.ll index 20f90efb9c..53f0c5a635 100644 --- a/intrinsics/probes/max_nan_bf16.ll +++ b/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_bf16x2.ll b/intrinsics/probes/max_nan_bf16x2.ll index 4938b05160..22ab75c3b2 100644 --- a/intrinsics/probes/max_nan_bf16x2.ll +++ b/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_f16.ll b/intrinsics/probes/max_nan_f16.ll index b48856a401..d6886d5fdc 100644 --- a/intrinsics/probes/max_nan_f16.ll +++ b/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_f16x2.ll b/intrinsics/probes/max_nan_f16x2.ll index f7840ca365..ca9314db30 100644 --- a/intrinsics/probes/max_nan_f16x2.ll +++ b/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index f234e4ba90..9c689c312c 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index 0e1a0e22ed..cb8a1fedee 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 7f977d3f5c..09f95c5333 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index ad91a7b8cd..b6887a3433 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 505dc8d7c1..b9ccc9093b 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_relu_s16x2.ll b/intrinsics/probes/max_relu_s16x2.ll index 25a6d49c6b..74d545e427 100644 --- a/intrinsics/probes/max_relu_s16x2.ll +++ b/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_relu_s32.ll b/intrinsics/probes/max_relu_s32.ll index d364154157..d11695ec35 100644 --- a/intrinsics/probes/max_relu_s32.ll +++ b/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_s16x2.ll b/intrinsics/probes/max_s16x2.ll index 1860d6088d..491502cf34 100644 --- a/intrinsics/probes/max_s16x2.ll +++ b/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_u16x2.ll b/intrinsics/probes/max_u16x2.ll index d2a27f7c79..498458c972 100644 --- a/intrinsics/probes/max_u16x2.ll +++ b/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_bf16.ll b/intrinsics/probes/max_xorsign_abs_bf16.ll index 174a9ad97b..e03768efa5 100644 --- a/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/intrinsics/probes/max_xorsign_abs_bf16x2.ll index 6c9213a7a4..7bda95779b 100644 --- a/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f16.ll b/intrinsics/probes/max_xorsign_abs_f16.ll index 12c7994b29..6100cf9892 100644 --- a/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f16x2.ll b/intrinsics/probes/max_xorsign_abs_f16x2.ll index b7ca356ab1..18b6a0257a 100644 --- a/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f32.ll b/intrinsics/probes/max_xorsign_abs_f32.ll index e0be80518f..a6701f4642 100644 --- a/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive.ll b/intrinsics/probes/mbarrier_arrive.ll index b592698137..9a4cb2dd94 100644 --- a/intrinsics/probes/mbarrier_arrive.ll +++ b/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_cluster.ll b/intrinsics/probes/mbarrier_arrive_cluster.ll index 93bb09121d..0879c09184 100644 --- a/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/intrinsics/probes/mbarrier_arrive_expect_tx.ll index c36d6aad5c..e3d6bdfd41 100644 --- a/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index ad001c0ebb..77d2a74d60 100644 --- a/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_no_complete.ll b/intrinsics/probes/mbarrier_arrive_no_complete.ll index fc17a2816f..cbcc00c6e8 100644 --- a/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_init.ll b/intrinsics/probes/mbarrier_init.ll index ab6038402e..b409f0116b 100644 --- a/intrinsics/probes/mbarrier_init.ll +++ b/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_inval.ll b/intrinsics/probes/mbarrier_inval.ll index 21d816dafa..1fc2a20e7a 100644 --- a/intrinsics/probes/mbarrier_inval.ll +++ b/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_test_wait.ll b/intrinsics/probes/mbarrier_test_wait.ll index e16b8dc05c..cd5e460ed5 100644 --- a/intrinsics/probes/mbarrier_test_wait.ll +++ b/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait.ll b/intrinsics/probes/mbarrier_try_wait.ll index 9ea2000fdc..77029f9abf 100644 --- a/intrinsics/probes/mbarrier_try_wait.ll +++ b/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait_parity.ll b/intrinsics/probes/mbarrier_try_wait_parity.ll index ddfe48da54..8fa9b85202 100644 --- a/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 609bfd5de7..81255e13ee 100644 --- a/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_bf16.ll b/intrinsics/probes/min_bf16.ll index 685c8beee2..27dc4f0fae 100644 --- a/intrinsics/probes/min_bf16.ll +++ b/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_bf16x2.ll b/intrinsics/probes/min_bf16x2.ll index 98d2a848a7..0819ac4a6b 100644 --- a/intrinsics/probes/min_bf16x2.ll +++ b/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_f16.ll b/intrinsics/probes/min_f16.ll index db6d4cf788..8103c31fb3 100644 --- a/intrinsics/probes/min_f16.ll +++ b/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_f16x2.ll b/intrinsics/probes/min_f16x2.ll index fd6aad9c0a..f03e162377 100644 --- a/intrinsics/probes/min_f16x2.ll +++ b/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_f16.ll b/intrinsics/probes/min_ftz_f16.ll index d7f60d85a2..bffb5a5288 100644 --- a/intrinsics/probes/min_ftz_f16.ll +++ b/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_f16x2.ll b/intrinsics/probes/min_ftz_f16x2.ll index 2c0522f034..4e1a0a8242 100644 --- a/intrinsics/probes/min_ftz_f16x2.ll +++ b/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_f16.ll b/intrinsics/probes/min_ftz_nan_f16.ll index 40383f5720..51b83c8712 100644 --- a/intrinsics/probes/min_ftz_nan_f16.ll +++ b/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_f16x2.ll b/intrinsics/probes/min_ftz_nan_f16x2.ll index 9d8379e621..be55d87dfe 100644 --- a/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index b97e669b2c..51bfc87af2 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index 984663ea51..45475e4160 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 5b1816dc56..8cf3d34d01 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index c6cd152c15..c202282666 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index 535f2934a3..4b979781af 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index d4b65452e4..041d8d0e71 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_bf16.ll b/intrinsics/probes/min_nan_bf16.ll index 800e9c770b..9851a91ba4 100644 --- a/intrinsics/probes/min_nan_bf16.ll +++ b/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_bf16x2.ll b/intrinsics/probes/min_nan_bf16x2.ll index bfc7d311a0..bc5f837349 100644 --- a/intrinsics/probes/min_nan_bf16x2.ll +++ b/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_f16.ll b/intrinsics/probes/min_nan_f16.ll index a591854fb3..ba0eae455c 100644 --- a/intrinsics/probes/min_nan_f16.ll +++ b/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_f16x2.ll b/intrinsics/probes/min_nan_f16x2.ll index 4761a5aef5..3699f1feaf 100644 --- a/intrinsics/probes/min_nan_f16x2.ll +++ b/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 92b345ae9e..502070fc3e 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index 8d77182d42..7e68669879 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 42ae49f2f1..610aea9cd0 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index 4f671739a8..4f3e2c9dd4 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/intrinsics/probes/min_nan_xorsign_abs_f32.ll index 86ecde55d6..e6a494857e 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_relu_s16x2.ll b/intrinsics/probes/min_relu_s16x2.ll index 114ba0ea2b..a4991b5033 100644 --- a/intrinsics/probes/min_relu_s16x2.ll +++ b/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_relu_s32.ll b/intrinsics/probes/min_relu_s32.ll index 72ec4e9862..06a889d697 100644 --- a/intrinsics/probes/min_relu_s32.ll +++ b/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_s16x2.ll b/intrinsics/probes/min_s16x2.ll index 024ac9a79d..0b14a9c3f5 100644 --- a/intrinsics/probes/min_s16x2.ll +++ b/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_u16x2.ll b/intrinsics/probes/min_u16x2.ll index 17efe05f7d..c59076e1c8 100644 --- a/intrinsics/probes/min_u16x2.ll +++ b/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_bf16.ll b/intrinsics/probes/min_xorsign_abs_bf16.ll index b2c98a1396..30738fc5b4 100644 --- a/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 5c5eda21cc..04a71495ca 100644 --- a/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f16.ll b/intrinsics/probes/min_xorsign_abs_f16.ll index fad3ee263c..3b445bcc71 100644 --- a/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f16x2.ll b/intrinsics/probes/min_xorsign_abs_f16x2.ll index 83e2d083f1..02be86e3be 100644 --- a/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f32.ll b/intrinsics/probes/min_xorsign_abs_f32.ll index 0600e57238..9c6f372d2b 100644 --- a/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index d963d02bf2..83049598df 100644 --- a/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index ff9b31661d..b22979e2e1 100644 --- a/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/intrinsics/probes/mma_m16n8k16_f16_f16.ll index 61856d680a..c86577cb32 100644 --- a/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index c7143f7d5c..0b23c8ae91 100644 --- a/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 7c10c34523..9bab346508 100644 --- a/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index 68c5760734..e85fe3c401 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 3787a8eaf1..574d72b208 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 7bbb1fb8f0..e551c01892 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index e93e734e66..be073937ab 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index e4996d459a..fc2ccc55e4 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index 0bcfef72ab..b8a4976fa2 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 83b9d915f4..b87ca0c52f 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index 9bbd1478a9..4fb95e1ee2 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 9f41f357ca..c289249a4e 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 0d8a8d7e62..53816a3797 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 26470d0556..d164766734 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index 6fa872212a..990668f32a 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/intrinsics/probes/mma_m16n8k16_s32_u8.ll index 1b7a2ab311..e0ba1958a7 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index e1d99bcc83..6f912c40b8 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index 561ca82bce..8cc53b0c99 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index b06c35609f..2216dec559 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index 89a369f4e3..f520f62a66 100644 --- a/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 715d24a899..5e73e8af7e 100644 --- a/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index e46e29201b..098615515b 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index b02d641a4e..3616aac825 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 671c967568..571ff839f0 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index 16314882c7..8304ccf7de 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 78fc634a97..fdf581e6cc 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index 60a7744103..9dcd29a2ee 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index 3b21922940..b2ed033a7a 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index dcd9c3f56a..7bf79d9f3b 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index 8a15719960..c430821730 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 975bc9ed11..79c57f3b41 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index caf1cf19bd..0e2ace763e 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 97a7ad5840..204ee5ccec 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index 04bc4acf53..064c0ffd12 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index 72f90eed95..8acc89885a 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index 202ca3ce08..9ec58d8919 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 2038d886f5..c4c62855b1 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index defcac001d..4f875429f7 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index cd73de8eb0..835a429bb6 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 4431e4d8a5..8459158adb 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index aac88c2777..c545b1d1fa 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index 56d53ae963..07ed057b55 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index 94f34125c8..0605730a59 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index cdd7e57398..7abfbec163 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index 1ad6123d34..e95c0856b0 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index a23db44a2c..8fed56158d 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index d13ede0396..09054c50a4 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index af6ae69ee0..6cab261088 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index b9fb90a7cf..7bc5ce17e5 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index b65a5e1edf..c493f30012 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index d8ae5eb98d..6061a00911 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index f76a320e85..c4b1cabf8d 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index e186591fc6..af2f3cc63e 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index 06fb18bdf6..e5d2b9d30d 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index 6330da3521..db26c79aa2 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index 90ac832473..4bdfb2428b 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index 89a311470b..18cf3bb51d 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index fc423843d7..0b28041ac7 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index 62adbab13e..59bdd8347d 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index 3c95f08757..bca712e6cc 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index 595955a2bf..a6b05e9c05 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index fe76080066..b45f95d479 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 355be7c4d3..3df6831fa4 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index 0ed0557252..b7d4b61bcb 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index 726e27d404..14f063c32f 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index a20d6e3566..073f678491 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index a05abbbcfe..181acc5b5b 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index e4227f1110..bd7059afc2 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index f3670b096d..b8f630a689 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index f26c6f95fe..5496a911fc 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index e97e238d87..74d3978a13 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index 02f46cc6bd..7fa24f436b 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index 0704365d94..df698df4b4 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 7fece7f172..ef80449d4a 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 76a03fbe61..567183e2b6 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index c17965b9e1..a72aef48d4 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index ec293d8e6c..61fbc28f59 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index fcfe25612f..7e6ef81683 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index d5222ca264..26f193194f 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index fcf989dcdb..3cd529f911 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index 497b1c61a8..7becba9d7f 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 201c6b4011..b3940cc7b1 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index a781318992..69892b755d 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index 2dcbc0291f..aea7e99702 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index 9ab7304639..c5d702c5e4 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 6d42709e72..c893455273 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 07fe418d71..d3ee34f967 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index 99db7845c6..e4dc9e2f1d 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index f0162e7b5b..f67c37660d 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index 8bc029fb6d..ec4c902fed 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index c814062020..6a31a53076 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index 48ea460094..f642d41251 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index 7177de07ab..58eccc06e1 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 2e3f5b56c2..e995523db7 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index 1baca9d509..4094119f1e 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index e373189c4d..975656f462 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index bb9b4d535d..fc5911adcc 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 9f8922664f..809d52e7e3 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index 83225dc35d..7320e8eeab 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index b895e7eaf8..057e84fdd9 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 87a1a68e0a..13e1f60aac 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index 6e971e4550..a47326c59a 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index de42fd30e4..96d8af7715 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index 4f37af1c1e..d3131adb02 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/intrinsics/probes/mma_m16n8k32_s32_s4.ll index d98377829a..cf15d993e7 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index 1547583cf9..6eda2fc416 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index 887e048030..477e8feeec 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 8f9a007242..f00d815412 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_m16n8k32_s32_s8.ll index 4870c39f1b..cac146e2a9 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index cfe37b5c89..b05090af90 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index 3925250b55..bd875d14d0 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index 428e6c42e2..81141bc94d 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/intrinsics/probes/mma_m16n8k32_s32_u4.ll index eedaf28c42..c5384ffe47 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index f5ecd80ef9..810747ceb3 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 9336a6d1f4..d1cdede423 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index 4dae4a3d74..391a649b1d 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_m16n8k32_s32_u8.ll index d0cd3714f7..2c3a23206d 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index f3ca3a052d..4629b615d3 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index 7af581e38b..27a084def1 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index 35145ad43d..8dc4bdaca9 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index 1bee661ae6..0b2bdc363c 100644 --- a/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_m16n8k64_s32_s4.ll index d0d6079f1f..46808cdb48 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index b22126cd85..d252b98e99 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index 39a6330ed7..60954632b2 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index e098daa90a..5a9e279780 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_m16n8k64_s32_u4.ll index f408f9bcf9..f9d108e80d 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index c233aa5d2f..3221a6eb2c 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index bfec98968c..e73fd73885 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index a5e0eb5719..fd2b3aee26 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/intrinsics/probes/mma_m16n8k8_f16_f16.ll index d71f481be1..42e8033462 100644 --- a/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index 0b66faadfd..fa786b8e10 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 2770e8712f..70b9c3f787 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index 4c3c6fccff..d7c79d23c1 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index e66d1165dd..0832d94e0a 100644 --- a/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 7a5cb727fe..5e39dff15e 100644 --- a/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/intrinsics/probes/mma_m8n8k16_s32_s8.ll index d6d1b5fa40..905ad9b518 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index feffa7e53e..60722a3c32 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 63488fee56..5ad415c196 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 72786dd844..689fe0ec56 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/intrinsics/probes/mma_m8n8k16_s32_u8.ll index 2a2d346a55..0277ce64a8 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index 2c10dac69b..0601049388 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index fa86ab8e8c..6833e7bd97 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index 6c6359518c..03277ec1c2 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/intrinsics/probes/mma_m8n8k32_s32_s4.ll index ab00bf2115..5ecd1f4f37 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index e7d49c1bef..af4c53dee8 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index 33d8a273c5..fd700c0822 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 01ce04f226..4ef04b7b4f 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/intrinsics/probes/mma_m8n8k32_s32_u4.ll index ced502c6b4..de2a04923b 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 895b9dfa45..bbb69b992c 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index 9aa33a996e..308b10ec1c 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index bf67485621..28f662eb09 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k4_f64.ll b/intrinsics/probes/mma_m8n8k4_f64.ll index dbf88178e5..0209f70277 100644 --- a/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index d7f31a77d6..67fa3b520a 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index 41d695ffe7..52a4bc35b6 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index b6e0ba0453..81b05f2c60 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index dd0f05e125..8e067221f5 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 537f839550..9dc60706cc 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index e5b354e0f4..37b798dbc3 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index a126f1005f..67203ead5e 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index 3e8feac629..05e817fdaf 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index d5c6f74a7f..6c8ed5edfb 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 2f4d672a76..cc1e9f3265 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index a7a9ea84d3..07e6fdf0cb 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index 7f8ea390a9..241eaf63b5 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index 9be435913b..d2c09779f4 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index db618bc374..1e029cf475 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index 2916d72356..b942d0eb18 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index 71f8ad3386..b8b8a4f803 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 0850246521..d7da87ac6d 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index 099f62af1d..e686523de0 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index c3f9e73766..0051364e72 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 2c966709c8..59f1cec3f4 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 964db6ffe5..d955509f3d 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index e9b7c2712a..a753506f65 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index 08e9934829..dd4f250277 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index 5aafcbe2c4..f5fc3fc101 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index b30f068331..b99948442d 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index 9da665295b..b87a163bf5 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index 71bb98fcab..fb5b7a2d68 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index d0636ec068..ff2fe8e14b 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index a4d698320b..dc5072303c 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index e7dee2e3e9..447ada4b8c 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index 483be82f9a..023b390ae1 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index a13a90508c..8b55260be3 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index 5227f4c146..01a019d115 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 31788fae16..98e990a51d 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 1f6165d13c..d13f2ad98b 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 1bf74ed3fb..8583a54742 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index d2c54f1295..28a816a03b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index e0bcd14872..221acd0463 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index 4ae9f848cb..3db768ef01 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index 5b5bbe3323..6d29c8f92d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 28870da031..68c925282a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 3495c366ea..e5e43c82fe 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index 6b0a05e818..9276805007 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index 25707ae659..d702a6981f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index 31d5a18c92..943acc557e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index f9f7ffa76d..e8d167f8fc 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index 5059228b34..9a31b67675 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index 12ce6f2a9a..852f5dd697 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index bff33f80c7..6d41855ae2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index 04ba90da8b..3e6fd9db2d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index fadc501d15..6b1ea0a9ac 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index 1ce1ad0e48..8b571ba285 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index c86c1ff5bd..43947d5faa 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index 439a394053..fdcad602fe 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index c99c3e53e8..3c12729bcb 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index f72f045c70..425fad628b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index cf8d239357..2ac0551a98 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index 4b2f0007bd..044e551b9b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 65679c06d1..91c0d7004e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 63e92f2a0e..7e4b30f52f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index f1857a7789..96e7a9a260 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 667fec5619..d2864517c6 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index 6a51ca130e..5ee91c99d3 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 2996ed8bfa..f603738cf2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 148beeaffe..964e772ac1 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 624fe97fe9..a59a7db76f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 77dde641da..ad2376189d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index 1efd307b51..43ab935132 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index ebf30bb152..73e5075361 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index be386d3014..13001a470b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index 8d28298011..895188d522 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 0a70815a88..c12108fea1 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 12d3322d7d..92afaa040f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index 5972bab465..c696348ec9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index 05bedd4ca7..480cd52535 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index 3c852f3458..b9076d83e5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index ad7aca5d6b..65d2761bca 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 5651f7057e..60baff636e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index 24b662e07c..30704ed82a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index b59f4b8ae5..c53f3e834a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index fc7f857d6f..ad9dd1fa7b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 2900a5a529..2e42080a19 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index 3aea58be20..a38ff01acc 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index ee13db537e..9ef5e03cab 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index 417b7daf67..1bd31c335e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index 15d50059d0..12eb37d5b0 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index ce60de9013..b8216d486a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 383c88c27d..5b8756bc71 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 4a78e54b7c..02b5cd7e41 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 7df3e1f8ec..0fc46ba488 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 7d0036c102..34e97b8888 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 1858a99fbe..a775a6c5f5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 2eb2690023..75743016d5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index dabdf1f912..0e9c5f8d07 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index db87bbc8d0..1b753dc6f7 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index 6a2747a9f2..dfd498d093 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index a520bf42fd..77c81cda93 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index d91345bf6d..3a4df46cdd 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 6a8a68f043..7e0df0befa 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index badc172fb8..f3420f05c4 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index d3ee678420..e8625886f9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index 3612d40921..148cf002b0 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 30048c40d0..18f45dbf2e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index 9d3a909982..42440060c9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index 4c81d8736c..45aa9bb55b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index b921dbc6b4..256803ae0e 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index 6cdbed3935..498b6e45f9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index a638854bab..cb585101ac 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 8b7d7805ca..12fac367b6 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index 1e3f034f45..ce586d3225 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index 1ebbafd176..9bf22d519c 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index 447bdcf885..ad12c53c50 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 39880d919a..25be2ba56d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 29e49dfb6a..1c81e6fd47 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index e2ca8eaa43..2fe5542408 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 24e55f62fb..a25d2e6089 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 7af0218094..2c1bdaa0cf 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index 0deadc777a..78c7c5b2c2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index b444e489a5..ca5b6616b2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index ee14d0a4f8..03cf05c3a8 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index c79a4f8547..9f89beeae8 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index 63ce3fa1c1..03a2fd1c03 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index f67d04342e..d671222c5d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index 7450c5e86b..262bc62934 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index c56e48c5f5..0bc8beafdc 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index fbbfec4ad9..9a040a62a2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/movmatrix_trans_b16.ll b/intrinsics/probes/movmatrix_trans_b16.ll index eb61aefd8c..c68b6c0237 100644 --- a/intrinsics/probes/movmatrix_trans_b16.ll +++ b/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_bf16x2.ll b/intrinsics/probes/mul_bf16x2.ll index c1aa9ab7fc..5f5f8c7e96 100644 --- a/intrinsics/probes/mul_bf16x2.ll +++ b/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_f16x2.ll b/intrinsics/probes/mul_f16x2.ll index 54d9358d3f..6cbebfc123 100644 --- a/intrinsics/probes/mul_f16x2.ll +++ b/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_f32x2.ll b/intrinsics/probes/mul_f32x2.ll index 6ba7f1a2cf..5a82e6f1fa 100644 --- a/intrinsics/probes/mul_f32x2.ll +++ b/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_ftz_f32x2.ll b/intrinsics/probes/mul_ftz_f32x2.ll index 92527a58e0..8c54c72e8a 100644 --- a/intrinsics/probes/mul_ftz_f32x2.ll +++ b/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_f32.ll b/intrinsics/probes/mul_rm_f32.ll index 2f85919620..9a87d2ca87 100644 --- a/intrinsics/probes/mul_rm_f32.ll +++ b/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_f64.ll b/intrinsics/probes/mul_rm_f64.ll index d30395c092..435a32beda 100644 --- a/intrinsics/probes/mul_rm_f64.ll +++ b/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_ftz_f32.ll b/intrinsics/probes/mul_rm_ftz_f32.ll index 1f81ed3f58..d627e91175 100644 --- a/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_f32.ll b/intrinsics/probes/mul_rn_f32.ll index 10edac5a56..12b775c8f3 100644 --- a/intrinsics/probes/mul_rn_f32.ll +++ b/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_f64.ll b/intrinsics/probes/mul_rn_f64.ll index 699ba1b364..cff1b6594a 100644 --- a/intrinsics/probes/mul_rn_f64.ll +++ b/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_ftz_f32.ll b/intrinsics/probes/mul_rn_ftz_f32.ll index 14b830cd6c..4846ebd3e5 100644 --- a/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_f32.ll b/intrinsics/probes/mul_rp_f32.ll index 6d4c8626fa..2b1c2ee416 100644 --- a/intrinsics/probes/mul_rp_f32.ll +++ b/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_f64.ll b/intrinsics/probes/mul_rp_f64.ll index 01943608a3..746cf7bb22 100644 --- a/intrinsics/probes/mul_rp_f64.ll +++ b/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_ftz_f32.ll b/intrinsics/probes/mul_rp_ftz_f32.ll index 88058a3757..bdcfc3409f 100644 --- a/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_f32.ll b/intrinsics/probes/mul_rz_f32.ll index 7094191e33..ef17f80856 100644 --- a/intrinsics/probes/mul_rz_f32.ll +++ b/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_f64.ll b/intrinsics/probes/mul_rz_f64.ll index e4c450ed82..068b6fe9d1 100644 --- a/intrinsics/probes/mul_rz_f64.ll +++ b/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_ftz_f32.ll b/intrinsics/probes/mul_rz_ftz_f32.ll index 80f9921d36..bf54c87576 100644 --- a/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nanosleep.ll b/intrinsics/probes/nanosleep.ll index 0e1169bca3..c33ba1ebfd 100644 --- a/intrinsics/probes/nanosleep.ll +++ b/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/neg_bf16x2.ll b/intrinsics/probes/neg_bf16x2.ll index c1b3631c78..4e244a3ac8 100644 --- a/intrinsics/probes/neg_bf16x2.ll +++ b/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/neg_f16x2.ll b/intrinsics/probes/neg_f16x2.ll index f80d43a2cc..2cca991b68 100644 --- a/intrinsics/probes/neg_f16x2.ll +++ b/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nsmid.lib_nvvm.ll b/intrinsics/probes/nsmid.lib_nvvm.ll index 7721ebafad..2ab2bb0c24 100644 --- a/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nsmid.llvm_nvptx.ll b/intrinsics/probes/nsmid.llvm_nvptx.ll index 4e45521e75..2f6e980026 100644 --- a/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nwarpid.lib_nvvm.ll b/intrinsics/probes/nwarpid.lib_nvvm.ll index 3e3bb08f14..e769a42f6c 100644 --- a/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nwarpid.llvm_nvptx.ll b/intrinsics/probes/nwarpid.llvm_nvptx.ll index 05c381ad9c..a58b7e8502 100644 --- a/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/packed_atomic_add_bf16x2.ll b/intrinsics/probes/packed_atomic_add_bf16x2.ll index 8ce88380d6..4541ec57e5 100644 --- a/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/packed_atomic_add_f16x2.ll b/intrinsics/probes/packed_atomic_add_f16x2.ll index d45b9003c9..7b1afc08e9 100644 --- a/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/pmevent.ll b/intrinsics/probes/pmevent.ll index 661b4b1ae9..6cafcc5fca 100644 --- a/intrinsics/probes/pmevent.ll +++ b/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prefetch_tma_descriptor.ll b/intrinsics/probes/prefetch_tma_descriptor.ll index 06c1ccb4b4..a43be39f1c 100644 --- a/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt.ll b/intrinsics/probes/prmt.ll index 8d363b04c1..a23d134535 100644 --- a/intrinsics/probes/prmt.ll +++ b/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_b4e.ll b/intrinsics/probes/prmt_b4e.ll index fdd4f96867..7934f62e3c 100644 --- a/intrinsics/probes/prmt_b4e.ll +++ b/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_ecl.ll b/intrinsics/probes/prmt_ecl.ll index d9f3640d85..25016a9c40 100644 --- a/intrinsics/probes/prmt_ecl.ll +++ b/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_ecr.ll b/intrinsics/probes/prmt_ecr.ll index 8f22bef597..82f79c13e1 100644 --- a/intrinsics/probes/prmt_ecr.ll +++ b/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_f4e.ll b/intrinsics/probes/prmt_f4e.ll index 45e26de539..94a80d925e 100644 --- a/intrinsics/probes/prmt_f4e.ll +++ b/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_rc16.ll b/intrinsics/probes/prmt_rc16.ll index 16b565c1d1..3bca678215 100644 --- a/intrinsics/probes/prmt_rc16.ll +++ b/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_rc8.ll b/intrinsics/probes/prmt_rc8.ll index 1fde8d5fe8..d5add80dda 100644 --- a/intrinsics/probes/prmt_rc8.ll +++ b/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_approx_ftz_f32.ll b/intrinsics/probes/rcp_approx_ftz_f32.ll index fef0dba0bd..6fd48f0c4c 100644 --- a/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_approx_ftz_f64.ll b/intrinsics/probes/rcp_approx_ftz_f64.ll index 105ef4a2e5..ab33775ffb 100644 --- a/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_f32.ll b/intrinsics/probes/rcp_rm_f32.ll index b441204b53..76d8c5beed 100644 --- a/intrinsics/probes/rcp_rm_f32.ll +++ b/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_f64.ll b/intrinsics/probes/rcp_rm_f64.ll index cf4b693faa..1899b9d026 100644 --- a/intrinsics/probes/rcp_rm_f64.ll +++ b/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_ftz_f32.ll b/intrinsics/probes/rcp_rm_ftz_f32.ll index d389565526..4c0802abd3 100644 --- a/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_f32.ll b/intrinsics/probes/rcp_rn_f32.ll index d6a387a560..3d5aad5f1e 100644 --- a/intrinsics/probes/rcp_rn_f32.ll +++ b/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_f64.ll b/intrinsics/probes/rcp_rn_f64.ll index da02ad2519..e2806b142f 100644 --- a/intrinsics/probes/rcp_rn_f64.ll +++ b/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_ftz_f32.ll b/intrinsics/probes/rcp_rn_ftz_f32.ll index 4d25aa11fe..820315b253 100644 --- a/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_f32.ll b/intrinsics/probes/rcp_rp_f32.ll index bebb01b8b7..b6ed0d090b 100644 --- a/intrinsics/probes/rcp_rp_f32.ll +++ b/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_f64.ll b/intrinsics/probes/rcp_rp_f64.ll index 61ff35afaf..c5de452c7e 100644 --- a/intrinsics/probes/rcp_rp_f64.ll +++ b/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_ftz_f32.ll b/intrinsics/probes/rcp_rp_ftz_f32.ll index 0263af7acd..509abc3e97 100644 --- a/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_f32.ll b/intrinsics/probes/rcp_rz_f32.ll index 923614eafa..938e3a3bf1 100644 --- a/intrinsics/probes/rcp_rz_f32.ll +++ b/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_f64.ll b/intrinsics/probes/rcp_rz_f64.ll index 607567956e..055d595866 100644 --- a/intrinsics/probes/rcp_rz_f64.ll +++ b/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_ftz_f32.ll b/intrinsics/probes/rcp_rz_ftz_f32.ll index 52af228886..a6dda5582f 100644 --- a/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_add.ll b/intrinsics/probes/redux_sync_add.ll index 6fa409fedf..f1f718ca72 100644 --- a/intrinsics/probes/redux_sync_add.ll +++ b/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_and.ll b/intrinsics/probes/redux_sync_and.ll index f636eaa949..ce3bb97bb9 100644 --- a/intrinsics/probes/redux_sync_and.ll +++ b/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_abs_f32.ll b/intrinsics/probes/redux_sync_max_abs_f32.ll index ed3d2b2b6b..6797ad8e60 100644 --- a/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 86c46f8c1b..502c1a0700 100644 --- a/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_f32.ll b/intrinsics/probes/redux_sync_max_f32.ll index 7df4271426..dc16847b6c 100644 --- a/intrinsics/probes/redux_sync_max_f32.ll +++ b/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_i32.ll b/intrinsics/probes/redux_sync_max_i32.ll index ffe9bdf596..99f3107c74 100644 --- a/intrinsics/probes/redux_sync_max_i32.ll +++ b/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_nan_f32.ll b/intrinsics/probes/redux_sync_max_nan_f32.ll index 1e8d732997..fdb90d8c4f 100644 --- a/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_u32.ll b/intrinsics/probes/redux_sync_max_u32.ll index 94c0e36224..339b613b11 100644 --- a/intrinsics/probes/redux_sync_max_u32.ll +++ b/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_abs_f32.ll b/intrinsics/probes/redux_sync_min_abs_f32.ll index acd58ca578..23612855f6 100644 --- a/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 024477177d..30a237e98c 100644 --- a/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_f32.ll b/intrinsics/probes/redux_sync_min_f32.ll index dadb7ee5ba..7272a8c45b 100644 --- a/intrinsics/probes/redux_sync_min_f32.ll +++ b/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_i32.ll b/intrinsics/probes/redux_sync_min_i32.ll index 1b68195374..b172a4cb82 100644 --- a/intrinsics/probes/redux_sync_min_i32.ll +++ b/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_nan_f32.ll b/intrinsics/probes/redux_sync_min_nan_f32.ll index 7acfed2132..35ae766d30 100644 --- a/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_u32.ll b/intrinsics/probes/redux_sync_min_u32.ll index 7ff8b78ae2..9d0cedbf58 100644 --- a/intrinsics/probes/redux_sync_min_u32.ll +++ b/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_or.ll b/intrinsics/probes/redux_sync_or.ll index e89a02af0a..77ca86f057 100644 --- a/intrinsics/probes/redux_sync_or.ll +++ b/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_xor.ll b/intrinsics/probes/redux_sync_xor.ll index 84b947e298..84eab884e6 100644 --- a/intrinsics/probes/redux_sync_xor.ll +++ b/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_f32.ll b/intrinsics/probes/rsqrt_approx_f32.ll index de5e025417..57e2661ca3 100644 --- a/intrinsics/probes/rsqrt_approx_f32.ll +++ b/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_f64.ll b/intrinsics/probes/rsqrt_approx_f64.ll index c381bb0c12..ae7dcf7bb9 100644 --- a/intrinsics/probes/rsqrt_approx_f64.ll +++ b/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/intrinsics/probes/rsqrt_approx_ftz_f32.ll index d798d5a19f..c6f5633403 100644 --- a/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/setmaxnreg_dec.ll b/intrinsics/probes/setmaxnreg_dec.ll index af37b39561..4b085545e7 100644 --- a/intrinsics/probes/setmaxnreg_dec.ll +++ b/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/setmaxnreg_inc.ll b/intrinsics/probes/setmaxnreg_inc.ll index 6a5a12f145..618d9a0731 100644 --- a/intrinsics/probes/setmaxnreg_inc.ll +++ b/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_f32_sync.ll b/intrinsics/probes/shuffle_down_f32_sync.ll index f0cc07b049..ac606b5fd9 100644 --- a/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_sync.ll b/intrinsics/probes/shuffle_down_sync.ll index 0c190bb994..120fbc88fd 100644 --- a/intrinsics/probes/shuffle_down_sync.ll +++ b/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_u64_sync.ll b/intrinsics/probes/shuffle_down_u64_sync.ll index 72e66c3ce4..1b9a4f72da 100644 --- a/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_f32_sync.ll b/intrinsics/probes/shuffle_f32_sync.ll index 4e49b2feb7..abd8ad8dde 100644 --- a/intrinsics/probes/shuffle_f32_sync.ll +++ b/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_sync.ll b/intrinsics/probes/shuffle_sync.ll index 42ecc23be7..0405b5121e 100644 --- a/intrinsics/probes/shuffle_sync.ll +++ b/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_u64_sync.ll b/intrinsics/probes/shuffle_u64_sync.ll index fe100ff054..48189f44e7 100644 --- a/intrinsics/probes/shuffle_u64_sync.ll +++ b/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_f32_sync.ll b/intrinsics/probes/shuffle_up_f32_sync.ll index 38729f4ba6..3a6833c8ce 100644 --- a/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_sync.ll b/intrinsics/probes/shuffle_up_sync.ll index 9e09f6f83a..3484863f62 100644 --- a/intrinsics/probes/shuffle_up_sync.ll +++ b/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_u64_sync.ll b/intrinsics/probes/shuffle_up_u64_sync.ll index 601928804b..76b300e500 100644 --- a/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_f32_sync.ll b/intrinsics/probes/shuffle_xor_f32_sync.ll index 95da54f491..453d8abcc8 100644 --- a/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_sync.ll b/intrinsics/probes/shuffle_xor_sync.ll index 4efa1e64ca..57e7230cff 100644 --- a/intrinsics/probes/shuffle_xor_sync.ll +++ b/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_u64_sync.ll b/intrinsics/probes/shuffle_xor_u64_sync.ll index e6862cc95c..790af42d3f 100644 --- a/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sin_approx_f32.ll b/intrinsics/probes/sin_approx_f32.ll index cb89352f47..c96f3ff69b 100644 --- a/intrinsics/probes/sin_approx_f32.ll +++ b/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sin_approx_ftz_f32.ll b/intrinsics/probes/sin_approx_ftz_f32.ll index b1fd971f5c..de90d16ef0 100644 --- a/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/smid.lib_nvvm.ll b/intrinsics/probes/smid.lib_nvvm.ll index a73644934d..340e4d9954 100644 --- a/intrinsics/probes/smid.lib_nvvm.ll +++ b/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/smid.llvm_nvptx.ll b/intrinsics/probes/smid.llvm_nvptx.ll index eb1d182e2d..6d21c82707 100644 --- a/intrinsics/probes/smid.llvm_nvptx.ll +++ b/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_approx_f32.ll b/intrinsics/probes/sqrt_approx_f32.ll index 72ff6b20c3..1c010bbd92 100644 --- a/intrinsics/probes/sqrt_approx_f32.ll +++ b/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_approx_ftz_f32.ll b/intrinsics/probes/sqrt_approx_ftz_f32.ll index 280aab1b93..4e4b3cf52a 100644 --- a/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_f32.ll b/intrinsics/probes/sqrt_rm_f32.ll index 8f985c9cd7..0a37b17acc 100644 --- a/intrinsics/probes/sqrt_rm_f32.ll +++ b/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_f64.ll b/intrinsics/probes/sqrt_rm_f64.ll index a83ca9bc74..b8ae4364ba 100644 --- a/intrinsics/probes/sqrt_rm_f64.ll +++ b/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_ftz_f32.ll b/intrinsics/probes/sqrt_rm_ftz_f32.ll index 15e453699e..b4eece6ca3 100644 --- a/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_f32.ll b/intrinsics/probes/sqrt_rn_f32.ll index 31fee94478..ae60fbfcc6 100644 --- a/intrinsics/probes/sqrt_rn_f32.ll +++ b/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_f64.ll b/intrinsics/probes/sqrt_rn_f64.ll index e9a3f1ffbe..ab8f956e7a 100644 --- a/intrinsics/probes/sqrt_rn_f64.ll +++ b/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_ftz_f32.ll b/intrinsics/probes/sqrt_rn_ftz_f32.ll index 3777543af4..189b8fb88d 100644 --- a/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_f32.ll b/intrinsics/probes/sqrt_rp_f32.ll index 82037c4860..6e30d42892 100644 --- a/intrinsics/probes/sqrt_rp_f32.ll +++ b/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_f64.ll b/intrinsics/probes/sqrt_rp_f64.ll index 9604d98682..c311ab7d74 100644 --- a/intrinsics/probes/sqrt_rp_f64.ll +++ b/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_ftz_f32.ll b/intrinsics/probes/sqrt_rp_ftz_f32.ll index 75ac3ae913..91ae91876f 100644 --- a/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_f32.ll b/intrinsics/probes/sqrt_rz_f32.ll index 367661169b..6e075115a9 100644 --- a/intrinsics/probes/sqrt_rz_f32.ll +++ b/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_f64.ll b/intrinsics/probes/sqrt_rz_f64.ll index b1b24314cc..1a64decad5 100644 --- a/intrinsics/probes/sqrt_rz_f64.ll +++ b/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_ftz_f32.ll b/intrinsics/probes/sqrt_rz_ftz_f32.ll index 01548379a0..c35cac71fe 100644 --- a/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 0d8744f163..4b01cfac10 100644 --- a/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index cd059aa51f..ff4f377fb0 100644 --- a/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 972f13e200..cdb8ead4ed 100644 --- a/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 79566adecb..a64f6a1d19 100644 --- a/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_bf16x2.ll b/intrinsics/probes/sub_bf16x2.ll index 02007ab138..5fd5853526 100644 --- a/intrinsics/probes/sub_bf16x2.ll +++ b/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_f16x2.ll b/intrinsics/probes/sub_f16x2.ll index ee084d0d9d..7e76597578 100644 --- a/intrinsics/probes/sub_f16x2.ll +++ b/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_f32x2.ll b/intrinsics/probes/sub_f32x2.ll index c4dbe9a3d3..d63be57e7f 100644 --- a/intrinsics/probes/sub_f32x2.ll +++ b/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_ftz_f32x2.ll b/intrinsics/probes/sub_ftz_f32x2.ll index 9043eb52e7..a4030b6bca 100644 --- a/intrinsics/probes/sub_ftz_f32x2.ll +++ b/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sync_mask.ll b/intrinsics/probes/sync_mask.ll index ee16870a76..0d3fb736de 100644 --- a/intrinsics/probes/sync_mask.ll +++ b/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sync_threads.ll b/intrinsics/probes/sync_threads.ll index a8130be45f..10de49de4d 100644 --- a/intrinsics/probes/sync_threads.ll +++ b/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tanh_approx_f32.ll b/intrinsics/probes/tanh_approx_f32.ll index e3e34ae1df..421f5b0832 100644 --- a/intrinsics/probes/tanh_approx_f32.ll +++ b/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_alloc.ll b/intrinsics/probes/tcgen05_alloc.ll index 6acfe45a1c..a3cf8c1b46 100644 --- a/intrinsics/probes/tcgen05_alloc.ll +++ b/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_alloc_cg2.ll b/intrinsics/probes/tcgen05_alloc_cg2.ll index dc0cbca220..73204db234 100644 --- a/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit.ll b/intrinsics/probes/tcgen05_commit.ll index beae2ad8f3..ec22dfacbd 100644 --- a/intrinsics/probes/tcgen05_commit.ll +++ b/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_cg2.ll b/intrinsics/probes/tcgen05_commit_cg2.ll index 023e507d30..6338f7728d 100644 --- a/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_multicast.ll b/intrinsics/probes/tcgen05_commit_multicast.ll index 699a249e90..0ce6cee933 100644 --- a/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index 4772d30167..c25918ca72 100644 --- a/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/intrinsics/probes/tcgen05_commit_shared_cluster.ll index acb1289973..fda0a5064c 100644 --- a/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index a1433b90ed..5c3c3eb161 100644 --- a/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b.ll b/intrinsics/probes/tcgen05_cp_128x128b.ll index 874d863547..c8d38f7017 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index f4e3048836..70183e921e 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index 718a763a12..7cac6e4a52 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index 2d06edfdf3..6531d83e45 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 035c14ab55..a27ad8d76f 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 57f39a88a9..a180c53135 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index be74fe61f7..c2951bc4d2 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index c011f0964b..0150bbf16f 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index d129867b93..c424095709 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 91b0ad59d5..246ea357cb 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index 388a3a637d..9b1d673b22 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index e3d340db5d..ca02866827 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index 544639637e..4ddd78659b 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index b88e27ade7..165a3ea7a7 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index 0f61ebfa29..c4d0ea878a 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index 75716157c3..922f138ad5 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b.ll b/intrinsics/probes/tcgen05_cp_4x256b.ll index d0477d394b..549e41ea22 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index bc8bcb758c..3893dcd7ee 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index b5008cc128..55f59e5cab 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 2404cfa452..0466a07651 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index 6346a9b18c..2a4989af12 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 0823b74275..12c967bfc2 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 851638d6ee..671c60d30a 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 6154c24fd5..5d674158c5 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index 9c81cd3854..7f8df44926 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index e6232c72ef..72dcba4181 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index 04eb00bd45..b72e23a3fb 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 913e0c9653..7beb8eb19d 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index cd6b4548c8..1f9f60aaa6 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index ca2f614a30..065e6ed67a 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index ea163ebf31..62011f1962 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 900231f06d..5212aad6a9 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index ee2c9aa47a..8675542d22 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 4b53f40f9a..87b6d21ede 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index f2927ff61e..0b7da8de6e 100644 --- a/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index 2c6d574821..4dcfcd2006 100644 --- a/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_dealloc.ll b/intrinsics/probes/tcgen05_dealloc.ll index 7bae53ecbc..d9e2a1c72e 100644 --- a/intrinsics/probes/tcgen05_dealloc.ll +++ b/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_dealloc_cg2.ll b/intrinsics/probes/tcgen05_dealloc_cg2.ll index 9c5c87deee..a1b5e6b72b 100644 --- a/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 4d99f51ecf..057134ea9a 100644 --- a/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index 2e85368df9..35c228e758 100644 --- a/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index 66eea7805c..5b7e83ed80 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index ebcd6d5324..39d5e63e6e 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index 122e4ea3af..302d196567 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index ed09506f35..e952ee063f 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index 2c16e905b2..24b8a328c5 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index cb6e7fafc3..e08ae53f02 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index 38733d57e4..c03fb1a3d4 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index 65fd5cc7b5..373664c92a 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index 0676b6a0f6..61d1025d3c 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index e40101f924..fa33041cb1 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index dedb0a0aa8..54ccfa859d 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index cc8df14d6f..505c9f6613 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 3fd33fb9fb..d4473692f8 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index ed89859528..3e2841ddc4 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index a54c69118f..4a9b5dcd2f 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index ea56b1fcc8..ea7315f3a3 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index 9a97fcdb32..0a86080491 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index e0edac159d..653e5ed2b2 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index e900efdf2c..99bbcbaf3b 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index 77aadce51d..7c6dd90a6d 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 8c31fbdd5f..e8d74ee084 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index 9e62b6aa48..a0faa997fb 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 62e4a4c350..d8e445b88e 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 26d54c0e9b..cf57e575cd 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 89f8e40fb3..5edaacc53b 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 6a77ff0120..9e0e5def2f 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index d50558f4de..6e67915e11 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 40dbcdecff..fa19e2b7f4 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index 4ffa096247..b5ee44a994 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index 80e486edc9..9361b0e12c 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index 3bc438421d..23e35ef762 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index b525a53143..aa4956d20e 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index ea91a76523..ac4ed61ca9 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 00da06b059..3b7eefb4e6 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index 4c8c9985a6..f67477317a 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 4b3daade89..d4935f5e4d 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index d11833737e..c3e6ff2435 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 0c7fcf9f74..dcc2fcd3ce 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index 03733b0c51..543a1576a9 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 43a58a1463..52f97d0758 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index cbbf017da2..81d7941291 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index d9a4c1e9b7..900fa9fdee 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index dd2114bcc3..4479292b37 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index be835e1829..88edb69cd1 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 1f9edb56a4..dc6c0d06a9 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index f18bbd320d..c203a4e760 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index e29c844fda..5f46cc897c 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 624bc96f51..fe3600dbc8 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index bda45b63ea..ed1f6077ad 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index 8f179c43fb..cda75e3814 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index 0c0110ad7c..8d3cd8d9da 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index 8b9844f114..7ede69949b 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 5f44b02772..05d87244cc 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index d2416f0b00..6da595172a 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index 93e718a265..4a9e137498 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index f1b7f0bfdb..1184874745 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 911d27c9de..c8c9536117 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index b2da613b08..2bd17d1918 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index f803c14f8c..15454b375c 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 41b56f833c..d8ebf008db 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index cf894651a3..7d0fa5b1d3 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index eb6d10e4ee..d48389f770 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 1de6de1955..15992e4a39 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 2a64bffd5b..47982f5d55 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index 2d5f8d1a89..ee093d1034 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index c00badb7f2..5c50bf4c71 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index baaf6a82bd..2a5356b982 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index 73e2290d58..f6b9faff17 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index a32e916295..1c04047d04 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index ea7f80426f..07e81018ef 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index eb7afc20a7..a4de93ca51 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index 114b18dc87..0ea71adc59 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index 78cfc8971e..bdce062d1c 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index 12f42ca913..722108c928 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 8695c8b553..fa96a95e76 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 690bb364d6..f0702f3157 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_load_wait.ll b/intrinsics/probes/tcgen05_load_wait.ll index 2bdc2cd00e..ba92f6600b 100644 --- a/intrinsics/probes/tcgen05_load_wait.ll +++ b/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e2m1.ll b/intrinsics/probes/tcgen05_mma_e2m1.ll index 3d86f3ec30..c0e097214e 100644 --- a/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e2m3.ll b/intrinsics/probes/tcgen05_mma_e2m3.ll index 35aa6aad96..a2ab42e0f2 100644 --- a/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e3m2.ll b/intrinsics/probes/tcgen05_mma_e3m2.ll index efddea9aba..0c885ef137 100644 --- a/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e4m3.ll b/intrinsics/probes/tcgen05_mma_e4m3.ll index a0d6aea99c..3edd9cb837 100644 --- a/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e5m2.ll b/intrinsics/probes/tcgen05_mma_e5m2.ll index 4b442a42ea..486699fd95 100644 --- a/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_f16.ll b/intrinsics/probes/tcgen05_mma_f16.ll index ea9cbec878..4c4ca798fa 100644 --- a/intrinsics/probes/tcgen05_mma_f16.ll +++ b/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 2ff1642e13..70dce3d08d 100644 --- a/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_shared.ll b/intrinsics/probes/tcgen05_mma_shared.ll index bd1a07c0b6..5e23943857 100644 --- a/intrinsics/probes/tcgen05_mma_shared.ll +++ b/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_shared.ll b/intrinsics/probes/tcgen05_mma_sp_shared.ll index 8e9bf92012..7b1f5e2b82 100644 --- a/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/intrinsics/probes/tcgen05_mma_sp_tensor.ll index 6cf4f5d3d3..7f3f2d78c5 100644 --- a/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index 7c7e82d3aa..f2983520d7 100644 --- a/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_tensor.ll b/intrinsics/probes/tcgen05_mma_tensor.ll index 16f6ee0987..b4774bbfd4 100644 --- a/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index 7c81061cd3..84adaa4cb9 100644 --- a/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 7bdec871ab..b3e09d2b3b 100644 --- a/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index 1926738c37..12ba6f21da 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index 6c8f8d352b..52e9d8f81e 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index edc85afaa4..e4e1acfa45 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index d9b5d3ed36..04278209c1 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 1e8ae5c0fc..4ec0eb7772 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_f16.ll b/intrinsics/probes/tcgen05_mma_ws_f16.ll index f1e5272d20..148ae88b1d 100644 --- a/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_shared.ll b/intrinsics/probes/tcgen05_mma_ws_shared.ll index 06651cf6d4..bc51da4659 100644 --- a/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 12b9e3e785..c1e7e42355 100644 --- a/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index 6ac3fab437..273ada5a64 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index bdfd06214a..a279da48c2 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index fda46dbb86..d1753727ef 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index e477013ea4..4a475b391b 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 8c0f0911a5..b95f0f19be 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index d0a1478f1e..77345ac7ef 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/intrinsics/probes/tcgen05_mma_ws_tf32.ll index 1ef72f1863..57823544dd 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index e4d3b3cb80..ca0ec589c2 100644 --- a/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index b018f24a45..1ecf6025d4 100644 --- a/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_shift_down.ll b/intrinsics/probes/tcgen05_shift_down.ll index a21d3b771b..113326f9a6 100644 --- a/intrinsics/probes/tcgen05_shift_down.ll +++ b/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_shift_down_cg2.ll b/intrinsics/probes/tcgen05_shift_down_cg2.ll index 10cb9cc4ae..aa1a9a8b05 100644 --- a/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index 7c429f8205..a5a5566676 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index a4d28ec63b..ff79f733bc 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 04cea28f4e..55b279a4be 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index 89567ebd34..914be4b346 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 6dc84f6ecb..1b4315c620 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 3aa981e65b..ae575010e6 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index 2ab4eca27c..9d8440ef9e 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index e8f73ec97e..a62a417a8d 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index 6991738971..5ad2c47471 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index 95f5fc3501..a1e240ab9f 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index b405eb8847..409d1afec7 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index 9214e24793..49d981d136 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index d6824be5c9..ca3841b8b5 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 9a193796f4..489accee2f 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index fda79237c9..9cc2159607 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 8f9d7122b3..038261228f 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index fdf8be91d4..4147dcc4d7 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 7e9d57b1d4..b48b07852d 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 51f68b428f..15c73baf23 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index 2a22da8dfb..b569e0a95c 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index 78d13c1601..9022813955 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index 4324fd6448..0df0ccf738 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index a02f312eac..e6f73753fb 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 445d595681..da1c6b1eab 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index 179fe1d8ce..21b04e6d52 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index 7f2e27bb63..6aabb84805 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index 94ba07bf66..c746637ee8 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 318fa8dfef..5a4659302c 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index bfc9cad2b2..4ea8e2e908 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index 68278293aa..962947595d 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index dd0ebe55da..b68b1d4e2e 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index c173bd426c..691702a9b1 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index 64704b9451..ab0a4f5699 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index 6c21d82c47..ef3409529d 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index c33a5a2262..eb50b60056 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index 78e1a39bb5..947605cf84 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 96bec07f91..d65bf330fc 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index d613c5f820..fe2676ea6e 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index 4095f9f91f..1f1094988b 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index f69630077d..23363fbf18 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 99b477f92b..b0f4451462 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index b88ab5dfe7..27068db98a 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index 230955bafa..cd50220933 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index 6e9e9b3ee1..71534a1641 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 19baa5f06f..fa1d589397 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 0133bde57f..f9c0fe49fd 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index e4dcd66474..c28d6a3191 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index 6c19cdad41..fff1af03b0 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index ee58d4837d..106c0add03 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 8c2b89be5b..a59d4a66f3 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index 97bac36f86..7151146bde 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 41b64f612c..24a142aa8e 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index 903a071085..7464bc0d6f 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index dd2c2a4782..5103bc81f3 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index f14f075df1..6b447b58c1 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index d93185a9d6..5f52003fb0 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index 5fe775d3cd..9a39f34898 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index a544ecb655..690ea0cf53 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index 60a076fe83..f614d27a96 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index 545930c809..1e217929f6 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index 827771e3d5..76b389f419 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index 1721568cf6..c2a631397d 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 15d598cec6..89d6d02f53 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index a698bdc75d..a54d9fddc7 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index ecedbf8056..2bf5b314f6 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 6a83680da1..f84ef4f5a0 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index beea9becc1..ac4596bf64 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 9b42f9b0a4..57170cdc6e 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index ccec22823a..88fd22d367 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index c8b9ddc19c..d22e5ccf92 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index b717e71d43..29b6e1b02f 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 25ac6dd11b..f0f4c1bd08 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index 97b923d8a6..6be4b3c273 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index b810bab3a5..fe427a081d 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_store_wait.ll b/intrinsics/probes/tcgen05_store_wait.ll index 7726764af4..e50dfe7c4d 100644 --- a/intrinsics/probes/tcgen05_store_wait.ll +++ b/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_box_dim.ll b/intrinsics/probes/tensormap_replace_box_dim.ll index 383215df50..e8fb8752a6 100644 --- a/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_element_stride.ll b/intrinsics/probes/tensormap_replace_element_stride.ll index aa4f704382..6baf9c55e3 100644 --- a/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_element_type.ll b/intrinsics/probes/tensormap_replace_element_type.ll index 10bd454b1a..d843c5a5e6 100644 --- a/intrinsics/probes/tensormap_replace_element_type.ll +++ b/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_fill_mode.ll b/intrinsics/probes/tensormap_replace_fill_mode.ll index 4a4137fb54..d57dc3af80 100644 --- a/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_address.ll b/intrinsics/probes/tensormap_replace_global_address.ll index 4ae2808853..0118898927 100644 --- a/intrinsics/probes/tensormap_replace_global_address.ll +++ b/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_dim.ll b/intrinsics/probes/tensormap_replace_global_dim.ll index 61aa6c25e2..907ee15ec7 100644 --- a/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_stride.ll b/intrinsics/probes/tensormap_replace_global_stride.ll index c0758977bb..abe1fbb081 100644 --- a/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_interleave_layout.ll b/intrinsics/probes/tensormap_replace_interleave_layout.ll index cd114237d2..11b1b29164 100644 --- a/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_rank.ll b/intrinsics/probes/tensormap_replace_rank.ll index fa0adf79fc..f336d97550 100644 --- a/intrinsics/probes/tensormap_replace_rank.ll +++ b/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index dc59ddd5bd..05b8fd7810 100644 --- a/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/intrinsics/probes/tensormap_replace_swizzle_mode.ll index 79573631db..c95da731e8 100644 --- a/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_x.ll b/intrinsics/probes/thread_idx_x.ll index 70f54c724c..a95947dbc2 100644 --- a/intrinsics/probes/thread_idx_x.ll +++ b/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_y.ll b/intrinsics/probes/thread_idx_y.ll index d8623f0ce3..341ac6b8f4 100644 --- a/intrinsics/probes/thread_idx_y.ll +++ b/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_z.ll b/intrinsics/probes/thread_idx_z.ll index 94614f0fda..dafbbf3924 100644 --- a/intrinsics/probes/thread_idx_z.ll +++ b/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence.ll b/intrinsics/probes/threadfence.ll index 4dcb217402..c413393053 100644 --- a/intrinsics/probes/threadfence.ll +++ b/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence_block.ll b/intrinsics/probes/threadfence_block.ll index cbb9417b54..3cdf14ab9f 100644 --- a/intrinsics/probes/threadfence_block.ll +++ b/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence_system.ll b/intrinsics/probes/threadfence_system.ll index c6aa00abc1..6257257a77 100644 --- a/intrinsics/probes/threadfence_system.ll +++ b/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/total_smem_size.lib_nvvm.ll b/intrinsics/probes/total_smem_size.lib_nvvm.ll index 56ddcc2e84..dfa5ac9686 100644 --- a/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 8ac62d912e..71e629812d 100644 --- a/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/trap.ll b/intrinsics/probes/trap.ll index 50e301518a..f69baf301e 100644 --- a/intrinsics/probes/trap.ll +++ b/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/uni_sync.ll b/intrinsics/probes/uni_sync.ll index c90bea50bf..e6f9c086bc 100644 --- a/intrinsics/probes/uni_sync.ll +++ b/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/warpid.lib_nvvm.ll b/intrinsics/probes/warpid.lib_nvvm.ll index 647cb3118a..1f5faebbbb 100644 --- a/intrinsics/probes/warpid.lib_nvvm.ll +++ b/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/warpid.llvm_nvptx.ll b/intrinsics/probes/warpid.llvm_nvptx.ll index a9453df671..675d0b2da8 100644 --- a/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_commit_group.ll b/intrinsics/probes/wgmma_commit_group.ll index e6c502f5d5..269c3e4717 100644 --- a/intrinsics/probes/wgmma_commit_group.ll +++ b/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_fence.ll b/intrinsics/probes/wgmma_fence.ll index fc90766de2..db81d48b8f 100644 --- a/intrinsics/probes/wgmma_fence.ll +++ b/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_wait_group.ll b/intrinsics/probes/wgmma_wait_group.ll index 8691bd31ae..0da1af7f01 100644 --- a/intrinsics/probes/wgmma_wait_group.ll +++ b/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" From 4fd958c6ba8997084c63f62f0e5ae20d34e0d899 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 19:28:06 +0530 Subject: [PATCH 03/17] fix(cuda-intrinsics-gen): gate the CTA-local TMA admission behind shard schema 65 The TMA overlay gained two required evidence-profile fields for the CTA-local G2S variants, so a shard without them stopped parsing and every fixture had to change. Follow the reduction-admission precedent instead: the two fields are optional with a serde default, a shard that lists the CTA-local variants must name both and must carry overlay shard schema 65, and a shard without those variants still parses and expands to the cluster family alone. The canonical-order check accepts either the full operation list or the list without the CTA-local variants. The overlay shard schema ceiling moves from 64 to 65 and tma.toml declares schema 65. Tests cover the schema gate in both directions, the missing-profile error, the CTA-free expansion and a partial CTA-local list. Signed-off-by: nihalpasham --- .../cuda-intrinsics-gen/src/model/overlay.rs | 18 ++++- .../cuda-intrinsics-gen/src/resolve/driver.rs | 4 +- .../src/resolve/families/tma.rs | 36 +++++++-- .../src/resolve/overlay.rs | 12 ++- .../src/resolve/tests/fixtures.rs | 4 +- .../src/resolve/tests/overlay_schema.rs | 11 +++ .../src/resolve/tests/tma_clc_cluster.rs | 75 +++++++++++++++++-- cuda-oxide/intrinsics/overlay/tma.toml | 2 +- 8 files changed, 142 insertions(+), 20 deletions(-) diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs index fbe3bf8c3d..91ac2f9fba 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs @@ -322,8 +322,12 @@ pub struct ClcAdmissionVariant { pub struct TmaAdmission { pub llvm_evidence_profile: String, pub libnvvm_evidence_profile: String, - pub cta_llvm_evidence_profile: String, - pub cta_libnvvm_evidence_profile: String, + /// Evidence profiles for the CTA-local G2S variants. A shard that lists + /// those variants must name both; an older shard without them parses. + #[serde(default)] + pub cta_llvm_evidence_profile: Option, + #[serde(default)] + pub cta_libnvvm_evidence_profile: Option, #[serde(default)] pub reduce_llvm_evidence_profile: Option, #[serde(default)] @@ -343,6 +347,16 @@ pub struct TmaAdmissionVariant { pub operation: TmaOperation, } +impl TmaAdmission { + /// Whether the shard lists the CTA-local G2S variants. Those variants + /// carry their own evidence profiles and need overlay shard schema 65. + pub fn admits_cta_g2s(&self) -> bool { + self.variants + .iter() + .any(|variant| variant.operation.is_cta_g2s()) + } +} + /// One reviewed tensor-reduction operation and its reserved ABI ID. #[derive(Debug, Clone, Deserialize)] #[serde(deny_unknown_fields)] diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/driver.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/driver.rs index 0273760be0..782dfd196f 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/driver.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/driver.rs @@ -243,8 +243,8 @@ pub(crate) fn test_catalog_with_tma(repo_root: &Path) -> Result { let admission = TmaAdmission { llvm_evidence_profile: "llvm-tma-test".into(), libnvvm_evidence_profile: "libnvvm-tma-test".into(), - cta_llvm_evidence_profile: "llvm-tma-test".into(), - cta_libnvvm_evidence_profile: "libnvvm-tma-test".into(), + cta_llvm_evidence_profile: Some("llvm-tma-test".into()), + cta_libnvvm_evidence_profile: Some("libnvvm-tma-test".into()), reduce_llvm_evidence_profile: Some("llvm-tma-reduce-test".into()), reduce_libnvvm_evidence_profile: Some("libnvvm-tma-reduce-test".into()), runtime_validation: RuntimeValidation::Unexecuted, diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs index 95acbb28fd..b782f5003c 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs @@ -1458,19 +1458,35 @@ pub(in crate::resolve) fn expand_tma_admission( ); ensure!( !admission.llvm_evidence_profile.trim().is_empty() - && !admission.libnvvm_evidence_profile.trim().is_empty() - && !admission.cta_llvm_evidence_profile.trim().is_empty() - && !admission.cta_libnvvm_evidence_profile.trim().is_empty(), + && !admission.libnvvm_evidence_profile.trim().is_empty(), "compact TMA admission requires both backend evidence profiles" ); + let admits_cta_g2s = admission.admits_cta_g2s(); + if admits_cta_g2s { + ensure!( + admission + .cta_llvm_evidence_profile + .as_deref() + .is_some_and(|profile| !profile.trim().is_empty()) + && admission + .cta_libnvvm_evidence_profile + .as_deref() + .is_some_and(|profile| !profile.trim().is_empty()), + "compact TMA admission with CTA-local G2S variants requires cta_llvm_evidence_profile and cta_libnvvm_evidence_profile" + ); + } + let expected_operations = TMA_OPERATIONS + .into_iter() + .filter(|operation| admits_cta_g2s || !operation.is_cta_g2s()) + .collect::>(); ensure!( admission .variants .iter() .map(|variant| variant.operation) - .eq(TMA_OPERATIONS), + .eq(expected_operations.iter().copied()), "compact TMA admission must list all {} operations in canonical order", - TMA_OPERATIONS.len() + expected_operations.len() ); let mut records = admission @@ -1485,12 +1501,18 @@ pub(in crate::resolve) fn expand_tma_admission( recipe.abi_id ); let llvm_evidence_profile = if recipe.operation.is_cta_g2s() { - &admission.cta_llvm_evidence_profile + admission + .cta_llvm_evidence_profile + .as_ref() + .expect("validated TMA CTA-local LLVM evidence profile") } else { &admission.llvm_evidence_profile }; let libnvvm_evidence_profile = if recipe.operation.is_cta_g2s() { - &admission.cta_libnvvm_evidence_profile + admission + .cta_libnvvm_evidence_profile + .as_ref() + .expect("validated TMA CTA-local libNVVM evidence profile") } else { &admission.libnvvm_evidence_profile }; diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs index 03a02dbea6..dd4f3dc373 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs @@ -19,7 +19,7 @@ use super::guards::*; pub(super) const OVERLAY_SCHEMA: u32 = 44; pub(super) const MINIMUM_OVERLAY_SHARD_SCHEMA: u32 = 26; -pub(super) const OVERLAY_SHARD_SCHEMA: u32 = 64; +pub(super) const OVERLAY_SHARD_SCHEMA: u32 = 65; pub(super) const REGISTER_MMA_F8F6F4_SHARD_SCHEMA: u32 = 46; pub(super) const REGISTER_MMA_F8F6F4_F16_SHARD_SCHEMA: u32 = 47; pub(super) const REGISTER_MMA_MXF8F6F4_SHARD_SCHEMA: u32 = 60; @@ -42,6 +42,7 @@ pub(super) const CLUSTER_MEMORY_SHARD_SCHEMA: u32 = 39; pub(super) const CLC_SHARD_SCHEMA: u32 = 40; pub(super) const TMA_SHARD_SCHEMA: u32 = 61; pub(super) const TMA_REDUCTION_SHARD_SCHEMA: u32 = 62; +pub(super) const TMA_CTA_SHARD_SCHEMA: u32 = 65; pub(super) const MBARRIER_EXTENDED_SHARD_SCHEMA: u32 = 40; pub(super) const WGMMA_CONTROL_SHARD_SCHEMA: u32 = 38; pub(super) const TCGEN05_SHARD_SCHEMA: u32 = 42; @@ -556,6 +557,15 @@ pub(super) fn validate_overlay_shard_schema_with_max( "compact TMA reduction admission requires overlay shard schema {}", TMA_REDUCTION_SHARD_SCHEMA ); + ensure!( + shard + .tma + .as_ref() + .is_none_or(|admission| !admission.admits_cta_g2s()) + || shard.schema >= TMA_CTA_SHARD_SCHEMA, + "compact TMA CTA-local G2S admission requires overlay shard schema {}", + TMA_CTA_SHARD_SCHEMA + ); ensure!( shard.mbarrier_extended.is_none() || shard.schema >= MBARRIER_EXTENDED_SHARD_SCHEMA, "compact extended-mbarrier admission requires overlay shard schema {}", diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs index 69e7c4b2fd..f52d307ad3 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs @@ -1520,8 +1520,8 @@ pub(super) fn test_tma_admission() -> TmaAdmission { TmaAdmission { llvm_evidence_profile: "llvm-tma-test".into(), libnvvm_evidence_profile: "libnvvm-tma-test".into(), - cta_llvm_evidence_profile: "llvm-tma-test".into(), - cta_libnvvm_evidence_profile: "libnvvm-tma-test".into(), + cta_llvm_evidence_profile: Some("llvm-tma-test".into()), + cta_libnvvm_evidence_profile: Some("libnvvm-tma-test".into()), reduce_llvm_evidence_profile: Some("llvm-tma-reduce-test".into()), reduce_libnvvm_evidence_profile: Some("libnvvm-tma-reduce-test".into()), runtime_validation: RuntimeValidation::Unexecuted, diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs index c09b698ac8..f2dcbe9554 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs @@ -343,6 +343,17 @@ fn overlay_shard_schema_range_is_composable_and_new_fields_fail_closed() { .to_string() .contains("requires overlay shard schema 47") ); + let mut cta_tma = shard(TMA_CTA_SHARD_SCHEMA, None, None); + cta_tma.family = "tma".into(); + cta_tma.tma = Some(test_tma_admission()); + validate_overlay_shard_schema(&cta_tma, path).unwrap(); + cta_tma.schema -= 1; + assert!( + validate_overlay_shard_schema(&cta_tma, path) + .unwrap_err() + .to_string() + .contains("compact TMA CTA-local G2S admission requires overlay shard schema 65") + ); let mut sparse_f16_mma = shard(SPARSE_MMA_F8F6F4_F16_SHARD_SCHEMA, None, None); sparse_f16_mma.sparse_mma_f8f6f4_f16 = Some(test_sparse_mma_f8f6f4_f16_admission()); diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs index 57b88f0e2b..8a0d197e03 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs @@ -322,6 +322,47 @@ fn compact_tma_admission_matches_llvm_and_fails_closed() { executed.runtime_validation = RuntimeValidation::Executed; assert!(expand_tma_admission(&executed).is_err()); + let mut cta_without_profiles = test_tma_admission(); + cta_without_profiles.cta_llvm_evidence_profile = None; + assert!( + expand_tma_admission(&cta_without_profiles) + .unwrap_err() + .to_string() + .contains("requires cta_llvm_evidence_profile and cta_libnvvm_evidence_profile") + ); + + let cta_count = TMA_OPERATIONS + .iter() + .filter(|operation| operation.is_cta_g2s()) + .count(); + let mut cluster_only = test_tma_admission(); + cluster_only + .variants + .retain(|variant| !variant.operation.is_cta_g2s()); + cluster_only.cta_llvm_evidence_profile = None; + cluster_only.cta_libnvvm_evidence_profile = None; + let cluster_only_records = expand_tma_admission(&cluster_only).unwrap(); + assert_eq!( + cluster_only_records.len(), + TMA_OPERATIONS.len() - cta_count + tma_reduction_matrix().len() + ); + assert!( + cluster_only_records + .iter() + .all(|record| !record.id.ends_with("_g2s_cta")) + ); + + let mut partial_cta = test_tma_admission(); + partial_cta.variants.retain(|variant| { + !variant.operation.is_cta_g2s() || variant.operation == TmaOperation::G2sCtaTile2d + }); + assert!( + expand_tma_admission(&partial_cta) + .unwrap_err() + .to_string() + .contains("canonical order") + ); + let declaration = declarations[records[0].source_record.as_deref().unwrap()]; let mut wrong_adapter = records[0].clone(); wrong_adapter.tma.as_mut().unwrap().adapter = TmaAdapter::NoOperands; @@ -330,11 +371,18 @@ fn compact_tma_admission_matches_llvm_and_fails_closed() { #[test] fn tma_compact_schema_is_reserved_for_aggregation() { - let shard = |schema: u32, include_reductions: bool| { + let shard = |schema: u32, include_reductions: bool, include_cta: bool| { let mut admission = test_tma_admission(); if !include_reductions { admission.reduce_variants.clear(); } + if !include_cta { + admission + .variants + .retain(|variant| !variant.operation.is_cta_g2s()); + admission.cta_llvm_evidence_profile = None; + admission.cta_libnvvm_evidence_profile = None; + } OverlayShardFile { schema, family: "tma".into(), @@ -376,14 +424,14 @@ fn tma_compact_schema_is_reserved_for_aggregation() { let path = Path::new("intrinsics/overlay/tma.toml"); validate_overlay_shard_schema_with_max( - &shard(TMA_SHARD_SCHEMA, false), + &shard(TMA_SHARD_SCHEMA, false, false), path, TMA_REDUCTION_SHARD_SCHEMA, ) .unwrap(); assert!( validate_overlay_shard_schema_with_max( - &shard(TMA_SHARD_SCHEMA - 1, false), + &shard(TMA_SHARD_SCHEMA - 1, false, false), path, TMA_REDUCTION_SHARD_SCHEMA, ) @@ -393,14 +441,14 @@ fn tma_compact_schema_is_reserved_for_aggregation() { ); validate_overlay_shard_schema_with_max( - &shard(TMA_REDUCTION_SHARD_SCHEMA, true), + &shard(TMA_REDUCTION_SHARD_SCHEMA, true, false), path, TMA_REDUCTION_SHARD_SCHEMA, ) .unwrap(); assert!( validate_overlay_shard_schema_with_max( - &shard(TMA_REDUCTION_SHARD_SCHEMA - 1, true), + &shard(TMA_REDUCTION_SHARD_SCHEMA - 1, true, false), path, TMA_REDUCTION_SHARD_SCHEMA, ) @@ -408,6 +456,23 @@ fn tma_compact_schema_is_reserved_for_aggregation() { .to_string() .contains("compact TMA reduction admission requires overlay shard schema 62") ); + + validate_overlay_shard_schema_with_max( + &shard(TMA_CTA_SHARD_SCHEMA, true, true), + path, + TMA_CTA_SHARD_SCHEMA, + ) + .unwrap(); + assert!( + validate_overlay_shard_schema_with_max( + &shard(TMA_CTA_SHARD_SCHEMA - 1, true, true), + path, + TMA_CTA_SHARD_SCHEMA, + ) + .unwrap_err() + .to_string() + .contains("compact TMA CTA-local G2S admission requires overlay shard schema 65") + ); } #[test] diff --git a/cuda-oxide/intrinsics/overlay/tma.toml b/cuda-oxide/intrinsics/overlay/tma.toml index 9948c5bbf9..dbb2a826ca 100644 --- a/cuda-oxide/intrinsics/overlay/tma.toml +++ b/cuda-oxide/intrinsics/overlay/tma.toml @@ -1,6 +1,6 @@ # Reviewed Tensor Memory Accelerator operations. -schema = 62 +schema = 65 family = "tma" [tma] From e786af00b8b431cec1fcdd832d32866f44135853 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 19:28:06 +0530 Subject: [PATCH 04/17] docs(tma): state CTA locality in the raw-layer safety text and refresh stale CTA comments The raw ABI entries for the CTA-local G2S forms keep a generic destination pointer that the lowering converts with cvta.to.shared, but their generated safety text only asked for valid shared-memory objects. The renderer now says that the destination and barrier must lie in the issuing CTA's own shared memory and that a cluster-shared pointer from map_shared_rank would yield a wrong local offset. The mir-lower doc comment for the CTA-local lowering now describes the destination as shared by type with the barrier as the caller's promise, and names the cast and from_raw gap. The importer arity diagnostic names the _cta form when that form is being emitted. Signed-off-by: nihalpasham --- .../cuda-intrinsics-gen/src/render/raw_abi.rs | 9 +++++++++ .../src/translator/terminator/intrinsics/tma.rs | 3 ++- .../mir-lower/src/convert/intrinsics/tma.rs | 16 +++++++++++----- 3 files changed, 22 insertions(+), 6 deletions(-) diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/raw_abi.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/raw_abi.rs index 6477e238cf..20fac09b2e 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/raw_abi.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/raw_abi.rs @@ -330,6 +330,15 @@ pub(super) fn render_raw_abi(catalog: &CatalogFile, hash: &str) -> Result + { + output.push_str( + "/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality.\n\ + /// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form.\n\ + /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer.\n", + ) + } TmaAdapter::G2sPointersCoordinatesBarrierInjectDefaults | TmaAdapter::G2sPointersCoordinatesBarrierMaskInjectDefaults => output .push_str( diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs index de907452d4..034567db76 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/tma.rs @@ -108,8 +108,9 @@ fn emit_tma_g2s_impl( return input_err!( loc.clone(), TranslationErr::unsupported(format!( - "cp_async_bulk_tensor_{}d_g2s expects {} arguments, got {}", + "cp_async_bulk_tensor_{}d_g2s{} expects {} arguments, got {}", dims, + if cta { "_cta" } else { "" }, expected_args, args.len() )) diff --git a/cuda-oxide/crates/mir-lower/src/convert/intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/intrinsics/tma.rs index 4f5b18b089..66a51c81eb 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/intrinsics/tma.rs @@ -85,11 +85,17 @@ fn pointer_address_space(ctx: &Context, value: pliron::value::Value) -> Option`, which lowers to an `addrspace(3)` pointer, while the raw +/// ABI entry still passes a generic pointer under the caller's promise of +/// CTA locality. The barrier is a caller promise on both layers, as for every +/// mbarrier operation. Generic and shared pointers are converted to the +/// shared window. A pointer known to be cluster shared (`addrspace(7)`, for +/// example from `map_shared_rank`) may name another CTA's shared memory, so +/// it is rejected rather than reinterpreted as a local offset; such copies +/// use the cluster form. The check sees only uncast operands: a pointer cast +/// or `SharedPtr::from_raw` applied to a cluster pointer is the caller's +/// unsafe contract. pub(crate) fn convert_g2s_cta( ctx: &mut Context, rewriter: &mut DialectConversionRewriter, From 0f42431f1f0c967e1bb8b79ce612475af8078663 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 19:28:06 +0530 Subject: [PATCH 05/17] chore(intrinsics): re-record the CTA TMA evidence under LLVM 23.1.3 and regenerate The CTA-local G2S evidence was recorded against the LLVM revision the previous nightly shipped, which the catalog no longer admits. Replace it with rust-llvm-23.1.3-aab11ee0-tma-cta.json, written from five candidate probe runs of the pinned nightly-2026-10-11 sysroot llc at sm_90 +ptx86, with CUDA 13.3 ptxas V13.3.73 as the terminal stage; every stage detail cites the artifact sizes and SHA-256 digests of those runs. Regenerate the catalog outputs on top of main's copies (catalog SHA-256 27a44de2: 581 operations, 555 generated, 1034 entries, tma 116) and refresh the book stamp at both sites. Signed-off-by: nihalpasham --- .../cuda-device/src/generated/async_copy.rs | 2 +- .../cuda-device/src/generated/atomic.rs | 2 +- .../crates/cuda-device/src/generated/bf16.rs | 2 +- .../cuda-device/src/generated/bf16x2.rs | 2 +- .../crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- .../cuda-device/src/generated/convert.rs | 2 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- .../cuda-device/src/generated/dotprod.rs | 2 +- .../crates/cuda-device/src/generated/f16.rs | 2 +- .../crates/cuda-device/src/generated/f16x2.rs | 2 +- .../crates/cuda-device/src/generated/f32x2.rs | 2 +- .../crates/cuda-device/src/generated/fence.rs | 2 +- .../crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- .../cuda-device/src/generated/grid_sreg.rs | 2 +- .../crates/cuda-device/src/generated/i16x2.rs | 2 +- .../crates/cuda-device/src/generated/int.rs | 2 +- .../cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- .../cuda-device/src/generated/movmatrix.rs | 2 +- .../crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- .../crates/cuda-device/src/generated/sreg.rs | 2 +- .../cuda-device/src/generated/stmatrix.rs | 2 +- .../cuda-device/src/generated/tcgen05.rs | 2 +- .../crates/cuda-device/src/generated/tma.rs | 100 +- .../cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- .../cuda-intrinsics/src/generated/abi_v1.rs | 113 +- .../cuda-intrinsics/src/generated/mod.rs | 47 +- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 2 +- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 347 ++- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 292 ++- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- .../src/ops/generated/active_mask.rs | 2 +- .../dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- .../dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 2 +- .../dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- .../dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- .../dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- .../dialect-nvvm/src/ops/generated/tma.rs | 82 +- .../dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 27 +- .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 2 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 2 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 182 +- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 37 +- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 2 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 72 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../src/generated_intrinsics.rs | 22 +- .../cuda-oxide-book/compiler/mlir-dialects.md | 6 +- cuda-oxide/intrinsics/catalog.json | 2042 ++++++++++++++--- .../rust-llvm-23.1.0-16696adc-tma-cta.json | 116 - .../rust-llvm-23.1.3-aab11ee0-tma-cta.json | 252 ++ cuda-oxide/intrinsics/generated-reference.md | 72 +- cuda-oxide/intrinsics/probes/abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/active_mask.ll | 2 +- cuda-oxide/intrinsics/probes/add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f64.ll | 2 +- .../intrinsics/probes/add_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f64.ll | 2 +- .../intrinsics/probes/add_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f64.ll | 2 +- .../intrinsics/probes/add_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f64.ll | 2 +- .../intrinsics/probes/add_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/ballot_sync.ll | 2 +- .../probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- .../intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_sync.ll | 2 +- .../probes/barrier_cta_sync_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- .../probes/clc_query_is_canceled.ll | 2 +- .../intrinsics/probes/clc_try_cancel.ll | 2 +- .../probes/clc_try_cancel_multicast.ll | 2 +- .../intrinsics/probes/clock.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock.llvm_nvptx.ll | 2 +- .../intrinsics/probes/clock64.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- .../intrinsics/probes/cluster_block_count.ll | 2 +- .../intrinsics/probes/cluster_block_idx_x.ll | 2 +- .../intrinsics/probes/cluster_block_idx_y.ll | 2 +- .../intrinsics/probes/cluster_block_idx_z.ll | 2 +- .../intrinsics/probes/cluster_block_rank.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_z.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_x.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_y.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_z.ll | 2 +- .../intrinsics/probes/cos_approx_f32.ll | 2 +- .../intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s_cta.ll | 4 +- .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_cta.ll | 4 +- .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s_cta.ll | 4 +- .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s_cta.ll | 4 +- .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s_cta.ll | 4 +- .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- .../probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- .../intrinsics/probes/cp_async_ca_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_8.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- .../intrinsics/probes/cp_async_cg_16.ll | 2 +- .../intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- .../probes/cp_async_commit_group.ll | 2 +- .../probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- .../intrinsics/probes/cp_async_wait_all.ll | 2 +- .../intrinsics/probes/cp_async_wait_group.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f64.ll | 2 +- .../intrinsics/probes/div_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f64.ll | 2 +- .../intrinsics/probes/div_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f64.ll | 2 +- .../intrinsics/probes/div_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f64.ll | 2 +- .../intrinsics/probes/div_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_u32.ll | 2 +- .../intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- .../intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- .../probes/elect_sync.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- .../intrinsics/probes/ex2_approx_f16.ll | 2 +- .../intrinsics/probes/ex2_approx_f32.ll | 2 +- .../intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- cuda-oxide/intrinsics/probes/fma_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll | 2 +- .../intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- .../intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- .../intrinsics/probes/fma_relu_bf16x2.ll | 2 +- .../intrinsics/probes/fma_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f64.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f64.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f64.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f64.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll | 2 +- .../intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- .../probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- .../intrinsics/probes/grid_dependency_wait.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_z.ll | 2 +- .../intrinsics/probes/gridid.lib_nvvm.ll | 2 +- .../intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/lane_id.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_eq.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_ge.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_gt.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_le.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- .../intrinsics/probes/lg2_approx_f32.ll | 2 +- .../intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/map_shared_rank.ll | 2 +- .../intrinsics/probes/match_all_i64_sync.ll | 2 +- .../intrinsics/probes/match_all_sync.ll | 2 +- .../intrinsics/probes/match_any_i64_sync.ll | 2 +- .../intrinsics/probes/match_any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16.ll | 2 +- .../intrinsics/probes/max_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/max_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/max_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_u16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- .../probes/max_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- .../probes/max_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/mbarrier_arrive.ll | 2 +- .../probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_init.ll | 2 +- .../intrinsics/probes/mbarrier_inval.ll | 2 +- .../intrinsics/probes/mbarrier_test_wait.ll | 2 +- .../intrinsics/probes/mbarrier_try_wait.ll | 2 +- .../probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16.ll | 2 +- .../intrinsics/probes/min_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/min_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/min_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_u16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- .../probes/min_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- .../probes/min_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- .../probes/mma_m16n8k16_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- .../intrinsics/probes/movmatrix_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/mul_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f64.ll | 2 +- .../intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f64.ll | 2 +- .../intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f64.ll | 2 +- .../intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f64.ll | 2 +- .../intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/nanosleep.ll | 2 +- cuda-oxide/intrinsics/probes/neg_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/neg_f16x2.ll | 2 +- .../intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- .../probes/packed_atomic_add_bf16x2.ll | 2 +- .../probes/packed_atomic_add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/pmevent.ll | 2 +- .../probes/prefetch_tma_descriptor.ll | 2 +- cuda-oxide/intrinsics/probes/prmt.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_b4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecl.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecr.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_f4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc16.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc8.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f64.ll | 2 +- .../intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f64.ll | 2 +- .../intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f64.ll | 2 +- .../intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f64.ll | 2 +- .../intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/redux_sync_add.ll | 2 +- .../intrinsics/probes/redux_sync_and.ll | 2 +- .../probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_i32.ll | 2 +- .../probes/redux_sync_max_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_u32.ll | 2 +- .../probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_i32.ll | 2 +- .../probes/redux_sync_min_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_or.ll | 2 +- .../intrinsics/probes/redux_sync_xor.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f32.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f64.ll | 2 +- .../intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/setmaxnreg_dec.ll | 2 +- .../intrinsics/probes/setmaxnreg_inc.ll | 2 +- .../probes/shuffle_down_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_down_sync.ll | 2 +- .../probes/shuffle_down_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_sync.ll | 2 +- .../intrinsics/probes/shuffle_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- .../intrinsics/probes/sin_approx_f32.ll | 2 +- .../intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll | 2 +- .../intrinsics/probes/smid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/sqrt_approx_f32.ll | 2 +- .../intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll | 2 +- .../intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll | 2 +- .../intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll | 2 +- .../intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll | 2 +- .../intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/sub_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sync_mask.ll | 2 +- cuda-oxide/intrinsics/probes/sync_threads.ll | 2 +- .../intrinsics/probes/tanh_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc.ll | 2 +- .../intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_commit.ll | 2 +- .../intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- .../probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_load_wait.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_shared.ll | 2 +- .../probes/tcgen05_mma_sp_shared.ll | 2 +- .../probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- .../probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_shift_down.ll | 2 +- .../probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- .../intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- .../probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence.ll | 2 +- .../intrinsics/probes/threadfence_block.ll | 2 +- .../intrinsics/probes/threadfence_system.ll | 2 +- .../probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/trap.ll | 2 +- cuda-oxide/intrinsics/probes/uni_sync.ll | 2 +- .../intrinsics/probes/warpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/wgmma_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_fence.ll | 2 +- .../intrinsics/probes/wgmma_wait_group.ll | 2 +- 1324 files changed, 4652 insertions(+), 1783 deletions(-) delete mode 100644 cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json create mode 100644 cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-tma-cta.json diff --git a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs index 10e6c3e0b2..318879e1f1 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs index 425aedc4ac..2c06168531 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs index 57652a5228..1962d80438 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs index aa95140971..3aae5a6142 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/clc.rs b/cuda-oxide/crates/cuda-device/src/generated/clc.rs index 0c9b7ff9aa..4918b2d05a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/clc.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs index a4b905b524..ee7530ce4f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs index 04d544f30c..677a31bfd7 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs index 6be477a270..273a8d2253 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::cluster`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/convert.rs b/cuda-oxide/crates/cuda-device/src/generated/convert.rs index 36a19cb533..17fddd7e93 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/convert.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::convert`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs index 872d91cee2..5c6a2ef0f0 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs index d6828a8aba..76e006e07d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs index ee39d4257d..d58c92b45d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs index 37d8248126..18963fed82 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16.rs b/cuda-oxide/crates/cuda-device/src/generated/f16.rs index 2ec7953393..d568cb01ac 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs index d985923438..9a1591173b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs index 349e4a4d35..11664b325e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/fence.rs b/cuda-oxide/crates/cuda-device/src/generated/fence.rs index 88ee2432f3..34cc985181 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/fence.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/float.rs b/cuda-oxide/crates/cuda-device/src/generated/float.rs index 4dcf1edb9b..60db1c80e9 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/float.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::float`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs index 3fb0c6490c..7457da7e6c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs index 9b98009d81..21db68db5d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs index e99a610b45..3aa25b8e8a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/int.rs b/cuda-oxide/crates/cuda-device/src/generated/int.rs index 3788f17e4b..b56a33ff0a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/int.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs index c90b321c91..1ce99c08cf 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs index f7e8c4e54a..b1d777d9fe 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs index 0f1714ad05..3e4104456a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs index 3b0f283d9b..10c1f3fdc8 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs index d7a9131ae3..7bc30d4521 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs index 86e1231142..1501078c06 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs index e643c970b3..cef0fb4d14 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs index 037c21f721..eb485fe174 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs index 08980edf29..31a91a170e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs index 412e03bf24..748a2c0495 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs index ece2a3992a..6594ee5105 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs index 0ad1f01a09..5a34a464ff 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tma.rs b/cuda-oxide/crates/cuda-device/src/generated/tma.rs index 92226b0c88..7da6e40ca6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tma` to keep its public API stable. @@ -233,6 +233,23 @@ pub unsafe fn cp_async_bulk_tensor_1d_g2s( unreachable!("cp_async_bulk_tensor_1d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_1d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, barrier); + unreachable!("cp_async_bulk_tensor_1d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -263,6 +280,24 @@ pub unsafe fn cp_async_bulk_tensor_2d_g2s( unreachable!("cp_async_bulk_tensor_2d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_2d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, barrier); + unreachable!("cp_async_bulk_tensor_2d_g2s_cta called outside CUDA kernel context") +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// # Safety @@ -329,6 +364,25 @@ pub unsafe fn cp_async_bulk_tensor_3d_g2s( unreachable!("cp_async_bulk_tensor_3d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_3d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, barrier); + unreachable!("cp_async_bulk_tensor_3d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -363,6 +417,26 @@ pub unsafe fn cp_async_bulk_tensor_4d_g2s( unreachable!("cp_async_bulk_tensor_4d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_4d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, coord3, barrier); + unreachable!("cp_async_bulk_tensor_4d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -402,6 +476,30 @@ pub unsafe fn cp_async_bulk_tensor_5d_g2s( unreachable!("cp_async_bulk_tensor_5d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +#[allow(clippy::too_many_arguments)] +pub unsafe fn cp_async_bulk_tensor_5d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + coord4: i32, + barrier: *mut Barrier, +) { + let _ = ( + dst, tensor_map, coord0, coord1, coord2, coord3, coord4, barrier, + ); + unreachable!("cp_async_bulk_tensor_5d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety diff --git a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs index b8f8444bac..b9bc56377f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs index 6a734aa7fb..9c6207060b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 3eccfe34de..7a928dc5b9 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Raw ABI functions recognized by cuda-oxide. @@ -1181,6 +1181,22 @@ pub unsafe fn i0328(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u6 ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_1d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1030(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u64) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_1d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1217,6 +1233,28 @@ pub unsafe fn i0329( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_2d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1031( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` executed outside device compilation" + ) +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// Catalog ID: `cp_async_bulk_tensor_2d_g2s_multicast`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;`. @@ -1298,6 +1336,29 @@ pub unsafe fn i0332( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_3d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1032( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_3d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1336,6 +1397,30 @@ pub unsafe fn i0333( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_4d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1033( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_4d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1383,6 +1468,32 @@ pub unsafe fn i0334( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_5d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[allow(clippy::too_many_arguments)] +#[inline(never)] +pub unsafe fn i1034( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: i32, + _arg7: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_5d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];`. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs index fb3435f9b8..aa7ca75dd5 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. #[doc(hidden)] @@ -1094,6 +1094,11 @@ pub mod tma { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0984 as cp_async_bulk_tensor_reduce_xor_im2col_3d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0985 as cp_async_bulk_tensor_reduce_xor_im2col_4d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0986 as cp_async_bulk_tensor_reduce_xor_im2col_5d; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1030 as cp_async_bulk_tensor_1d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1031 as cp_async_bulk_tensor_2d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1032 as cp_async_bulk_tensor_3d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1033 as cp_async_bulk_tensor_4d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1034 as cp_async_bulk_tensor_5d_g2s_cta; } /// Generated `warp` intrinsic source API. @@ -1690,6 +1695,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_1d_g2s_cta_reexports_abi_i1030() { + let public: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_1d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1030; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_1d_s2g_reexports_abi_i0335() { let public: unsafe fn(*const u8, *const u8, i32) -> () = super::tma::cp_async_bulk_tensor_1d_s2g; @@ -1706,6 +1719,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_2d_g2s_cta_reexports_abi_i1031() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_2d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1031; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_2d_g2s_multicast_reexports_abi_i0330() { let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64, u16) -> () = super::tma::cp_async_bulk_tensor_2d_g2s_multicast; @@ -1738,6 +1759,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_3d_g2s_cta_reexports_abi_i1032() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_3d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1032; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_3d_s2g_reexports_abi_i0337() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_3d_s2g; @@ -1754,6 +1783,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_4d_g2s_cta_reexports_abi_i1033() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_4d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1033; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_4d_s2g_reexports_abi_i0338() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_4d_s2g; @@ -1770,6 +1807,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_5d_g2s_cta_reexports_abi_i1034() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_5d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1034; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_5d_s2g_reexports_abi_i0339() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_5d_s2g; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index 6dfd69169a..e85efe8e33 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index aadf572896..442a332616 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index fe36a8aff8..143cea136f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index 1e1509cef6..ed83253c5d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index ead108a430..a7af5344ae 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index 8cf09104c6..6e6cd6834a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index 7020044fa2..132e42f57d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 72c7864a4f..19de0d5a88 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index 1953d9b216..c03ac79fbb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index fcb91feaf8..23841b8439 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index 89d677a028..6fb2e2ef85 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index c6c7586c8c..d99779ccdc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index 3e1feecc57..548860aefb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index 5b15bf968e..5c3eb52a1f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index fb9ad47daa..67d3dece72 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index 9530f36fbc..4ba5f99824 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 9dca2e8bfa..7820e75597 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index 76f6352d37..0605e8cc67 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index 41849141b1..597007a6a0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 26829231f7..3146ae0269 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index 643b0122d8..6919d0b0dd 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index d21b5037bf..f133c9c2a0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index f714ace2f2..b191571ea6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 26c2b9101e..54094d2c3b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index ca977feaf9..48e9e8e2d7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 60066e6a95..3b658ccf7b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index ced2ff35e1..dc889eb5cd 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index 6987c58ffe..14864ab794 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index e6589fd47f..78e53e6363 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index e6dfa4e62d..63b3fbc03d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index a71b4e90ce..c568a6a44e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index 2c5b7977ea..3562de4fd2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index 7519ac479c..db644a1558 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index 8101ef71f2..1ca21e6ee8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index 942b0d8d20..4204ab5120 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index b4388a4208..baef4fc35d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index b05ce7cbf9..b593775d06 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index d13cbf86fb..7f632bddd7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index a35c18829c..c0f768cf0b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index 4ea27e9777..f6ca650831 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index be7f8fbead..5bf34eec50 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index e56d77b272..d89b227403 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index c681aa4b93..914de29661 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 458a3d4b8f..069c13135a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index f1503abbd0..125d097eb0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 9f808f3507..59175e4a3d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index e6d33d9513..947e6bd8b3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 678bbd7350..070bcdb18e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index 971f157365..9942389730 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index eacc5c5122..d9a27b2b3c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests, grouped per family shard. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index f0e0a4f733..08811cb41b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 0945d8c1f6..72c32ac38f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index 7ed4a44ef6..50c6718100 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index 6f86271313..c94dbe9d6d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index 491b841ab5..2e035456c9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index 8647954b1c..5438d3374d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 9865a8fa47..07d5a64bf5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index 478197ec78..aaa6d1adfc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index 60e4b4a4f0..266c54837e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index d775d2d9c0..2b704752c4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index da31b53b2c..42bd2becf8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index b2477c9d74..67c9b93ad7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index 8a1758029b..8038b4d5fb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index c136d5c526..971a086586 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index aa7f15a2bb..5d4ed2a92c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index 0659c6ce7e..61f2a0edae 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index 67318df319..f0f98ff0df 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 322035b874..3831e729f1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index 1aecf409b0..b44991f20e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index 5a2af67e32..068adac85d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index 0d5606e517..4de3cb9a94 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index 764bc401f0..2a9631b1b9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tma` intrinsics. @@ -6643,3 +6643,348 @@ fn cp_async_bulk_tensor_reduce_xor_im2col_5d_target_matches_the_catalog() { ] as &[&str] ); } + +#[test] +fn cp_async_bulk_tensor_1d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1030").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_1d_g2s_cta"); + assert_eq!(target.abi_id, "i1030"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_2d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1031").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_2d_g2s_cta"); + assert_eq!(target.abi_id, "i1031"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_3d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1032").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_3d_g2s_cta"); + assert_eq!(target.abi_id, "i1032"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_4d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1033").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_4d_g2s_cta"); + assert_eq!(target.abi_id, "i1033"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_5d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1034").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_5d_g2s_cta"); + assert_eq!(target.abi_id, "i1034"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 04c7f5e3a7..07de9882e0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index ce2d1695bb..c6d9df42ec 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index f20f4c4491..1fbc940cb7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index 36db24882a..b08c5491bd 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 447579f72e..14caa5adcb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index 129f12b861..1ace05c9e8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tma` intrinsics. @@ -5670,4 +5670,294 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ result_range: None, }), }, + GeneratedIntrinsicTarget { + marker: "v1:i1030", + id: "cp_async_bulk_tensor_1d_g2s_cta", + abi_id: "i1030", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1031", + id: "cp_async_bulk_tensor_2d_g2s_cta", + abi_id: "i1031", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1032", + id: "cp_async_bulk_tensor_3d_g2s_cta", + abi_id: "i1032", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1033", + id: "cp_async_bulk_tensor_4d_g2s_cta", + abi_id: "i1033", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1034", + id: "cp_async_bulk_tensor_5d_g2s_cta", + abi_id: "i1034", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, ]; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 033064166b..ccdd65c2fa 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index d45478d0ce..2559ea2ca8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index 7e1c4672ee..d977330505 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index f44e7d1502..4bfaba20e1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index 90d30ad393..4199dbf817 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 2c3f93f347..2ca1a95d6d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for the generated active warp mask. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs index 616d65ca24..9773039800 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Cluster Launch Control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index 1e84152b60..1ec533f00b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated operations for the closed cluster-barrier family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index f475f9c8a1..c1ad4b86f5 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs index 373648c5be..25758a6591 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for classic `cp.async` instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs index 812c00a924..68869c4b6a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated PTX debug controls. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs index b4eed5cbbd..41b5a740b5 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed integer dot products. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs index 0975f2887d..a42f65a7a9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated warp leader-election operation. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs index 7ef65fad25..4efec68715 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index e8d6750774..b30f35f1be 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for extended floating-point min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 6bc73ad2ec..dc17124a6d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated extended integer min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index 5643285302..b575b3e8f1 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index 8d2447c25d..66e29afc37 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index 37a6d2de5f..1e629a1179 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs index 6346816724..38d74b564c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. mod active_mask; diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index 82ebc991f6..3256807492 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated in-register matrix transpose. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index db1dbf890f..c616df1189 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed floating-point arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index 96379c594d..e6095e28e0 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated packed-atomic family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index fe611f1857..686f008188 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated packed conversion. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs index 84759078dc..4ee1b29577 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated `prmt` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs index 9692546253..b210840413 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 8fef6c41f9..97a01da4e0 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index e30a341c2f..178890b226 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index f13122c5cd..f086e0cefb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index ff68d7ff74..efbd87b1c8 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar math. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 0ab9e5d669..84b10189e9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated sparse MMA variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs index 19a5061577..d516b35ed9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for generated special-register reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index ebb671fe9a..75e73d66fb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the four generated `stmatrix` stores. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs index 90a3220586..dbbc71bf58 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated CTA synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index f59d9a2360..febb5d5d3e 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Core Generation 5 operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs index ea10d67f42..1e2cfc266c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Memory Accelerator operations. @@ -227,6 +227,21 @@ impl CpAsyncBulkTensorG2sTile1dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<5>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile1dOp; + +impl CpAsyncBulkTensorG2sCtaTile1dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_1d", @@ -257,6 +272,21 @@ impl CpAsyncBulkTensorG2sTile2dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<6>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile2dOp; + +impl CpAsyncBulkTensorG2sCtaTile2dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast", @@ -317,6 +347,21 @@ impl CpAsyncBulkTensorG2sTile3dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<7>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile3dOp; + +impl CpAsyncBulkTensorG2sCtaTile3dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_3d", @@ -347,6 +392,21 @@ impl CpAsyncBulkTensorG2sTile4dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<8>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile4dOp; + +impl CpAsyncBulkTensorG2sCtaTile4dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_4d", @@ -377,6 +437,21 @@ impl CpAsyncBulkTensorG2sTile5dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<9>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile5dOp; + +impl CpAsyncBulkTensorG2sCtaTile5dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_5d", @@ -1697,16 +1772,21 @@ pub(super) fn register(ctx: &mut Context) { CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op::register(ctx); CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp::register(ctx); CpAsyncBulkTensorG2sTile1dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile1dOp::register(ctx); CpAsyncBulkTensorS2gTile1dOp::register(ctx); CpAsyncBulkTensorG2sTile2dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastCg2Op::register(ctx); CpAsyncBulkTensorS2gTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile3dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile3dOp::register(ctx); CpAsyncBulkTensorS2gTile3dOp::register(ctx); CpAsyncBulkTensorG2sTile4dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile4dOp::register(ctx); CpAsyncBulkTensorS2gTile4dOp::register(ctx); CpAsyncBulkTensorG2sTile5dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile5dOp::register(ctx); CpAsyncBulkTensorS2gTile5dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col3dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col4dOp::register(ctx); diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs index 082a1c4a04..9eee8930ac 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `vote.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 28109e6e97..15cc9dfb5f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated warp synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs index b0e5ae390f..5a9e9a05eb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `match.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 8066f000b3..918913cd6c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `shfl.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index 079d2f4753..be46657dbf 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated WGMMA controls. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index dbf3e84396..6b5617476c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index ade0190400..e8725fc71b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index 54b62778ae..192d3b3cfe 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index e3996216fd..37472270b6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index b3af25763b..ec5e550f53 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index a04d390784..67937e6e70 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 9fc04194b7..7959b71a2a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 3b6afc8dde..0425331510 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 265c0112f4..61726de0c4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 4ab825e358..6de2f7b601 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index dca4ce9c0b..3ca948b7ef 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 40595ac88b..7e1c55badd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index 3178c8f319..67dfb8596b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index d6a9cdc83f..1cb6e084fd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index b74612cfa4..c1a3afbdbb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index 531e5c92af..a5b9506522 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -313,10 +313,14 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -325,14 +329,20 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" @@ -2301,16 +2311,21 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3431,10 +3446,14 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" => Some("v1:i0328"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => Some("v1:i1030"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => Some("v1:i0335"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" => Some("v1:i0329"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => Some("v1:i1031"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => Some("v1:i0330"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -3443,14 +3462,20 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" => Some("v1:i0336"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" => Some("v1:i0332"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => Some("v1:i1032"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => Some("v1:i0337"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" => Some("v1:i0333"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => Some("v1:i1033"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => Some("v1:i0338"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" => Some("v1:i0334"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => Some("v1:i1034"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => Some("v1:i0339"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 1510f4da7a..826ec0a0c0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 2f359a9d70..f80f9ae917 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index f66702fb4f..954d37ba58 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index 346a6badbc..f49208b6e6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index 69eaf65b5d..d6157b608f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index 269ecd2f67..0e0210e837 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index a741546fde..b3c9deec18 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 8db6935f06..2f30a26ef7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index 0ae8377b8d..44afc5b403 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index 62df911084..75a6ffec1c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 86c8b04476..85972feb8f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 042d22fe9a..951e3223db 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index 4ec1b4dc49..d8401d06c8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index a9533ccfd6..9e39168315 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 959c46be61..3650d4129e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index 2514ff7c0b..db4c304654 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 024eacb481..7dd180761f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index b48fe1d28e..268d20d8b9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index ffb9d6def0..97cd1350cc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index c4ab0cc802..2a0776d80f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index 84bf97d723..b16e300858 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index 9f0bcb6654..f96cf5828c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index b0a5c03ee5..ebb760e4a8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index 75edecc7e8..c5ed120dc5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index c0b0a46333..81cd65f3bf 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index ebf3d5ae86..0c3554e3e0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index 4f2a4b0dba..ad267da768 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index bfe29b8e9a..36590deb45 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index fc08e24bd9..6f4948b624 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index 4e168a9be0..b42f2eb906 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index bcd1ae8633..c89000a507 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index 4c4e7a84fc..fc3841f6ed 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 2cfdad1ac9..bf37827147 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index 773f7f2f0c..5fbb5c81ee 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index dada114a26..5e99c193ee 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 41d58a16a0..e217cc5837 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index c0c0d792ba..e662e47869 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 7c515139fa..3ea7569a39 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index e614d48ac4..ada5b75ba3 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index 3539b87869..5cdcc95c39 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index 766e88b2f5..bf1a26c17b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index 43676af093..b3184a9cdf 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index 45ccb2d478..103f192168 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index ea1a765cb8..02c6581dc2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 3e48ff5889..35068db47b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 4e2fa3ace6..9cc3214c21 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 77cf005293..cd573bda93 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 7c67982816..187557dafe 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 4dc9842f60..0336e432ad 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 2acc775bc3..ddd6c8c816 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tma` intrinsics. @@ -529,6 +529,42 @@ fn cp_async_bulk_tensor_1d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_1d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030"), + Some("v1:i1030") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta"), + Some("v1:i1030") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1030" + )); +} +#[test] fn cp_async_bulk_tensor_1d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" @@ -601,6 +637,42 @@ fn cp_async_bulk_tensor_2d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_2d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031"), + Some("v1:i1031") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta"), + Some("v1:i1031") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1031" + )); +} +#[test] fn cp_async_bulk_tensor_2d_g2s_multicast_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" @@ -745,6 +817,42 @@ fn cp_async_bulk_tensor_3d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_3d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032"), + Some("v1:i1032") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta"), + Some("v1:i1032") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1032" + )); +} +#[test] fn cp_async_bulk_tensor_3d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" @@ -817,6 +925,42 @@ fn cp_async_bulk_tensor_4d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_4d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033"), + Some("v1:i1033") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta"), + Some("v1:i1033") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1033" + )); +} +#[test] fn cp_async_bulk_tensor_4d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" @@ -889,6 +1033,42 @@ fn cp_async_bulk_tensor_5d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_5d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034"), + Some("v1:i1034") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta"), + Some("v1:i1034") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1034" + )); +} +#[test] fn cp_async_bulk_tensor_5d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index 2a3002383a..afbfce5f2c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 98ad159499..5e3b17d8df 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index f33fee4ce7..bf74c33b4a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index fe43a1c911..95eacbc9b2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index 5da7982e82..c4e9373c17 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 6fb539a5f9..38708c024c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. @@ -587,6 +587,13 @@ pub(super) fn try_dispatch( "v1:i0328", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 1, + "v1:i1030", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -601,6 +608,13 @@ pub(super) fn try_dispatch( "v1:i0329", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 2, + "v1:i1031", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => { Ok(Some(super::super::tma::emit_tma_g2s_multicast( @@ -627,6 +641,13 @@ pub(super) fn try_dispatch( "v1:i0332", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 3, + "v1:i1032", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -641,6 +662,13 @@ pub(super) fn try_dispatch( "v1:i0333", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 4, + "v1:i1033", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -655,6 +683,13 @@ pub(super) fn try_dispatch( "v1:i0334", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 5, + "v1:i1034", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 171dae8c5a..742fd66b6d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index 5b440dda69..a0f4db2d57 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 7f967c28c7..1d5832299a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index 9e852b7a6c..fd4f4683ca 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 99f9c2491c..9e88dd3302 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 8406c947cb..4eb027ff4d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index 7234063462..62f363da29 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index b3a91fd5e6..ec6caee406 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 3e1aa64684..b5574be1bc 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index c57daa967f..b93d87380c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index 576e50e4e9..76ce652f5f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 10bb6fe166..76dd9fcdc1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index e4d3bd35ca..2e84715962 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index 89dbae811c..a0c461a455 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index 969c0b61a3..92751f538d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index 435dd821d3..7c46d1f3b2 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index 4f53978669..c464f61053 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index 3555905110..a8fe431b0f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 75541afe5d..5155e53e3c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index 3215b3beee..b038ef4d33 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index 22f41b615a..c86adea87f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index 1c15e63abb..3662af1e3c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index ab105b7637..f58780d93f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index cd88761a89..c0ad431bd5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 065a2989e9..7354f3e94a 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index ba3730d318..b1a042093a 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index edb62cf830..67972807d0 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index a2a9b9bc70..f0cfa2ce46 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index a579f8bb6a..a0ffcd8439 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 3e23e8f32c..630a6f2a48 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index a1469bfddf..2b0cdbe397 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index 441e699695..bb8d589b59 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 8e465421a3..00da470c06 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 05fd89a47e..4386a0a0e7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index 45cfb5e9c1..ee45bab7f6 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index 3ba049ba8f..879420d1c9 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tma` intrinsics. @@ -12,9 +12,9 @@ use crate::conversion_interface::MirToLlvmConversion; use crate::convert::intrinsics::{ tma::PrefetchTileConfig, tma::ReduceConfig, tma::convert_control, tma::convert_g2s, - tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, tma::convert_prefetch_tile, - tma::convert_reduce_s2g, tma::convert_s2g, tma::convert_tensormap_fence, - tma::convert_tensormap_replace, + tma::convert_g2s_cta, tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, + tma::convert_prefetch_tile, tma::convert_reduce_s2g, tma::convert_s2g, + tma::convert_tensormap_fence, tma::convert_tensormap_replace, }; use dialect_nvvm::ops::{ CpAsyncBulkCommitGroupOp, CpAsyncBulkPrefetchTensor1dL2CacheHintOp, @@ -23,7 +23,9 @@ use dialect_nvvm::ops::{ CpAsyncBulkPrefetchTensor3dL2Op, CpAsyncBulkPrefetchTensor4dL2CacheHintOp, CpAsyncBulkPrefetchTensor4dL2Op, CpAsyncBulkPrefetchTensor5dL2CacheHintOp, CpAsyncBulkPrefetchTensor5dL2Op, CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp, - CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sTile1dOp, + CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sCtaTile1dOp, + CpAsyncBulkTensorG2sCtaTile2dOp, CpAsyncBulkTensorG2sCtaTile3dOp, + CpAsyncBulkTensorG2sCtaTile4dOp, CpAsyncBulkTensorG2sCtaTile5dOp, CpAsyncBulkTensorG2sTile1dOp, CpAsyncBulkTensorG2sTile2dMulticastCg2Op, CpAsyncBulkTensorG2sTile2dMulticastOp, CpAsyncBulkTensorG2sTile2dOp, CpAsyncBulkTensorG2sTile3dOp, CpAsyncBulkTensorG2sTile4dOp, CpAsyncBulkTensorG2sTile5dOp, CpAsyncBulkTensorReduceAddIm2col3dOp, @@ -383,6 +385,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile1dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile1dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 1) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile1dOp { fn convert( @@ -407,6 +421,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile2dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 2) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dMulticastOp { fn convert( @@ -455,6 +481,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile3dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile3dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 3) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile3dOp { fn convert( @@ -479,6 +517,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile4dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile4dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 4) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile4dOp { fn convert( @@ -503,6 +553,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile5dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile5dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 5) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile5dOp { fn convert( diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index 9de55faf9b..bda450a06a 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index 7e15564c79..71e6b5f00f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index b29139f764..1efa2de713 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index b84d82d837..90063d0f7c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index 5b6f6313f4..caf17366b3 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index d0f9932736..36399b6c9e 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated collector predicates for intrinsic placeholder functions. @@ -99,16 +99,21 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929", @@ -1131,16 +1136,21 @@ pub(crate) const GENERATED_INTRINSIC_PUBLIC_PATHS: &[&str] = &[ "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2", "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d", @@ -2169,16 +2179,21 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3208,16 +3223,21 @@ pub(crate) fn is_generated_intrinsic_compatibility_path(path: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2" | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_3d" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_4d" diff --git a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md index 5a99a3d72a..30f8dc7882 100644 --- a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `c243e568` (the stamp in every `ops/generated/` file -header), the dialect holds 576 operations across 42 modules, and they come +At catalog SHA-256 `27a44de2` (the stamp in every `ops/generated/` file +header), the dialect holds 581 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `c243e568`, the counts on this page predate the +stamp no longer starts with `27a44de2`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `cuda-oxide/crates/dialect-nvvm/src/ops/`. These are the diff --git a/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index 9c57efa92b..dc3dadac9e 100644 --- a/cuda-oxide/intrinsics/catalog.json +++ b/cuda-oxide/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "54d47b8e281dd953dc95fe16f834ff38164d9809171c9a1af097231b0bc45cab", - "overlay_sha256": "08f7a25510022bf48902cb4889b9f711a1cff31d1c5e8ce21c669d61afbb0310", - "abi_ledger_sha256": "fc378e74ef6da50df31de62c3fafcb3264dfb414034e5852147359fcad4184eb", + "overlay_sha256": "1c92502edd54de3fca6e7ab332654262ef9ecd27b60210550f7c856e51e31fa9", + "abi_ledger_sha256": "ce40abc882afa0a313cc657758fd07aef21dd98338d8015f9cccf4edf85eb83f", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -70,6 +70,7 @@ "cuda-13.3-libnvvm-13.3.73-execution-control.json:f66216bc7aaccca9ea66fc557e19fdeeaac8c277781f669ee969e727e203aac5", "cuda-13.3-libnvvm-13.3.73-register-mma-mxf8f6f4-f32.json:e437bf84570f4df0a28a743a132e0b93f359f9431d1af03b443f324f62b715ba", "cuda-13.3-libnvvm-13.3.73-sparse-mma-fp8.json:8a57d207ffd510b8c1bb40643bdb828d3d22e5d1c5ab24e1bc1b697538e9bc21", + "cuda-13.3-libnvvm-13.3.73-tma-cta.json:352f3e82f0f1c190d99e12e717ef5357a05729521a1fbab100b40964bc9526e7", "cuda-13.3-libnvvm-13.3.73-tma-descriptor-control.json:ac44f485122404f9a900cfb94bff658160a6c23d9e2f8ef74bb751f044b848fc", "cuda-13.3-libnvvm-13.3.73-tma-reduce.json:f5df5bb51068966594660cb01b1b28c3b72cb004c456fcd8dff377ff0bee4fc0", "cuda-13.3-libnvvm-13.3.73-tma.json:8607f93605f215858a98075a0ea1394f2718d090538bcaf0a795e11b4cb66d47", @@ -127,6 +128,7 @@ "rust-llvm-23.1.3-aab11ee0-tcgen05-st.json:e559aa775e3e964f9889483e9d9d8c45b46628931ff78050c46f33e448f69e86", "rust-llvm-23.1.3-aab11ee0-tcgen05.json:b6cf11d81c0880877ac5ad8958231222c728d5ab0243ddc84f1db901fbc8608b", "rust-llvm-23.1.3-aab11ee0-threadfence.json:087343e0247e7f9571c74fe763c0d48fbd26e246a21a9f10e5466a9a4457f83e", + "rust-llvm-23.1.3-aab11ee0-tma-cta.json:5aa68833705b5a98b393c4c752739a93cb3360d8d1c2c130bc28e09e475a208c", "rust-llvm-23.1.3-aab11ee0-tma-descriptor-control.json:7b6deaedfb8f8ac4b6551c47701cdefd04867431ac4dc5c3fbaf8f8bc604c9e9", "rust-llvm-23.1.3-aab11ee0-tma-reduce.json:e472b96476cad1d406fdac6dd8209645357f068cbd545669e2f09a8c6a232b79", "rust-llvm-23.1.3-aab11ee0-tma.json:97fb372aa7c6966bd01667da32d9d0857e385776131cb548f7d70f73b52718bc", @@ -18549,6 +18551,274 @@ }, "summary": "Starts a TMA tile copy from global to cluster shared memory." }, + { + "id": "cp_async_bulk_tensor_1d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.1d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d" + }, + "selections": [ + { + "source_record": "TMA_G2S_CTA_TILE_1D", + "asm": "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + "predicates": [ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 5, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i1030", + "module": "tma", + "name": "cp_async_bulk_tensor_1d_g2s_cta", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sCtaTile1dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "arguments": [ + "shared_ptr", + "shared_ptr", + "ptr", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probe for the 1D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_1d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,007-byte source had SHA-256 253d3ea39f9d26f5000f83febb9f72a15af2bed6ad0f668ab206976ce4c6d6ba; the 2,640-byte bitcode had SHA-256 9f42944557579c5c9c4e38716cce52607a07c51e5f2272393948fdc40cfb8961; the 1,234-byte canonical IR had SHA-256 835c2efc3b53d5333ec6a8260fcad29d5d0ce889194d0523feb57f7b22281fff.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "typed LLVM call for the 1D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,165-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 b0b4469f60fea3342f58e26eb4698834c6cb0714aac17fd3ffaed79cbd4f226a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the 1D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_cta_tile1d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "1d", + "shared::cta", + "global", + "tile", + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." + }, { "id": "cp_async_bulk_tensor_1d_s2g", "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.1d", @@ -19093,29 +19363,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "id": "cp_async_bulk_tensor_2d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.2d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_CTA_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 7, - "value": -1 - }, - { - "argument_index": 8, + "argument_index": 6, "value": 0 } ] @@ -19123,7 +19389,282 @@ } ], "rust": { - "abi_id": "i0330", + "abi_id": "i1031", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_cta", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sCtaTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "arguments": [ + "shared_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probe for the 2D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_2d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,038-byte source had SHA-256 f0d47c1b6670854dfddce5bc45206d8f4ad16e9a05236dd4e1d4036b518666ea; the 2,648-byte bitcode had SHA-256 6ab53bfd2185eb8b6d7f7ffb3793d72cb7b48a07f61846a5317604700bdd3225; the 1,265-byte canonical IR had SHA-256 2c00613205683c093c5002f0f7347d3b2d6b82e46c5bc00072fb5a781f7d9d44.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "typed LLVM call for the 2D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,299-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 59fa9746f48c669cbff59faffea6039a796055f2721b8647a6331d78b5a26521.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the 2D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_cta_tile2d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cta", + "global", + "tile", + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0330", "module": "tma", "name": "cp_async_bulk_tensor_2d_g2s_multicast", "arguments": [ @@ -19159,13 +19700,889 @@ "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->hasTMABlackwellSupport()" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0331", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64", + "u16" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + }, + "ptx_result": "()", + "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100a", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_100a" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 065ffbad7cc3f4a868f3de9e7f86d3a4fb1b0ea7047e1ddcdd578192e1fe101c.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 c6f52db1ccb491246b76348fb43a73fc961823d087d80432a2cec08dc36439f1.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 eea1fea3e3abbce7a6d377fafd8ad8e2c8fec998a990afbd664b7ce4f3856dff.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_100a", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx93" + ], + "representation": "linked cubin retaining the CTA-pair TMA instruction", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast_cg2", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster", + "cta_group::2" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_TENSOR_S2G_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 5, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0336", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_s2g", + "arguments": [ + "*const u8", + "*const u8", + "i32", + "i32" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorS2gTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "operands": [ + "ptr", + "ptr", + "i32", + "i32" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "arguments": [ + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "ReadOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "s2g_tile2d", + "adapter": "s2g_pointers_coordinates_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "global", + "shared::cta", + "tile", + "bulk_group" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from shared to global memory." + }, + { + "id": "cp_async_bulk_tensor_3d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 8, + "value": 0 + }, + { + "argument_index": 9, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0332", + "module": "tma", + "name": "cp_async_bulk_tensor_3d_g2s", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", "arguments": [ "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", + "i32", "i16", "i64", "i1", @@ -19174,11 +20591,11 @@ ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "ImmArg", "ImmArg", "IntrConvergent", - "Range", + "Range", "ReadOnly", "WriteOnly" ], @@ -19344,8 +20761,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19355,12 +20772,11 @@ "async", "bulk", "tensor", - "2d", + "3d", "shared::cluster", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster" + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19371,37 +20787,31 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "id": "cp_async_bulk_tensor_3d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_CTA_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", "predicates": [ - "Subtarget->hasTMABlackwellSupport()" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { "argument_index": 7, - "value": -1 - }, - { - "argument_index": 8, "value": 0 } ] @@ -19409,62 +20819,57 @@ } ], "rust": { - "abi_id": "i0331", + "abi_id": "i1032", "module": "tma", - "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "name": "cp_async_bulk_tensor_3d_g2s_cta", "arguments": [ "*mut u8", "*const u8", "i32", "i32", - "*mut u64", - "u16" + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "op_type": "CpAsyncBulkTensorG2sCtaTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", "operands": [ "ptr", "ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ "ImmArg", - "ImmArg", - "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -19485,63 +20890,39 @@ "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] }, "ptx_result": "()", - "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "targets": "all", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100a", + "gpu_target": "sm_90", "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19554,13 +20935,13 @@ "stages": [ { "targets": [ - "sm_100a" + "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 3D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 065ffbad7cc3f4a868f3de9e7f86d3a4fb1b0ea7047e1ddcdd578192e1fe101c.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_3d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,069-byte source had SHA-256 159ab50347ad4f81c8a110df7de35db35da9eb099938874523e5ec10ea867c8c; the 2,656-byte bitcode had SHA-256 3b5f19c93cffadf279efad33b501519f09cacbc660784f195ab3d007b5bca854; the 1,296-byte canonical IR had SHA-256 223dd701edae4ecf145ca18afc659dd5a50785cf8e492f6ba5c7b7f8299dc030.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19568,14 +20949,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 3D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 c6f52db1ccb491246b76348fb43a73fc961823d087d80432a2cec08dc36439f1.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,433-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 e9bab5f8e2c2a697ca435f86eaff468984a03470ae4bf28f975c2c89fc6676b9.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19583,14 +20964,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 3D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 eea1fea3e3abbce7a6d377fafd8ad8e2c8fec998a990afbd664b7ce4f3856dff.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -19601,27 +20982,15 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19634,14 +21003,14 @@ "stages": [ { "targets": [ - "compute_100a", + "compute_90", "ptx93" ], - "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19649,14 +21018,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx93" ], - "representation": "linked cubin retaining the CTA-pair TMA instruction", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -19666,8 +21035,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast_cg2", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_cta_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19677,13 +21046,11 @@ "async", "bulk", "tensor", - "2d", - "shared::cluster", + "3d", + "shared::cta", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster", - "cta_group::2" + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19694,26 +21061,23 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_2d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "id": "cp_async_bulk_tensor_3d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_2D", - "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "source_record": "TMA_TENSOR_S2G_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19721,7 +21085,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 5, + "argument_index": 6, "value": 0 } ] @@ -19729,49 +21093,52 @@ } ], "rust": { - "abi_id": "i0336", + "abi_id": "i0337", "module": "tma", - "name": "cp_async_bulk_tensor_2d_s2g", + "name": "cp_async_bulk_tensor_3d_s2g", "arguments": [ "*const u8", "*const u8", "i32", + "i32", "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile2dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "op_type": "CpAsyncBulkTensorS2gTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", "operands": [ "ptr", "ptr", "i32", + "i32", "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", "arguments": [ "shared_ptr", "ptr", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", "ReadOnly", "ReadOnly" @@ -19938,7 +21305,7 @@ } ], "tma": { - "operation": "s2g_tile2d", + "operation": "s2g_tile3d", "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, @@ -19949,7 +21316,7 @@ "async", "bulk", "tensor", - "2d", + "3d", "global", "shared::cta", "tile", @@ -19967,17 +21334,17 @@ "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_3d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_3D", - "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "source_record": "TMA_G2S_TILE_CG0_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19985,11 +21352,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 8, + "argument_index": 9, "value": 0 }, { - "argument_index": 9, + "argument_index": 10, "value": 0 } ] @@ -19997,30 +21364,31 @@ } ], "rust": { - "abi_id": "i0332", + "abi_id": "i0333", "module": "tma", - "name": "cp_async_bulk_tensor_3d_g2s", + "name": "cp_async_bulk_tensor_4d_g2s", "arguments": [ "*mut u8", "*const u8", "i32", "i32", "i32", + "i32", "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "op_type": "CpAsyncBulkTensorG2sTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", "operands": [ "ptr", "ptr", @@ -20028,13 +21396,14 @@ "i32", "i32", "i32", + "i32", "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", "arguments": [ "shared_cluster_ptr", "shared_ptr", @@ -20042,6 +21411,7 @@ "i32", "i32", "i32", + "i32", "i16", "i64", "i1", @@ -20051,10 +21421,10 @@ "results": [], "properties": [ "ImmArg", - "ImmArg", + "ImmArg", "ImmArg", "IntrConvergent", - "Range", + "Range", "ReadOnly", "WriteOnly" ], @@ -20220,7 +21590,7 @@ } ], "tma": { - "operation": "g2s_tile3d", + "operation": "g2s_tile4d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -20231,7 +21601,7 @@ "async", "bulk", "tensor", - "3d", + "4d", "shared::cluster", "global", "tile", @@ -20252,25 +21622,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_3d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_3D", - "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", + "source_record": "TMA_G2S_CTA_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 6, + "argument_index": 8, "value": 0 } ] @@ -20278,55 +21648,62 @@ } ], "rust": { - "abi_id": "i0337", + "abi_id": "i1033", "module": "tma", - "name": "cp_async_bulk_tensor_3d_s2g", + "name": "cp_async_bulk_tensor_4d_g2s_cta", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", + "op_type": "CpAsyncBulkTensorG2sCtaTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", - "i32" + "i32", + "i32", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", "arguments": [ + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -20340,7 +21717,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20357,21 +21734,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20392,11 +21769,11 @@ "targets": [ "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 4D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_4d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,100-byte source had SHA-256 becebd363b47d76e388fccf46646db4d8bba5225b4eb23cc7cae1673a5411ddd; the 2,664-byte bitcode had SHA-256 77f23ed44102d830fb8a576a8ff62f947e96b2b15871e2984083851d1e82f63e; the 1,327-byte canonical IR had SHA-256 6fd6a38abbb214d1e1fc5a4c8aa800ab963d1f53350afc3a7ec965501e467fb4.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20405,13 +21782,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 4D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,567-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 0607dfca39829f02dbda9367602cb27d28e6dc5dd8f5909a451fa4648442629a.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20420,13 +21797,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 4D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -20437,9 +21814,9 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20461,11 +21838,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20476,11 +21853,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -20490,8 +21867,8 @@ } ], "tma": { - "operation": "s2g_tile3d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_cta_tile4d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20501,35 +21878,38 @@ "async", "bulk", "tensor", - "3d", - "global", + "4d", "shared::cta", + "global", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_4d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", + "id": "cp_async_bulk_tensor_4d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_4D", - "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + "source_record": "TMA_TENSOR_S2G_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20537,11 +21917,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 9, - "value": 0 - }, - { - "argument_index": 10, + "argument_index": 7, "value": 0 } ] @@ -20549,69 +21925,58 @@ } ], "rust": { - "abi_id": "i0333", + "abi_id": "i0338", "module": "tma", - "name": "cp_async_bulk_tensor_4d_g2s", + "name": "cp_async_bulk_tensor_4d_s2g", "arguments": [ - "*mut u8", + "*const u8", "*const u8", "i32", "i32", "i32", - "i32", - "*mut u64" + "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", + "op_type": "CpAsyncBulkTensorS2gTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", "operands": [ - "ptr", "ptr", "ptr", "i32", "i32", "i32", - "i32", - "i16", - "i64" + "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", "arguments": [ - "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", - "ReadOnly", - "WriteOnly" + "ReadOnly", + "ReadOnly" ], "result_facts": { "no_undef": false, @@ -20775,8 +22140,8 @@ } ], "tma": { - "operation": "g2s_tile4d", - "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "operation": "s2g_tile4d", + "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20787,37 +22152,34 @@ "bulk", "tensor", "4d", - "shared::cluster", "global", + "shared::cta", "tile", - "mbarrier::complete_tx::bytes" + "bulk_group" ], "operands": [ { "kind": "address" }, - { - "kind": "address" - }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_4d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", + "id": "cp_async_bulk_tensor_5d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_4D", - "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", + "source_record": "TMA_G2S_TILE_CG0_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20825,7 +22187,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 7, + "argument_index": 10, + "value": 0 + }, + { + "argument_index": 11, "value": 0 } ] @@ -20833,58 +22199,72 @@ } ], "rust": { - "abi_id": "i0338", + "abi_id": "i0334", "module": "tma", - "name": "cp_async_bulk_tensor_4d_s2g", + "name": "cp_async_bulk_tensor_5d_g2s", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", + "op_type": "CpAsyncBulkTensorG2sTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "i16", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", "arguments": [ + "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", + "i32", + "i16", "i64", - "i1" + "i1", + "i1", + "i32" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", + "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "Range", + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -21048,8 +22428,8 @@ } ], "tma": { - "operation": "s2g_tile4d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_tile5d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -21059,47 +22439,46 @@ "async", "bulk", "tensor", - "4d", + "5d", + "shared::cluster", "global", - "shared::cta", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_5d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", + "id": "cp_async_bulk_tensor_5d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_5D", - "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + "source_record": "TMA_G2S_CTA_TILE_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 10, - "value": 0 - }, - { - "argument_index": 11, + "argument_index": 9, "value": 0 } ] @@ -21107,9 +22486,9 @@ } ], "rust": { - "abi_id": "i0334", + "abi_id": "i1034", "module": "tma", - "name": "cp_async_bulk_tensor_5d_g2s", + "name": "cp_async_bulk_tensor_5d_g2s_cta", "arguments": [ "*mut u8", "*const u8", @@ -21124,15 +22503,15 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile5dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", + "op_type": "CpAsyncBulkTensorG2sCtaTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", "operands": [ "ptr", "ptr", @@ -21142,15 +22521,14 @@ "i32", "i32", "i32", - "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", @@ -21158,19 +22536,13 @@ "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -21186,7 +22558,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21203,21 +22575,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21238,11 +22610,11 @@ "targets": [ "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 5D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_5d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,131-byte source had SHA-256 da8bf49e04c4250626e97150345c704f33802066582d2e80d1c17dd58c0fa0b0; the 2,668-byte bitcode had SHA-256 d1d535b89cc23b4983c8fb5b00562e06266454177d8eb3a81684859ac5430cb5; the 1,358-byte canonical IR had SHA-256 e4862fd8477461d79f65871ae6dc8750d273a0e50911602e0bfea91bae76df7c.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21251,13 +22623,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 5D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,701-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 2f87ab7dc1ed68335e4ad29e4bcc7bd264779b33deb26f1d18e0c955f6bf8223.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21266,13 +22638,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 5D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -21283,9 +22655,9 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21307,11 +22679,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21322,11 +22694,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -21336,7 +22708,7 @@ } ], "tma": { - "operation": "g2s_tile5d", + "operation": "g2s_cta_tile5d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -21348,7 +22720,7 @@ "bulk", "tensor", "5d", - "shared::cluster", + "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes" @@ -21365,7 +22737,7 @@ } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { "id": "cp_async_bulk_tensor_5d_s2g", diff --git a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json deleted file mode 100644 index 90dc9dd32e..0000000000 --- a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-tma-cta.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "schema": 6, - "backend_profile": "rust-llvm-23.1.0-16696adc-tma-cta", - "backend_kind": "llvm_nvptx", - "llvm_revision": "16696adcd119e6ba9cc175207d984d7021211acb", - "backend_version": "LLVM version 23.1.0-rust-1.100.0-nightly", - "backend_sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", - "defaults": { - "llvm_results": [], - "concrete_llvm_results": [], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_90", - "ptx_feature": "+ptx86", - "status": "validated", - "declaration_attributes_canonicalized": true, - "runtime_validation": "unexecuted" - }, - "fixtures": [ - { - "id": "tma_cta_typed_sm90", - "coverage_count": 5, - "stages": [ - { - "targets": ["sm_90"], - "representation": "one LLVM module containing the five CTA-local TMA G2S declarations and calls", - "stage": "declaration_canonicalization", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "llvm-as canonicalized the intrinsic attributes. The 3,115-byte source had SHA-256 1a93a64ff20db145c5aaa3c16bc0adcb3a2c5e0892ab3b94c6d06ab878cea30a; the 3,520-byte bitcode had SHA-256 36af1844b0aeca593dd70f32e3102513da337d17197e96b31ab98948d0066747." - }, - { - "targets": ["sm_90", "ptx86"], - "representation": "typed LLVM calls for the five CTA-local TMA G2S operations", - "stage": "backend_codegen", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "The pinned LLVM emitted all 5 exact 1D-5D instructions with no extern calls. The 6,805-byte PTX had SHA-256 932c64f28fd32c73fb6486448f133795ecac5d41443c97410d1e47a3b9b0d8fb." - }, - { - "targets": ["sm_90", "ptx86"], - "representation": "LLVM-produced PTX retaining the five CTA-local TMA G2S instructions", - "stage": "ptx_assembly", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "artifact_kind": "cubin", - "detail": "CUDA 13.3 ptxas produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", - "tool_path": "ptxas", - "tool_version": "CUDA 13.3 V13.3.73 build cuda_13.3.r13.3/compiler.38244171_0", - "tool_sha256": "b1afa6add5264e7507863c1789ee97c9835e6dfecc8dbd3c3529fab36083f7da" - } - ] - } - ], - "matrices": [ - { - "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_1d_g2s_cta"]}], - "product_count": 1, - "fixtures": ["tma_cta_typed_sm90"], - "template": { - "id": "${id}", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", - "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", - "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "1d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, - "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i64", "i1"]} - } - }, - { - "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_2d_g2s_cta"]}], - "product_count": 1, - "fixtures": ["tma_cta_typed_sm90"], - "template": { - "id": "${id}", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", - "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", - "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "2d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, - "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i64", "i1"]} - } - }, - { - "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_3d_g2s_cta"]}], - "product_count": 1, - "fixtures": ["tma_cta_typed_sm90"], - "template": { - "id": "${id}", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", - "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", - "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "3d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, - "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i64", "i1"]} - } - }, - { - "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_4d_g2s_cta"]}], - "product_count": 1, - "fixtures": ["tma_cta_typed_sm90"], - "template": { - "id": "${id}", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", - "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", - "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "4d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, - "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"]} - } - }, - { - "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_5d_g2s_cta"]}], - "product_count": 1, - "fixtures": ["tma_cta_typed_sm90"], - "template": { - "id": "${id}", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", - "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", - "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "5d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, - "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"]} - } - } - ] -} diff --git a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-tma-cta.json b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-tma-cta.json new file mode 100644 index 0000000000..a4ea137c6c --- /dev/null +++ b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-tma-cta.json @@ -0,0 +1,252 @@ +{ + "schema": 6, + "backend_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "backend_kind": "llvm_nvptx", + "llvm_revision": "aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019", + "backend_version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "backend_sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "defaults": { + "llvm_results": [], + "concrete_llvm_results": [], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86", + "status": "validated", + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted" + }, + "fixtures": [ + { + "id": "tma_cta_typed_sm90_1d", + "coverage_count": 1, + "stages": [ + { + "targets": ["sm_90"], + "representation": "separate typed LLVM declaration probe for the 1D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_1d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,007-byte source had SHA-256 253d3ea39f9d26f5000f83febb9f72a15af2bed6ad0f668ab206976ce4c6d6ba; the 2,640-byte bitcode had SHA-256 9f42944557579c5c9c4e38716cce52607a07c51e5f2272393948fdc40cfb8961; the 1,234-byte canonical IR had SHA-256 835c2efc3b53d5333ec6a8260fcad29d5d0ce889194d0523feb57f7b22281fff." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM call for the 1D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,165-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 b0b4469f60fea3342f58e26eb4698834c6cb0714aac17fd3ffaed79cbd4f226a." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the 1D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "id": "tma_cta_typed_sm90_2d", + "coverage_count": 1, + "stages": [ + { + "targets": ["sm_90"], + "representation": "separate typed LLVM declaration probe for the 2D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_2d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,038-byte source had SHA-256 f0d47c1b6670854dfddce5bc45206d8f4ad16e9a05236dd4e1d4036b518666ea; the 2,648-byte bitcode had SHA-256 6ab53bfd2185eb8b6d7f7ffb3793d72cb7b48a07f61846a5317604700bdd3225; the 1,265-byte canonical IR had SHA-256 2c00613205683c093c5002f0f7347d3b2d6b82e46c5bc00072fb5a781f7d9d44." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM call for the 2D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,299-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 59fa9746f48c669cbff59faffea6039a796055f2721b8647a6331d78b5a26521." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the 2D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "id": "tma_cta_typed_sm90_3d", + "coverage_count": 1, + "stages": [ + { + "targets": ["sm_90"], + "representation": "separate typed LLVM declaration probe for the 3D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_3d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,069-byte source had SHA-256 159ab50347ad4f81c8a110df7de35db35da9eb099938874523e5ec10ea867c8c; the 2,656-byte bitcode had SHA-256 3b5f19c93cffadf279efad33b501519f09cacbc660784f195ab3d007b5bca854; the 1,296-byte canonical IR had SHA-256 223dd701edae4ecf145ca18afc659dd5a50785cf8e492f6ba5c7b7f8299dc030." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM call for the 3D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,433-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 e9bab5f8e2c2a697ca435f86eaff468984a03470ae4bf28f975c2c89fc6676b9." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the 3D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "id": "tma_cta_typed_sm90_4d", + "coverage_count": 1, + "stages": [ + { + "targets": ["sm_90"], + "representation": "separate typed LLVM declaration probe for the 4D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_4d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,100-byte source had SHA-256 becebd363b47d76e388fccf46646db4d8bba5225b4eb23cc7cae1673a5411ddd; the 2,664-byte bitcode had SHA-256 77f23ed44102d830fb8a576a8ff62f947e96b2b15871e2984083851d1e82f63e; the 1,327-byte canonical IR had SHA-256 6fd6a38abbb214d1e1fc5a4c8aa800ab963d1f53350afc3a7ec965501e467fb4." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM call for the 4D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,567-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 0607dfca39829f02dbda9367602cb27d28e6dc5dd8f5909a451fa4648442629a." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the 4D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "id": "tma_cta_typed_sm90_5d", + "coverage_count": 1, + "stages": [ + { + "targets": ["sm_90"], + "representation": "separate typed LLVM declaration probe for the 5D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_5d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,131-byte source had SHA-256 da8bf49e04c4250626e97150345c704f33802066582d2e80d1c17dd58c0fa0b0; the 2,668-byte bitcode had SHA-256 d1d535b89cc23b4983c8fb5b00562e06266454177d8eb3a81684859ac5430cb5; the 1,358-byte canonical IR had SHA-256 e4862fd8477461d79f65871ae6dc8750d273a0e50911602e0bfea91bae76df7c." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "typed LLVM call for the 5D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,701-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 2f87ab7dc1ed68335e4ad29e4bcc7bd264779b33deb26f1d18e0c955f6bf8223." + }, + { + "targets": ["sm_90", "ptx86"], + "representation": "LLVM-produced PTX retaining the 5D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + } + ], + "matrices": [ + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_1d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90_1d"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "1d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_2d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90_2d"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "2d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_3d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90_3d"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "3d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_4d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90_4d"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "4d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + }, + { + "axes": [{"name": "id", "values": ["cp_async_bulk_tensor_5d_g2s_cta"]}], + "product_count": 1, + "fixtures": ["tma_cta_typed_sm90_5d"], + "template": { + "id": "${id}", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", + "llvm_symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", + "expected_ptx": {"mnemonic": "cp", "modifiers": ["async", "bulk", "tensor", "5d", "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes"], "operands": [{"kind": "address"}, {"kind": "address"}, {"kind": "address"}]}, + "facts": {"llvm_arguments": ["shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"], "concrete_llvm_arguments": ["ptr addrspace(3)", "ptr addrspace(3)", "ptr", "i32", "i32", "i32", "i32", "i32", "i64", "i1"]} + } + } + ] +} diff --git a/cuda-oxide/intrinsics/generated-reference.md b/cuda-oxide/intrinsics/generated-reference.md index d9fa06d84b..6b5f87f27d 100644 --- a/cuda-oxide/intrinsics/generated-reference.md +++ b/cuda-oxide/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. --> @@ -97,16 +97,21 @@ This table is generated from the resolved catalog. Evidence stages below disting | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4 [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4.L2::cache_hint [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.1d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster.cta_group::2 [
], [
], [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-reduce`; expects `cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.4d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-reduce`; expects `cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | @@ -1388,6 +1393,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s` +- `cp_async_bulk_tensor_1d_g2s_cta` (`i1030`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta` - `cp_async_bulk_tensor_1d_s2g` (`i0335`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` @@ -1396,6 +1405,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s` +- `cp_async_bulk_tensor_2d_g2s_cta` (`i1031`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta` - `cp_async_bulk_tensor_2d_g2s_multicast` (`i0330`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` @@ -1412,6 +1425,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s` +- `cp_async_bulk_tensor_3d_g2s_cta` (`i1032`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta` - `cp_async_bulk_tensor_3d_s2g` (`i0337`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` @@ -1420,6 +1437,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s` +- `cp_async_bulk_tensor_4d_g2s_cta` (`i1033`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta` - `cp_async_bulk_tensor_4d_s2g` (`i0338`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` @@ -1428,6 +1449,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s` +- `cp_async_bulk_tensor_5d_g2s_cta` (`i1034`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta` - `cp_async_bulk_tensor_5d_s2g` (`i0339`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` @@ -5803,16 +5828,21 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cp_async_bulk_prefetch_tensor_gather4_2d_l2`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D_CH`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_1d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_1d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_1d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_2d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D_MC`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast_cg2`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_2D_MC`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_2d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_3d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_4d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_5d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_reduce_add_im2col_3d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cp_async_bulk_tensor_reduce_add_im2col_4d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none @@ -7280,6 +7310,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_1d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_1d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,007-byte source had SHA-256 253d3ea39f9d26f5000f83febb9f72a15af2bed6ad0f668ab206976ce4c6d6ba; the 2,640-byte bitcode had SHA-256 9f42944557579c5c9c4e38716cce52607a07c51e5f2272393948fdc40cfb8961; the 1,234-byte canonical IR had SHA-256 835c2efc3b53d5333ec6a8260fcad29d5d0ce889194d0523feb57f7b22281fff. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,165-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 b0b4469f60fea3342f58e26eb4698834c6cb0714aac17fd3ffaed79cbd4f226a. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_1d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7296,6 +7334,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_2d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_2d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,038-byte source had SHA-256 f0d47c1b6670854dfddce5bc45206d8f4ad16e9a05236dd4e1d4036b518666ea; the 2,648-byte bitcode had SHA-256 6ab53bfd2185eb8b6d7f7ffb3793d72cb7b48a07f61846a5317604700bdd3225; the 1,265-byte canonical IR had SHA-256 2c00613205683c093c5002f0f7347d3b2d6b82e46c5bc00072fb5a781f7d9d44. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,299-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 59fa9746f48c669cbff59faffea6039a796055f2721b8647a6331d78b5a26521. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_2d_g2s_multicast`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7328,6 +7374,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_3d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_3d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,069-byte source had SHA-256 159ab50347ad4f81c8a110df7de35db35da9eb099938874523e5ec10ea867c8c; the 2,656-byte bitcode had SHA-256 3b5f19c93cffadf279efad33b501519f09cacbc660784f195ab3d007b5bca854; the 1,296-byte canonical IR had SHA-256 223dd701edae4ecf145ca18afc659dd5a50785cf8e492f6ba5c7b7f8299dc030. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,433-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 e9bab5f8e2c2a697ca435f86eaff468984a03470ae4bf28f975c2c89fc6676b9. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_3d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7344,6 +7398,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_4d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_4d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,100-byte source had SHA-256 becebd363b47d76e388fccf46646db4d8bba5225b4eb23cc7cae1673a5411ddd; the 2,664-byte bitcode had SHA-256 77f23ed44102d830fb8a576a8ff62f947e96b2b15871e2984083851d1e82f63e; the 1,327-byte canonical IR had SHA-256 6fd6a38abbb214d1e1fc5a4c8aa800ab963d1f53350afc3a7ec965501e467fb4. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,567-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 0607dfca39829f02dbda9367602cb27d28e6dc5dd8f5909a451fa4648442629a. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_4d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7360,6 +7422,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_5d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_5d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,131-byte source had SHA-256 da8bf49e04c4250626e97150345c704f33802066582d2e80d1c17dd58c0fa0b0; the 2,668-byte bitcode had SHA-256 d1d535b89cc23b4983c8fb5b00562e06266454177d8eb3a81684859ac5430cb5; the 1,358-byte canonical IR had SHA-256 e4862fd8477461d79f65871ae6dc8750d273a0e50911602e0bfea91bae76df7c. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,701-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 2f87ab7dc1ed68335e4ad29e4bcc7bd264779b33deb26f1d18e0c955f6bf8223. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_5d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index e4f90dbf4f..22c75c57d5 100644 --- a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/abs_f16x2.ll b/cuda-oxide/intrinsics/probes/abs_f16x2.ll index d282bddf00..f05702a307 100644 --- a/cuda-oxide/intrinsics/probes/abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/active_mask.ll b/cuda-oxide/intrinsics/probes/active_mask.ll index cfb28da0a0..42dd121c29 100644 --- a/cuda-oxide/intrinsics/probes/active_mask.ll +++ b/cuda-oxide/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_bf16x2.ll b/cuda-oxide/intrinsics/probes/add_bf16x2.ll index 9e8b4b623c..6bc104a082 100644 --- a/cuda-oxide/intrinsics/probes/add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f16x2.ll b/cuda-oxide/intrinsics/probes/add_f16x2.ll index 9c9be96eed..db1855991f 100644 --- a/cuda-oxide/intrinsics/probes/add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f32x2.ll b/cuda-oxide/intrinsics/probes/add_f32x2.ll index 8bd18c88a6..81850b270c 100644 --- a/cuda-oxide/intrinsics/probes/add_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll index 15b70e709e..660aaf8dfa 100644 --- a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_f32.ll index 6db9ffcbc0..557fea27f6 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f64.ll b/cuda-oxide/intrinsics/probes/add_rm_f64.ll index 969c43674e..fe9b35b4d3 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll index aa286f0f62..bfa507a089 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll index b6b6565745..4d8160a13c 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll index a03d96f4a1..89c1657c54 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_f32.ll index 1c618b8bff..dc6097fd89 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f64.ll b/cuda-oxide/intrinsics/probes/add_rn_f64.ll index cdc0104e4b..ef25a8e97c 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll index 5af9974339..15db69bf9e 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll index 12355cdf98..55083142a1 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll index 31e4fc10a6..6bca2816d5 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_f32.ll index 2e8a575e27..c99c790038 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f64.ll b/cuda-oxide/intrinsics/probes/add_rp_f64.ll index 5e098e67bc..4423f406cf 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll index 6daf152ec6..94226648c5 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll index 731fb3e96a..8cd1f677a2 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll index 4922207ad2..142e1ff569 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_f32.ll index 1d6bc06f27..89efd5c6fd 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f64.ll b/cuda-oxide/intrinsics/probes/add_rz_f64.ll index 66bfa0676a..7d4309fdd6 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll index de5dedd64b..8fa33e1082 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll index 9182470492..01ba575d9c 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll index cceab23750..a26f5b0998 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/all_sync.ll b/cuda-oxide/intrinsics/probes/all_sync.ll index cb66ad10f4..448caf2240 100644 --- a/cuda-oxide/intrinsics/probes/all_sync.ll +++ b/cuda-oxide/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/any_sync.ll b/cuda-oxide/intrinsics/probes/any_sync.ll index 97b18cd57a..fde7a50a98 100644 --- a/cuda-oxide/intrinsics/probes/any_sync.ll +++ b/cuda-oxide/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ballot_sync.ll b/cuda-oxide/intrinsics/probes/ballot_sync.ll index 96081557da..c004732aa7 100644 --- a/cuda-oxide/intrinsics/probes/ballot_sync.ll +++ b/cuda-oxide/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll index 304516396b..f24bde55cd 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll index cfc169549d..124194fc83 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index c03df80219..d000befa61 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 08e8e7e515..542c7f34b1 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll index 4c91be89ba..06bb231758 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll index ad7e518834..2f74988842 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll index ccaed8aec9..20afac7dc8 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll index 97fb497976..04afa560ab 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll index 24e0cf60ac..2cf36f959b 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll index b1d298dd80..4a2aca105c 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_x.ll b/cuda-oxide/intrinsics/probes/block_dim_x.ll index ad845532eb..5ab6f284ca 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_y.ll b/cuda-oxide/intrinsics/probes/block_dim_y.ll index 3d78afddb1..72655eb3bc 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_z.ll b/cuda-oxide/intrinsics/probes/block_dim_z.ll index 7048c46326..a070dc03f4 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_x.ll b/cuda-oxide/intrinsics/probes/block_idx_x.ll index 2066c67b2a..fbffad8ccc 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_y.ll b/cuda-oxide/intrinsics/probes/block_idx_y.ll index 56b24a2ca6..4df5a14364 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_z.ll b/cuda-oxide/intrinsics/probes/block_idx_z.ll index 69ec9340ab..09d56bca9e 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/breakpoint.ll b/cuda-oxide/intrinsics/probes/breakpoint.ll index fee6f3fa0c..5c5467928f 100644 --- a/cuda-oxide/intrinsics/probes/breakpoint.ll +++ b/cuda-oxide/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll index 4920f565fa..bf7c783ce3 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll index d3dc26d79b..91433e0a1e 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll index 418a1c73fa..eba3308558 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll index aec4b27149..fa7566b9f2 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll index a44a1cfa20..de912d3cab 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll index 38ffb0ce73..c7ce3a2e50 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll index 56dfe3d6b4..1d6fa7f9c4 100644 --- a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll index aed88470f6..19d0d341c7 100644 --- a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll index f51209ab7f..940294df54 100644 --- a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll index 3148c69953..73121c473a 100644 --- a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_count.ll b/cuda-oxide/intrinsics/probes/cluster_block_count.ll index 39fcd3f775..d90a8b42ab 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_count.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll index 9384df8463..30beb12cfb 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll index 72edce8ff2..0de6c3fd98 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll index d3ee68d839..5c9b53d002 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll index b5e1ebc766..d603980e31 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll index dcb424b2de..33c350224f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll index 08cb319243..320431d4d3 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll index 0f89613939..cdbd4e1d36 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll index d3007ef56d..d0d53a9eb2 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll index be7a7a08aa..c2272581d4 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll index 51b8879cae..031bd8bacb 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll index b0f443649b..9454a1719b 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll index 69fc747df9..70b2bd6f0d 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll index c72d22ab0d..61ce34a055 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll index 0e3115de83..2606c44604 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll index 6e16ebc12f..a56bbecfc4 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll index 405c6b104e..d2a5850eb8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index a1bfe4e31c..d4caf5921c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 72da0da001..afd44930b4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index f9247ae03f..fbe7eba277 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index a49fab07b3..52e5d9ed12 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 734a7ae722..78e3c1117e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index a927796f63..3304feb633 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index 51a44d497d..732ea47b3a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index 5962e7a5a3..a12c241f4e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index b4e0d8cfd5..fa0d528aa2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index 42ba75f968..3a3bc93e7b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index 0c97d6057b..ede7a1bf5c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index dcf767ab90..7a7237c6cb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index 03600b606f..a089a31afa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll index fd4a2694ff..9b75d5a086 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 899b146871..13fe996e4b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 1f7da55135..994e80d1ee 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll index 560f306d9e..fae02897dc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index c990b39f80..c60553dbb3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index b2266ee055..1b3145914d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 8b9b870c67..96823baabe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 39f9357b2c..41d378e0d4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll index 7b669f098c..afa8455a52 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index 83dd948659..1c5ebdf882 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index f726d34eb9..7353d629c6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll index 2d7bc8dd40..eff7592546 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index bf08d0169b..52699a5158 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index 7515b63d4c..ffae735162 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll index 6af4b57c94..438093f702 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 07180a97aef8b4054fbd09cd86b90bfe27cb19a3e9c19ef2120aa0de5d5bb63d. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 8b3c1839f9..3a20538fdc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index 5c3dedff9e..1428b5fb32 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index 96feaf9b06..251dc5ef6a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 045c4388c4..312505000c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 35062c7b02..2b7520302a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index a529f2b873..3a3e914ce5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index d2977dfc28..6f40fec34b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index e18e0ecee3..a445dae82f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index 9471b296a5..257dbe4f56 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index b7ea799c65..cc29493543 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index cd741e34cd..888c4926e6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index 91d6f48365..5e01d90ea2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 07e0943450..20eb1bc1ac 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index ca6edd125f..6235a46125 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index 0aa4d4b10c..585bcf9d69 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index 0b1a2fe822..ad5841ee4b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index 3806624727..f37883a4ec 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index a922f25e39..2a26025a46 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index 0252d6f7e0..cfe092807a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 1b81e77aeb..5f8d9e35b9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index d5c2922c22..58a7dfb768 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 1486b3d709..074be49ad0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index 41dc2529ea..51f3e5b913 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index ae03b1fceb..266a905d8e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index 9619620a5c..1c27053af6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index 2dd47ef1bc..f5ce586791 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index b2793b839a..439ff09386 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index f59d0ab3dd..555ad5baec 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index bb49fcd1a4..3e0b615d9c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index a6c45925a5..d3f07485db 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index c885e75016..73522d3826 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index 6a1a53cdaf..5d6e87c7a6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 9a84e9a095..c78abd7393 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index d1ed727d67..56cca2d412 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index db43ebc6d7..0ab6ad5cf6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index c65c97f089..b522066a2c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index 180c30a72c..fd1a20a8bf 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index 9e49134245..1e7159e43c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index 6888b3f9fc..fe602942f9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 2727eb9958..487072d180 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index b26850ff5f..8b7ae43ec4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index 4a303f2f91..be1c7e96a5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index f8d523f2c7..85c1eb72cd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index f1104484c8..1232f397a6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index d25cd2aa46..f7d49c2382 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index c2fdb514b6..4e8ecbc060 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index a77f118b64..75f8e732f5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index a171fba27f..22b4993b9a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index 1538221ea6..d46eed99cd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index d3832a0185..5710ac2b7f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index 9d7a79214c..58778f1522 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 2b8f449c46..91bfe4eaa1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index ea00c76c5b..186c812e21 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index b259dd25a2..5e4900563b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index d485f438b6..fee9d07818 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index 2f33dd8e47..c7f33e3fee 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 6238356da1..6072e03954 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index a9cf6db63d..c49ad65204 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index d14f710cef..078ae8fd13 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 299203b666..1418decf8e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 8d20bef888..ffdfe59af0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index e3ea390a0b..2d1734a339 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index 7c785b4521..20571b8eb2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 5f0fadf127..32c641fe24 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index cfa609bfc4..46dd5f540b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll index d107ebada2..100d1f9e65 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll index 9a8167a5fa..ce36c37f48 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll index fd68791d7b..e9c68a1daa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll index 4d7856699e..8048442a7b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll index e002dcbaf2..6f7c12524b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll index c7dd69139f..25b22f0fef 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll index 73ea5503d5..e10f101783 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll index 378d524a1d..cc0c6470a2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll index 9ed6a52162..72c3855b99 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll index e285d27077..ff8cd71b86 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll index 0f8eb28414..0e217411fa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll index 0addb10c1a..64773c4221 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index 61f4b8c68b..b172541c99 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 99de175278..b3df3c2e79 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index d49656c4e0..60f4ce600d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll index 26ce456bc6..5b59eb64fc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll index 347c99ea5c..7ae12d53f4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll index e68eafcffa..210505c4f1 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll index aa9ccb605e..f5923f5fef 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 5feca63319..1f03339fbb 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index daab8d78d9..a1fcd253ba 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index 9e8688dea8..f9aa05c9dc 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 920123763b..6a68f8bbc8 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index 6e4b6697e5..d6f6d2444a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index 2c100257bb..dadd77c9a3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index 72f2068054..d09ec7a29f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index 798279c8b3..0b208cdaea 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index d90d9e288a..77279877cf 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index 2df0cc6f42..201563e358 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index fc62e942db..f541a66b16 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index 694cc51b18..5fdcd37e90 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index 1df9b892fc..45882334fe 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index 45cdaeb7a7..cb18c1355f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index b006511c9a..c47d68f479 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index cac2a150ff..f12c4175e4 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index fa711fd95f..2b88617837 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll index a1ddc82b97..434240fcf9 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index 764d2a6e71..f93950a912 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll index a5c282a4c9..59d2e34a80 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 6a946aed71..15c80e7bff 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll index 000b7054c0..bc70035b0d 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index f7a12ec46c..37dcac43b4 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 733dda6360..83d01fe815 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 5b6c24128c..0ba17e1ce7 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll index f8f39abb08..08f1d73053 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_f32.ll index e8f91ad578..d0c10b1566 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f64.ll b/cuda-oxide/intrinsics/probes/div_rm_f64.ll index 09f00888e4..8e5dd2663c 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll index 0ea9c58769..b669ec53c6 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_f32.ll index 228cc0aa87..a505743c46 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f64.ll b/cuda-oxide/intrinsics/probes/div_rn_f64.ll index bd90372579..4f51c9f531 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll index 985d40669e..63f90518e0 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_f32.ll index 264c51acac..9c885c48a3 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f64.ll b/cuda-oxide/intrinsics/probes/div_rp_f64.ll index 2a63516b60..5a0ebd2416 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll index 2afa9efd1d..0fded723ca 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_f32.ll index e0e3aa902a..c7a5e7c086 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f64.ll b/cuda-oxide/intrinsics/probes/div_rz_f64.ll index a0efcae75c..20af939ae4 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll index 656c119c87..f6565a1147 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_s32.ll b/cuda-oxide/intrinsics/probes/dp2a_s32.ll index c876a0c3b8..2552e3c735 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_u32.ll b/cuda-oxide/intrinsics/probes/dp2a_u32.ll index f7c3e2f850..9072debb72 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_s32.ll b/cuda-oxide/intrinsics/probes/dp4a_s32.ll index ac0329ea6f..bcb82f9d1f 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_u32.ll b/cuda-oxide/intrinsics/probes/dp4a_u32.ll index 9a6bbb84c9..9abe1882cd 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll index 693c323247..9ed3b9eec0 100644 --- a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll +++ b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index 89000848ba..75f90a8b0c 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index b0ae0bf16e..3bdfd77862 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll index 3ceb7bacc1..8058afafcb 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll index 56c243c066..edf8632c4f 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll index 2084941712..9db9e66331 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll index c26aa21244..c85a6c1453 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll index 57c4d39162..ab8fe7aac7 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll index af884de851..9da1e74dab 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll index dc1d7a4a5b..8cf07f8cbe 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll index b0d7997a38..d52bc8f975 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll index 7306e10b93..c325cceb21 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index ad98279bac..c314f1fde2 100644 --- a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index 54f2db264f..caa9468382 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index fd539cbdcd..2668eeabcd 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll index 07cdb52a2a..826bdc8bb7 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 1af7bc56ea..2f4f46971e 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index a1e3f909a1..e9790caf9e 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index b2b7da6b0a..2bbf20e3f6 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index d02c4efd94..d09629dccc 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index d63f549758..7fc8b5e11f 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index dbdb54654f..a986d4cb39 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index c63b505f08..03e56c88a4 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index a48d2cb019..2274f93cc2 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll index 0227a87cf5..ebb0ecd5f8 100644 --- a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_f16x2.ll index 3006ca5867..69e1f76e38 100644 --- a/cuda-oxide/intrinsics/probes/fma_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_f32x2.ll index 4178195aff..aee367833b 100644 --- a/cuda-oxide/intrinsics/probes/fma_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll index 1dab638241..3cac204bca 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll index 88f236546c..526c77c6c9 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll index e6ce5730d2..f4925604b5 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll index 69c4856a3f..bf198406c7 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll index 1f06311c75..7cbe757e91 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll index 85fc23c5e9..a4d097c36c 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll index bd78e7fec2..96e41dab8f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll index 172f5a198f..9eab5d1401 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll index 15c7ff5a84..818b9a706a 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll index d5d90f504e..f4ee34bbfc 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll index f6a0afb978..df944e33e0 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll index fe95c6605a..b61350ba89 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll index aa8de7458c..fb975cd360 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll index 7f8f6bcdb5..4d98ebc5e7 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll index 7675704ac7..c449d8edca 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll index abd1aa265f..754f683cf2 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll index c2c322df2a..44f57ce5cc 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll index 64bc22b5ab..dc084850d6 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll index e938739361..df62539eb1 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll index 39143c43a8..c35a86676f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll index 795d4c65bd..d8bfb0d310 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll index 4d0f422f23..2d8bec1780 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll index 22b01235a1..f76da2feea 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll index ed8b184bc3..2a431bd332 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 03b8c41659..4e603e3fb7 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll index ef0b109944..49e77a67cc 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll index bb61385a0f..8b19eb9c43 100644 --- a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll index 341ecde82b..237fc3f613 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll index c562b17969..45fbb4bf5f 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll index b90c665711..be91431981 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll index 2cbb8fa061..1ced567037 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_x.ll b/cuda-oxide/intrinsics/probes/grid_dim_x.ll index d46c40c165..dcce8d5b74 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_y.ll b/cuda-oxide/intrinsics/probes/grid_dim_y.ll index 00341c087b..a1c25a1d24 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_z.ll b/cuda-oxide/intrinsics/probes/grid_dim_z.ll index 7cfe56e7bc..02f49fdff5 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll index a2b3c2796c..3e47d5521d 100644 --- a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll index 4bf2bb73f3..bd406bd15d 100644 --- a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lane_id.ll b/cuda-oxide/intrinsics/probes/lane_id.ll index 3810a36af7..1d4b745281 100644 --- a/cuda-oxide/intrinsics/probes/lane_id.ll +++ b/cuda-oxide/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_eq.ll b/cuda-oxide/intrinsics/probes/lanemask_eq.ll index 20799aed61..5047ef6ad0 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_eq.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_ge.ll b/cuda-oxide/intrinsics/probes/lanemask_ge.ll index f29b87e61f..04f1297f08 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_ge.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_gt.ll b/cuda-oxide/intrinsics/probes/lanemask_gt.ll index 53a7542517..455c1aadae 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_gt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_le.ll b/cuda-oxide/intrinsics/probes/lanemask_le.ll index 924858d48a..5ab9b4fdd9 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_le.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_lt.ll b/cuda-oxide/intrinsics/probes/lanemask_lt.ll index 189d4d8182..6d90bbae05 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_lt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index 2ba6ad2898..5f408b68fd 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 6056d52646..803918f385 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 7eeffa64bd..05f665c244 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index c64347b04d..f0975ec155 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index 0333251475..b86fb0b5a1 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index 5f2ccf3356..dcad6d4265 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index 1819f1a48a..dfb1b3af01 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index fb06db92da..e8d4814b95 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index d610e0ceba..5db68b7995 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 4ddddf4817..7d198d1f58 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index e6ea8ec6e2..10e166855c 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 216a433048..bb7d6929c0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 3f3b32a4b7..134f2609dd 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 4c4f562852..6271b8abc6 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index 5437865344..0aadec29e9 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index 38c1e31d97..28c14e69e8 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index f823d1f56d..239fe1faa0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index 2c7a254706..9ead3e37c9 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll index d20f9b8100..de7861d7ce 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll index 3150cdcf51..c459a1039e 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/map_shared_rank.ll b/cuda-oxide/intrinsics/probes/map_shared_rank.ll index 17ee6027fd..6c3be6cf3c 100644 --- a/cuda-oxide/intrinsics/probes/map_shared_rank.ll +++ b/cuda-oxide/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll index c0c2733487..e45b6954dd 100644 --- a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_sync.ll b/cuda-oxide/intrinsics/probes/match_all_sync.ll index b076a21e01..64d6424904 100644 --- a/cuda-oxide/intrinsics/probes/match_all_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll index cfc9acc3e1..4dd9cbb891 100644 --- a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_sync.ll b/cuda-oxide/intrinsics/probes/match_any_sync.ll index e4e65e322a..1f34f2e120 100644 --- a/cuda-oxide/intrinsics/probes/match_any_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16.ll b/cuda-oxide/intrinsics/probes/max_bf16.ll index 77be96d793..db05412337 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_bf16x2.ll index 7150683ad3..bf1c1867ce 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16.ll b/cuda-oxide/intrinsics/probes/max_f16.ll index d12d32818b..e082a8b657 100644 --- a/cuda-oxide/intrinsics/probes/max_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16x2.ll b/cuda-oxide/intrinsics/probes/max_f16x2.ll index eb099b1f93..5f1e374ab6 100644 --- a/cuda-oxide/intrinsics/probes/max_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll index 9ce110e2e6..5d6e777d78 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll index 99431d33f4..1b09f6813e 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll index 5b041b2ded..ef07624785 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll index 3fce4e7b67..e38ca068f9 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index 5ccb70e93a..fdcd13f9a8 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index d1b271ce91..58edbdd1b2 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 0663c3c250..8e15a2f8ce 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 6c218589e8..5fd11083eb 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 289a6f955e..9c54f200c9 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index 3fa48796ac..7109c240a9 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll index 9239b9fe52..43cfbd8382 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll index 0da393f4ff..a692839b5c 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_f16.ll index f91129746f..6da1f602fb 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll index 4656697104..70cd3ded78 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index 974357c43b..9cf88a5331 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index 437daaa6bb..e563f2d82d 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 5d73b53155..65da1f724c 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index ab4af618e8..5c2061cb9f 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 4f558e905c..f6f393403e 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll index bd8265fe52..33bb7babea 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s32.ll b/cuda-oxide/intrinsics/probes/max_relu_s32.ll index 5e8e6ed0d7..1c208a2c5f 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_s16x2.ll b/cuda-oxide/intrinsics/probes/max_s16x2.ll index db5370ba75..f210671cf8 100644 --- a/cuda-oxide/intrinsics/probes/max_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_u16x2.ll b/cuda-oxide/intrinsics/probes/max_u16x2.ll index 6a2e6c31c9..6807d26946 100644 --- a/cuda-oxide/intrinsics/probes/max_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll index 03e42f432b..7110a3443f 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll index a73ef96512..2af08f61ba 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll index f053059555..be1be97ccf 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll index 04aefd53a3..41b35d900d 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll index ca09146c29..756af2f216 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll index 1b0d581793..2305c066ff 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll index aff83d50be..21c70f5f46 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll index 4e35deeb04..e43a23c267 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index 65fe02b980..4fe1b80dd0 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll index 3cf7ee744f..6704a787f1 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_init.ll b/cuda-oxide/intrinsics/probes/mbarrier_init.ll index 910ac349d8..b7483c4e59 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_init.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll index 2c7fae458f..ce543bcaad 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll index 0e21dc951d..f86daed164 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll index df497af274..72631f9c34 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll index 8374649481..7bcdfcdeab 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 128985b3ea..4c7304058e 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16.ll b/cuda-oxide/intrinsics/probes/min_bf16.ll index 3bbd2641ca..cbb4690428 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_bf16x2.ll index f57634c72f..5c3658c6b6 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16.ll b/cuda-oxide/intrinsics/probes/min_f16.ll index a944cfb4dc..0ab84d0b4e 100644 --- a/cuda-oxide/intrinsics/probes/min_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16x2.ll b/cuda-oxide/intrinsics/probes/min_f16x2.ll index f295381bb6..0f546aa5c4 100644 --- a/cuda-oxide/intrinsics/probes/min_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll index 3d9104e413..dfd984cae3 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll index b06fa76805..9b93732ff1 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll index 00d289910f..d967d8adbf 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll index b583273e3c..b9aa82f5e6 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index aa8217ccf9..4a0f1fd17e 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index 53ceb1f1a5..c197a01a64 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 037e2349a9..3968b53633 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index 153ba0e6d6..04fec34cf2 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index d2e8fa9463..db0f19956f 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index 27fd6d5ccc..3c091f240e 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll index c533c9544b..1a657290a9 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll index 20513a48e4..9cd78c860a 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_f16.ll index 1a5d56cda6..24c77655a8 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll index 16d01c1151..d77ddc12a1 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 1607ff0343..50672b4b2f 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index e65f385040..d94710799b 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 35b19c4fd1..207e88d559 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index f87889e445..ef17e34a36 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll index 653e8bf18d..e9227ac768 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll index c08d69fe08..6948ccec01 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s32.ll b/cuda-oxide/intrinsics/probes/min_relu_s32.ll index ab32ac384a..1219f67785 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_s16x2.ll b/cuda-oxide/intrinsics/probes/min_s16x2.ll index 2f9c14a0d2..d0dbe7c39a 100644 --- a/cuda-oxide/intrinsics/probes/min_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_u16x2.ll b/cuda-oxide/intrinsics/probes/min_u16x2.ll index 211614bc5d..49c37c57c9 100644 --- a/cuda-oxide/intrinsics/probes/min_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll index c4e6e964e5..0077b89619 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 3b46362412..fd80aa42a3 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll index 9e9415875f..5b88aae4e0 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll index 0c5a76f1d3..46dc722397 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll index 9486508752..0bcbbc2c15 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index 1ea6db4179..6f5ee9ae51 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index f343c80ca5..fc6f947505 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll index f147489e24..dda5519043 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index 7776405699..10658b63eb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 8e3187734e..19dd5ae3ec 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index 2f60560537..5d358fada9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 0107ddc2be..95c4e5e682 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 348ca5a36b..261978dc8e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index 044208ada0..2a6c11c1ca 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index da44fee12a..9d1bb41ffb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index cb86d24427..eb4d3913b2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 10859f722d..8dbf49f5b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index 0ba1467481..c848e0c9df 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 92879daa3a..2eee1dd31a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 07c76884b2..4649851dba 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 3fa61efac6..55c786e4d7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index a5032ec5cb..f271ff4cf5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll index 70397939ac..c3ad6feffd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index 39b5948174..50fcbc9846 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index f6f7f63e30..23788b4c5c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index 3038cd2516..2071535076 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index e7ed2d289f..964389d77c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 7db1a4bb35..2b8a4f7e01 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index 4f85a99466..f569a2fda6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index d2104e7fb4..cf4d3fbdd9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 8ca6cb55a3..8e03cd6e53 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index db9953f57d..d7bda206b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 87a760c4ff..6cd7e352ef 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index dfd197e064..0d3b3ecd55 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index 735b58026a..1d39a71df2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index eaa10620d8..881d5fc1b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index e44cce86e8..1ea82e15d8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 8fbc3419a0..2b8c996052 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index e3f1714a86..8cb87a89b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 76c3c32edf..9d0f00e9e2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index 3c8be80062..aa95263375 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index eee0fafb8f..199e2aaf64 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index fd52e2f349..93e717b25d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 3baa32de0e..4f8cfdd3f4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index add7cc4475..b39bbd31f0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index 9e7c95d969..7b3133fcdb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 4b7122373b..8de2266327 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index c4ec2080fb..29deee89de 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index 630f56bf62..bd6f5d1fc1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index fecfeab6b8..14a14a2993 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index 1a0bc3f45b..660e4f85da 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index bc46a827c3..46e2b7ccea 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index 951a05b0f6..89194d3674 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index 96c95b15a5..c2e764e77b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index b0b2bcc4a1..8f7cdd2974 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index 52db2b582e..fe321897b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index 84c6daf029..6ccac258cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index 022877ae35..e7d6cdcd3c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index 96c70b0d9e..6d1fada2c9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index 97fe74ef02..481da75e52 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index 69f1fabc6e..4d42704022 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index de067d5495..025c429574 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index ed9138286c..1522651fb2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index 1c775c4e24..c400dea514 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index 4cc5ed515c..dba9b759e1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index 6f432aa964..be0019ac64 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index 5141ec0786..c0e4301aa1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index ffccc1412f..243617e60c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index b4d7276108..44b454f6ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 1221b46de6..67c43b1f2b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index 779b69c37b..c68bc1f5ff 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index b51a697a76..a6f3e439b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index 767763cb26..2e7f838334 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index 92c8be40b4..45a9d81eb1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index 980aa082e0..0155176864 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index 2101d232ac..e219694466 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index 46d5d57412..411c8a8bcc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index 4bec7816f6..4a22d1584d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index c6f7ee644a..faec6219a1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index 38c8822a13..ed89625875 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 823495e8ab..261eb4a4eb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 00aa64998e..59d833b503 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index fd74bfbf88..fa2c2c433b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index 751d6de387..ef9c385462 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index e62f38b979..b7ae243457 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index 9bb91ff4be..c8e31165b3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index 6ac77efe23..7356926c5f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index a29c45e41e..e706da145e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 8b6865a36d..2484931724 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index f1a47b1431..0aa111bb26 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index c141a5ed75..2b07d65e1f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index bcd74f3058..d95ef9dc93 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 15c81cafb8..22d1d2918b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 1090808214..2067f3c98e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index d8a60527d0..35cc3aa8ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index 9a10044a3c..bc7bb21ec6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index c9839d36d0..8443b6dd66 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index 0ad35af407..c73edae0bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index b3433362fc..7db34fcea0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index d009f0a87d..7d62c8bc52 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 7f33262008..7942ba545d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index c3de14683a..1cea7459a3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index db4ba8e839..d59cc377d7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index e220edb8f2..ca23ddcde9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 7900e8e9e6..5b64dbb711 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index f879eaf258..7b7cce682d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index 7afaf17ef0..574fb486dc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 6b40349352..895269fc56 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index 699d72600e..feb9590963 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index df4a91ccd7..dd4ca4a42e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index 7af2a65bb3..152f4f02c0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll index 7b171cbe43..274cb21c55 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index 7eb4d7daae..1021d5bcbb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index c6017a7c15..bbb55c1f02 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 051cc6c569..9806945eff 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll index b7f1b1db9e..d7b2cd90b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index b206c7932c..6492380e40 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index ac519e39fc..68d814d7c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index d4f89e99bb..df1c9cddfd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll index 209a3d9989..fed6ef6127 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index a681419efb..6c6b47052f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 5c96b8193e..091017fc77 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index cecdf1f30f..21e9e7a5a0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll index 7c350cf31b..264b567276 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index 0bebe9da87..9806ca7a66 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index 980f0166b2..b102d27bdf 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index 0efe91ff83..e35b05fe13 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index 6e70e7d1ec..e06cbeaa83 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll index e9890ce30d..73fb61c7b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index 9ee23211c7..9a357be363 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index 14d8201f7e..d9b582f55f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index 19343fb6de..1be1a63244 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll index 8826984e26..cadd11116c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index 3344055106..d13841e3bf 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index a33954d5ee..2f58e6ff99 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index 934c12dff5..e7717999e4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll index b021c14fac..353abfb5b5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index e64504fff5..076e37fd64 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 7caff5c91c..c7fac51337 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index d030101330..12ecca3203 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index 473c2e79de..c1bc8df8d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 3a432cfb15..5304775a25 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll index adddb22d13..54527b33c0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index 4a8d18d3da..d349bc9d25 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 054ea5e279..b865520e57 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 9bf0b3be33..4d0b6c40b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll index 50bc3245d2..79db5d385d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index 6e4ffb7eb5..fc3eb2907f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index e2160e822c..bbc1e54413 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index dd72002d02..10038b2094 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll index b1d3623e4d..b0e0ddd869 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index 2ad3a82c38..4c611e8f8f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index 530fd956cb..e347edd3c2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 53b117b1c6..494c3640d6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll index 7b524c3780..ec13655a77 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 598c261e2d..102ee68860 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index b69f73456e..91cc530325 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index d3892576bf..e5a0d710a2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll index 2e369de46c..9a0972aa4c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index 96d692c1ac..b8af8a4f4d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index 87cb08ca39..aedff8fe82 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index cfe144b4f5..3b90323bc3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index 07e3b09849..a81e346a1b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 54beec52dd..8e5d43894b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index 3ae0c3ab34..7e47db81c5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index 33b8a3e468..a043e62730 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index 3d7fdf7538..234953e449 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index c476d84a31..ba660953d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 7114cc098e..7db0f082f9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index ee3ac32bed..4f14f5dbce 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index 5cc5b3a4f4..bfada891bf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index 97dda1e812..cf8400d86e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index f05b0a4760..0f9baa8a5e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index 00a0005305..baf7438f5b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index f60dcffd1b..23939b8515 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 7352440cd0..919c9152fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index 64747e411f..bb5fb3798f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index 99f94d3258..9422c377d1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 5cebce211f..3ed16d0737 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 1a2729aa84..1e7b5cf0de 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index 4ffcd5c1b4..fc068dd512 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index 74262a9bd4..c77801ff82 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index c798abd261..dfa24c9725 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index f685ca86ab..59cb522573 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index 6d046a2fa9..731e5ac301 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index b0be097754..d4e98bfc41 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index 1f9c965c2d..31aa2569a7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index 4240c1d1fa..15be9a1895 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index 3011a63159..634b4de537 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index 82307523e8..b2298b4786 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index 35d18671bf..8840e0d960 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index d70ff12bf7..efb7e85706 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 83026e728b..557edc105f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 1598d12615..74dbcf3733 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 4332a7521b..95a6910cfc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index 684f3a2e70..479a2b5845 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index 21bb6d0c6d..1f52bdc1cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index c54884e450..99aaaee2de 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index cb2ddfa7aa..ada0ffc173 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 01817551d1..fade8cf58a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 525df8de25..e39f6abfe5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index b8fd3dde7e..996dbe6940 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index 31643145aa..565aee28e9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index 4c4533956b..9f040d901b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index b284e897b7..dc89194270 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index f0f1dbd4a4..904deac0c4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index be5dde36bc..f954ff465c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index 85056d3498..37db99db47 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index cbba60c4ba..861963f194 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index 15c849298e..ff876190ee 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index 8c42464754..b75b90b980 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index b1050defb6..4d108a2f71 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index d88fe3e3eb..59ffa3d3e4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index 0fd3ba60e1..bf5969ea56 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index 4d1c1830ae..7e5adfe19e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index b3dab246b3..d78adeb8ac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index c9e1a911e9..865ed77c10 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 021358fd1d..3a143163e1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 70b7321527..1e95077965 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index 4914fd926b..51302dfa38 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 6fd7141112..43252d5679 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index 6e85ddd82a..c51b7672f3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 02a23a1169..0c8d5fee4e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 618904c371..40e222a3ad 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 2d9b3994b7..726d51d545 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 456c9a0483..961c8067ba 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index c5f468014c..3a01802e62 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index 9fc3d9b536..cd9fe6286e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index 30466576ba..47cc75bf1e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index e56812c604..704620d22c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 83c991c331..1f777d0f24 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 5ba1ecb190..175aac78ca 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index c3fbfa9f94..d2e902677d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index d32068231a..533854175a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index da1a18a093..acaa0fb446 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index 4e2a6b84ec..6990466bcb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 60d12b5bba..8662c2d00e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index a0b26fc7e0..f9ff8eb12c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index 7f1f38bee5..25851abca0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index e51fcbf0e9..8bf84ebbd2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 78bffdbec0..f8c13e1af2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index eb254e0845..977056eb6a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index 961ec8b463..03a20a933b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index cccadb49d9..838cc7cf74 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index 13c28b8e69..ca9df15bac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index 87433ab492..0a2ebb5c6d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 377d250f54..85f9b9678e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 4274cd36df..3aa5e8d7ea 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 2f26e87cc5..17d3acf829 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 78a15d17ca..da6687a9d4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 3b5459eae9..08c9116ab8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 080082879d..6cc54042cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index 2973d38341..6a01c4971f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index 5fd43812e4..03b28035f9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index eb3ff37c98..de7e2b3232 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index 25662838f7..7bc8dbca4d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index e4554753f0..3489016362 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 86de360fe1..8703acd89b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index 017859b219..94f5c8cd77 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index e1e2970290..3c344b62ad 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index ede1412f6d..a19a847946 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 6e33033b3a..d5aaaa8491 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index ed5694e626..3f546f610b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index b034f065e4..e6d644d332 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index 3f4d201382..6ba9934e4f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index bca458a7c3..fb92595a29 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index 3a2cc89340..5bec047651 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 2173d4db76..5f1d6176ac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index 4a9fdedf9a..96a0c57fe2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index 20e4d53b8a..aa72f4833e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index ddbe82e11f..e0f69dfa9e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 6aff5c6f88..8ea2df6939 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 0421f0ffc6..0686cb1967 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index 1d59d7a529..e23e51179a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 1a9cce88d0..0c33ea62a9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 17d4263041..70668b88aa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index 9c8a310db5..261c676925 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index 491a194d31..83b0603519 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index 1d1a5e1c84..8c72eee160 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index d4e56c6b32..84ac5400be 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index 254cf0452f..efd541f6dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index a03a94023e..25741d3ae2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index f235abf757..887af17cc2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index 0e68928c1e..552d047b6c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index 10b9927d32..628572fbbe 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll index 204e98b791..93c710e4b7 100644 --- a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll index 726b7ea310..ccb6479f46 100644 --- a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f16x2.ll b/cuda-oxide/intrinsics/probes/mul_f16x2.ll index 092afdea4a..887e069f78 100644 --- a/cuda-oxide/intrinsics/probes/mul_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_f32x2.ll index 7a4daaeeb6..d8bd454149 100644 --- a/cuda-oxide/intrinsics/probes/mul_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll index e0e2b089d4..96c9eabc25 100644 --- a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll index 1e194d6780..30816998c2 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll index 513893e676..5d25b27bd2 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll index 86b0f26e72..654ef5c4d0 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll index c936466501..8ddb2deada 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll index d6b9062a8a..ded2a2550a 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll index e43e931133..851a9b7723 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll index 98b013e2c0..c6496c2dfc 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll index a0b25e06c2..7f2e7256f2 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll index b5932f240f..d7f0800313 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll index 1edbbaaef0..8e8f68927f 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll index a8f00b6bd3..a2342939eb 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll index 9fa2605615..1d6b2645d8 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nanosleep.ll b/cuda-oxide/intrinsics/probes/nanosleep.ll index ef3c013ae5..cfbef4a781 100644 --- a/cuda-oxide/intrinsics/probes/nanosleep.ll +++ b/cuda-oxide/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll index dc2600aad3..0ffa3e0736 100644 --- a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_f16x2.ll b/cuda-oxide/intrinsics/probes/neg_f16x2.ll index ee5840d2d9..cadb800301 100644 --- a/cuda-oxide/intrinsics/probes/neg_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll index 7d0749f632..72afddc32c 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll index 69cffffa8a..bb4cee13f4 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll index b9f83e5cb8..f707f5154b 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll index ee1c99f101..c294e0a568 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll index 1111cab0e8..d32ec4c9df 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll index 0ec238e0da..b100e2bbd6 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/pmevent.ll b/cuda-oxide/intrinsics/probes/pmevent.ll index ff9af628e0..2688caca3a 100644 --- a/cuda-oxide/intrinsics/probes/pmevent.ll +++ b/cuda-oxide/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll index dc36ea55be..f8fe3e88c2 100644 --- a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt.ll b/cuda-oxide/intrinsics/probes/prmt.ll index e7d238a19f..6089a48da6 100644 --- a/cuda-oxide/intrinsics/probes/prmt.ll +++ b/cuda-oxide/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_b4e.ll b/cuda-oxide/intrinsics/probes/prmt_b4e.ll index a1218e54b3..e3f84c4991 100644 --- a/cuda-oxide/intrinsics/probes/prmt_b4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecl.ll b/cuda-oxide/intrinsics/probes/prmt_ecl.ll index bbb2e73e77..6f81144609 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecl.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecr.ll b/cuda-oxide/intrinsics/probes/prmt_ecr.ll index 48b9949bd5..27b3245aee 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecr.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_f4e.ll b/cuda-oxide/intrinsics/probes/prmt_f4e.ll index bbcd5d8d54..91d513b69d 100644 --- a/cuda-oxide/intrinsics/probes/prmt_f4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc16.ll b/cuda-oxide/intrinsics/probes/prmt_rc16.ll index aae7467217..b8953a6771 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc16.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc8.ll b/cuda-oxide/intrinsics/probes/prmt_rc8.ll index 8830088f2e..023cb5aa25 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc8.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll index cb9e316c01..07ff28f22a 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll index ad7ad7aaef..b21318cddc 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll index e16fed6685..839b1d26ef 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll index e52fc61d12..85be84f8b1 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll index 818212973b..c5df1f1172 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll index e0357d9a35..7db89bd26d 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll index f1cf96b1e4..285c55022c 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll index d534303ff7..9c43f1c8f4 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll index d116a4b0fd..15e5bfcc34 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll index 4e6231b9b1..f6a53d62ba 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll index 91617991d7..c4ad38e6c0 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll index 42337793ae..923fd5bef0 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll index 211d1bb57f..faa83b9b4c 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll index ce89a31458..6b6ab37629 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_add.ll b/cuda-oxide/intrinsics/probes/redux_sync_add.ll index 0c2e753bb3..e6a43e9c8a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_add.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_and.ll b/cuda-oxide/intrinsics/probes/redux_sync_and.ll index b8c214ab64..0a1756b9e7 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_and.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll index 71ed83b2fc..51c36d8914 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 26e5f081cb..518743e908 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll index fbc23a8b97..d9e9c8d9ee 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll index 119729a5fb..e34284de22 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll index 602124e3a7..13004abce0 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll index 26e1140870..3d1f655cbf 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll index e9770680f2..34afa8d841 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 0ca86b7845..4435152d52 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll index f96e6ddf2f..251b1e8fdb 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll index ffe5fd9168..b98fc2784e 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll index ee0fc0d5a4..8dca4e944a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll index 457e80576a..1414367cf3 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_or.ll b/cuda-oxide/intrinsics/probes/redux_sync_or.ll index ed161fd7e1..f41f3f9209 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_or.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll index c03c50cf02..f3d621bcf9 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll index e4d7228376..8e14421257 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll index 4f48c804f1..c497f14225 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll index 828ad30b00..bcf761b484 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll index d2cba3359d..c55691a78b 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll index e1953f3cfa..474925191e 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll index db2087224f..81ebd917e3 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll index 02a95a332e..04c0c5ca57 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll index 6b0e005a30..a4ecb95a64 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll index c627065725..0263f8f204 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_sync.ll index e6f72c28ae..8feab8c6ba 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll index 2a7a9a7a70..46ba3a1f2b 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll index c14e1bc459..06927e75cb 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll index bb43fd0b6d..fe5ed42e22 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll index 13be404b92..abfe57c1df 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll index 37030a5272..e845d30470 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll index b25cc907c1..9a89c9db6b 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll index 46b9532127..3429d65598 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll index 56238c93d7..9ef2269cdf 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll index e46f7410da..9de516e3f1 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll index 4d5c293ef1..1aeb3808a2 100644 --- a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll index 597b0dcc44..42a3ae72e3 100644 --- a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll index 66cae7bf2d..3dee05eb88 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll index f30d3a602f..4efe659c3d 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll index e4f66f1f98..693a8621dd 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll index 2ffca499c0..035c300b0a 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll index ae698906d1..73c7495ee5 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll index a0f3e67b69..f5e0fcafea 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll index 84f69a4c2b..61ebdf30c4 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll index 2d18411bf8..e9ee3e5195 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll index e582bd4225..453ab3634c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll index e36ad71f3d..a5c09b6ef6 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll index d77884c98c..960fd3dfed 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll index 08308e002c..ef80ae6040 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll index abe907ac33..973f251ded 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll index e6498b0c18..0548c4b775 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 8b1fbcef6e..d9b0a35832 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index 32b43dd8c2..995874eb30 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 5db2851f97..83deea6de7 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 4f7d355392..2a43e14ca8 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll index 8970e79012..0d3df158de 100644 --- a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f16x2.ll b/cuda-oxide/intrinsics/probes/sub_f16x2.ll index d77114a98d..abc5a2bbb6 100644 --- a/cuda-oxide/intrinsics/probes/sub_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_f32x2.ll index 686258e982..48dedd923d 100644 --- a/cuda-oxide/intrinsics/probes/sub_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll index a402a2c3f7..fba9396876 100644 --- a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_mask.ll b/cuda-oxide/intrinsics/probes/sync_mask.ll index 683b1be50a..7d39ac816d 100644 --- a/cuda-oxide/intrinsics/probes/sync_mask.ll +++ b/cuda-oxide/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_threads.ll b/cuda-oxide/intrinsics/probes/sync_threads.ll index bfa844406b..782acc0e7c 100644 --- a/cuda-oxide/intrinsics/probes/sync_threads.ll +++ b/cuda-oxide/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll index 5f0d7e4493..2bc026ef17 100644 --- a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll index b823ac413b..9e426ace52 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll index dfb302cecb..ecfb457c13 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll index 779a70d70e..d757224555 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll index cb0f8096d4..31908a726a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll index 2e30ccc923..c725fb7e5b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index f92286e1f9..dd393d6993 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll index 59af03b61b..5603bb550e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index df4b9c0626..55e47e9c86 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll index a56fb45429..fce3746ee6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index 31a987f4e0..21d2b928ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index 1b5002737d..cf5480440d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index b2ff7cd266..6ad52ec2a3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 5a92491c7c..33d066f9e5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 40c0f29f78..69a8be76f3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index 12f23ad13d..79b136506f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index 09e05d6ef3..402206f242 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index 431b81da7a..c8abd2de4c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 09774f7173..0af9e40376 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index e6128ec64b..bbc0451a8f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index a729bc01d9..3c4d174ed4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index e63edfab81..489d1e088b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index dca98ba632..e1fede9e1d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index 3f05d3a6e2..d477f22827 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index c09e87a586..8c9209308d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll index 8d4d634cbb..e650245bf0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index 822e77604e..7bc1b9cad0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index 1d376316dc..8037a5e6b7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 272e8ef0e1..d7735bf157 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index bf72fbf488..1c9bee8b8d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 5b4bd94c8d..be9cbf8077 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 7837962693..7cc251b960 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 0c8ef2cfd3..61ab14a67a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index a78f8aa6d6..a1559f4107 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index a0b7e81098..11400ef08b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index 02a4d9562c..a5f40b211c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 14d660a9cc..d89e33b39b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index 8656971789..5f386d01ab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index 7bd6d95497..804faa5dea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index d646da8df8..2756e25059 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 82a1b27449..d97f8e4c75 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index 69ca6f3776..4cabb5222c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 82d525ada3..f900f76cea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index 21adeb6aec..b0ce19f2a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index f335cec4d1..6a8f751c55 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll index 34170acc77..f85928e972 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll index 6c4b3c8e84..24ba197fdc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 463b96a822..4df227d8cc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index 7eb0ca8654..e4f50ce3df 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index 5ae9bd0f15..fd5abaa846 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index 586087f6a9..1e85632e7d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index fed4990a1b..e6b75d13c5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index 647f31d933..737ccc389d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index 650cea233f..b8a5638417 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index 3c1b0e92fd..764e135816 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index 910354cb3f..ac267979cf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index eca0f558cb..51dd5eaf39 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index e9bbbe0942..31f425756e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index 0231c8cee7..b10a37a164 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index 18d52449a3..4148f67e4a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index 52c3931b6d..7e06f46c65 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 5c14240d1c..e4633628d1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index d5780d2b58..0e1eb56461 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index e473742565..aec768bc94 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index 4446a354ea..88acff2bc9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index b45bcd8f64..edbb0849cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index bfd93db9ea..6a0af12de8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index 61e93d7b86..0f2629f72a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index b970b32f99..2e865247f4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 5e6ec4657f..b50ed4ddca 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index f869e87f30..07abd729d3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 197489f82c..01dd2998d8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 070622fe59..ba20cf8a28 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 03fb61e7fb..95bfd49fb2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 4a51274d3b..14f6b3a1ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index 90cb0d5fe4..6c440250a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 2a94f9813d..c5cb8baf7a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index 4e2ab19a27..b1df59ac7f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index b034aee05b..5bffe4aebe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index 5fe0a8efed..5522b69b5c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index 83cac7b1e1..9c33a63b82 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index 71aa4fc632..c1588cb3f6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 43fac69148..9cb4b58d5c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index 22f0043a56..d6b411d814 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 261adbdbc9..2eb66024af 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index e5f9e24f90..38ed797246 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 02be0e8cad..13c7606ccb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index e74d6957ac..f4ce146c59 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 371d6f78b0..9f2519ff32 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index 71c6cc8313..666ca70830 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index 722bbb9938..b079fa0c34 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index 68d013a0dc..599f83de66 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index d9db4b69fd..2fb81fb674 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 578be1d4dd..3139f585e0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index 730e54dbfb..e04d72393e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index 71c61e2deb..8469b8ab44 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 183b1767e6..82669815a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index bb25fecc3c..84dd7fbdaf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index e527fb2524..a5b9920669 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index 15dcf1f873..a6fc0a6a95 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index f59cd79143..e2785354f8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 8947077001..9eba9bb84b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index 23f0103b48..ebc358fe30 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index e2003259a6..53e6e8f82e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index 83f8e8f89c..c428cf74db 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 01c6816cfb..9582d2da89 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index 1f54053c1c..93e1f17a8b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index 8c24485dba..7d8e6b45bb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 4165c98728..50e554379f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index dcf3eb1268..249c70b23d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index bd0569970f..fd719613f3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 830163dda2..61e0134fe0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 41bf253cc1..4d3b637caa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index fa2725a12d..060aa01112 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index 89ca9a137d..8491197808 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index f290ab329c..fed611b36e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index f1bbfeef13..c2e1ef6832 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index 1f4b9e89b4..9ee5e83e39 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index e46f0ca74e..9529df77c6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index a7d62aaf59..b46f73307c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index c94cd41efc..ebbed8efb3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index dbf59bafdd..381c7762da 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index 6dead57bca..0f3416adc3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 9ea197a58a..2cb5c40793 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 74df38e7c9..f31f2f0632 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll index d3fd44cf7b..d01db69f1f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll index 2c1773e38b..2502e6114e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll index e975159ebf..3653aba278 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll index 54fbc8fca1..0eaf5cefa8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll index a160caea4e..4c79b78dab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll index fa6921722a..8b55a03290 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll index 8de63e138d..48a0025ed3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 2f264a9b12..d71d368b81 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll index 0f32015fb1..c674ca9109 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll index 35f7bc7167..8d6523dd25 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll index 9117cdfd9e..71e2929e83 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index 5facb4658a..26065d41c7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll index 748f593bc6..66ba1ea8d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index e3a2ddf422..de696158d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 6bb3038653..e227cca3bd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index aedbde911b..6e75c79401 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index c9211ebf93..3280019bf6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index ff4d689ed2..efda75b2a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index 6119c1b78a..a07c0f525a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 3587b55966..251eaa1315 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll index 3d1513573e..821bd62781 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll index db6f697ace..0d851cb5f9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 94ffd5f6aa..9ab2388ad1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index e9445ec605..dac0aa39aa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index e6d9cdf552..bc486b1c3f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index 13650f29e6..083f52c4d4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index 9881f97c0a..b8deac351d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 981f53b0f8..607db6e054 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index e80ecc14d9..6c492622f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll index a341605a9e..b7e67809ef 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index 61a4985f62..c02c3c8abe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index 405b42e46a..062364fb49 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll index b6cd43c4ed..b6e5f90be5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll index edf0f0f845..ef1f15226a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index ba656768e3..0388b6262d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index 325dcf58ec..bef6918b9d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 06939a001b..645e9d5bf4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index fe2ce75b77..a115cdeab0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 3f7835672d..09527c4636 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 9ce2e04316..08b9728159 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index e0399b1b80..4566d85d7a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index ea382a3e4c..bcea3b049a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index eb14712d1a..3a4407f174 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index cef55806f6..9f922b78f2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index cce1d76678..e0457e8ad5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index ae71ecb02d..011c4186b5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index 1309cf9a45..fcfdd5386b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 561ad2b226..6cd6472ad0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index 4a7a422acf..102ee1c3f4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 102e728e71..66afeb2014 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index 77e2d52b11..5fef1eeee0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 2d6e832e41..4b5f30eeb5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 8f76bfac3b..f188ef695d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index 8ed6bcc5aa..cc48c50c1e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index c3d4ae1412..e6ba731c0c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index 88b8146e7f..e1e653ec61 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index e7c2de1788..d87779c277 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 734f56972d..48e4683eaa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index 9368684e3e..3a7b3b4c93 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index cd978dbea9..0fd1258151 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index b452b20c0d..3e90947285 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 291d882495..f14b322892 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index 334d590622..6e8236d176 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index a1cc78d36b..b125d4fb6f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index 3fc331675c..6455b2b272 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index 351a96ea42..fb4d5268c8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index 33179a9019..cbd301cd65 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index c9e00cdf2e..e8e5843ec0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index 8969ab99f5..f7745999ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index 13ae0c1322..cdb2de9c6d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 5615fdc594..8943d86cb7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index ccc1986b6d..d14208da2f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index 76ba355871..6ecadfef8f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index 93ce5bfb45..abd195d2d0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 1d98697b7d..def46cdb6b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index 6ec802a078..9303d27bfa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index 54408310a9..a40ace1810 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index fce5aab34f..692958f0d4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 579190ce50..9123e8d811 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 945c41b3a1..58775f57ab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index 6c3d0cbc2b..a8d59e3580 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index ebe4861d94..72f728955b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index cc93a7369e..8af53799b5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 8d9474b696..6265704980 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index f5a16d9133..bd25989d24 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 693fc0b853..90527b7716 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index 2c73baa378..4927ad9af9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index 5d0cf33455..8ac21fe1d4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index 4f3dbb60ef..e86dbe0b34 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index 57ce4c0886..d673714d2d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index 5e85f3a09d..f7159562ee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index 9b709f422e..226884b03c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index 5a9c9373e0..f79457cd7a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index bee9bffa8c..4432602fe4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index 8c77ac107a..12e49a7878 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index c8cc9d74ed..feb7d5bbc6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 333f2b5205..25423f664b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index ac0554d50e..05cc163a49 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index d9eb449d59..12feccf87f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 610f0af819..c536310f02 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index 33a2ee4102..61c4d3a8de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 548bd59e8d..2768ccc918 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index baad23fe6a..4ae2194a9e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index 98d7bbbb7e..67aa3025b4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index 0ec73445c8..36c4b2a2b1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 5a3eb0984c..1387967b63 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index bca6fc8d35..2c4c88bb40 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index b811beb1f0..eb54aeb09a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll index 43185ffd6a..55dc3933b4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll index d0f0f31dd9..3f9bc48017 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll index cb489ffa7c..5f4269e53f 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll index b65e7f27aa..95a4a45603 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll index 02e4af589f..9bc1cc0423 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll index 8c289fca96..72d0906084 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll index 1e1f8ae263..d61623d712 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll index aee07673e4..8ee24ea81b 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll index ce6ff316b2..543fb4b31b 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll index e335b2b467..c11170ecc2 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index afb08c6673..b91c145697 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll index d63b34a14c..ccec4214ac 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_x.ll b/cuda-oxide/intrinsics/probes/thread_idx_x.ll index 406589b9ea..1d94d11e56 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_y.ll b/cuda-oxide/intrinsics/probes/thread_idx_y.ll index df4ebf993a..7e65898e09 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_z.ll b/cuda-oxide/intrinsics/probes/thread_idx_z.ll index 04e6cb9ae7..abce05daa7 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence.ll b/cuda-oxide/intrinsics/probes/threadfence.ll index 3e40aeed62..86d8442c25 100644 --- a/cuda-oxide/intrinsics/probes/threadfence.ll +++ b/cuda-oxide/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_block.ll b/cuda-oxide/intrinsics/probes/threadfence_block.ll index c75307a800..be7273dce8 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_block.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_system.ll b/cuda-oxide/intrinsics/probes/threadfence_system.ll index 826e9530a3..0d614edec0 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_system.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll index 4d92137b25..1a48217b60 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 5707a1d084..f59485847e 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/trap.ll b/cuda-oxide/intrinsics/probes/trap.ll index 6da7e92ab2..2e09de8887 100644 --- a/cuda-oxide/intrinsics/probes/trap.ll +++ b/cuda-oxide/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/uni_sync.ll b/cuda-oxide/intrinsics/probes/uni_sync.ll index a33eef123c..df1c4d05c7 100644 --- a/cuda-oxide/intrinsics/probes/uni_sync.ll +++ b/cuda-oxide/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll index 3106842b64..c2e3f4f1f0 100644 --- a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll index 5a050dcd77..bdab62f7f4 100644 --- a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll index e9bf779030..651e3114aa 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_fence.ll b/cuda-oxide/intrinsics/probes/wgmma_fence.ll index 0f230c57dd..ed11cb5c62 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_fence.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll index 25ac4f95cc..96d32d7b59 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" From 74609ee4de3a358bf83ac152cb9c9781f24e397e Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 19:28:07 +0530 Subject: [PATCH 06/17] test(examples): cover two-field and returned SharedPtr shapes in shared_memory_helper Nothing in tree exercised a struct holding two SharedPtr fields passed by value to a non-inlined helper, or a SharedPtr returned from a non-inlined helper and used as a store target. A third kernel runs both shapes over 64 lanes with partner-lane overwrites, so the final values prove the stores went through the scalarized parameters and the returned pointer. The code-shape check now also requires both Pair fields to arrive as ptr addrspace(3) parameters and select_slot to return ptr addrspace(3), and the README describes the probe and the checks. Signed-off-by: nihalpasham --- .../examples/shared_memory_helper/README.md | 25 +++- .../examples/shared_memory_helper/src/main.rs | 109 ++++++++++++++++++ .../shared_memory_helper/verify-code-shape.sh | 20 +++- 3 files changed, 147 insertions(+), 7 deletions(-) diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md index b764d8a0f3..922d21e73d 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md @@ -29,9 +29,24 @@ unsafe fn shared_ptr_probe(region: Region) -> u64 { `SharedPtr` lowers to an `addrspace(3)` pointer by its type, so the generic helper's parameter, nested in a struct, is one too, and the call -needs no address-space adaptation. `verify-code-shape.sh` checks that the -helper takes `ptr addrspace(3)` and that no PTX converts a generic address to a -shared one. +needs no address-space adaptation. + +Two more shapes run after that probe, in `shared_ptr_shapes_probe`: a struct +with two `SharedPtr` fields passed by value to the non-inlined +`write_pair`, which writes through both, and a `SharedPtr` returned from +the non-inlined `select_slot` and used as a store target. Each of 64 lanes +writes its own slots, then overwrites one slot of its partner lane through the +returned pointer, so the final values prove both stores landed. + +`verify-code-shape.sh` checks the LLVM IR and PTX the build leaves next to +this file: + +| Helper | Required code shape | +| --- | --- | +| `shared_ptr_probe` | first parameter is `ptr addrspace(3)` | +| `write_pair` | both `Pair` fields arrive as `ptr addrspace(3)` parameters | +| `select_slot` | return type is `ptr addrspace(3)` | +| whole module | no `cvta.to.shared` in the PTX | The launch supplies 2048 dynamic shared bytes. Four independent CTAs each use one thread. Nine launches cover three data offsets and three input values, @@ -100,7 +115,9 @@ RTX 5090, NVIDIA driver 580.173.02, and CUDA tools 13.3. | Dependency helper, all three inline policies | Same rejection | With no features, the `SharedPtr` probe runs after the direct control and -reports `PASS shared-ptr: 9 launches, 36 CTA data/barrier checks`. +reports `PASS shared-ptr: 9 launches, 36 CTA data/barrier checks`, and the +shapes probe reports `PASS shared-ptr-shapes: 3 launches, 384 slot checks`, +on the LLVM NVPTX route and with `--emit-nvvm-ir --arch sm_120`. Both Compute Sanitizer runs of the direct control reported: diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs index 44c1a0fd3a..6b25119cb0 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/src/main.rs @@ -67,6 +67,43 @@ unsafe fn shared_ptr_probe(region: Region) -> u64 { } } +/// Two typed shared pointers in one struct, passed by value. +#[derive(Clone, Copy)] +struct Pair { + a: SharedPtr, + b: SharedPtr, +} + +/// Writes through both fields of a by-value `Pair`. The call boundary splits +/// the struct into two parameters; both must stay CTA-shared pointers. +/// +/// # Safety +/// Both fields must point at live shared memory with at least `index + 1` +/// `f32` slots. +#[inline(never)] +unsafe fn write_pair(pair: Pair, index: usize, value: f32) { + unsafe { + pair.a.add(index).as_ptr().write(value); + pair.b.add(index).as_ptr().write(value + 1000.0); + } +} + +/// Returns one slot of a `Pair` as a typed shared pointer. The return type +/// must stay a CTA-shared pointer so the caller's store needs no conversion. +/// +/// # Safety +/// As for `write_pair`. +#[inline(never)] +unsafe fn select_slot(pair: Pair, index: usize, second: bool) -> SharedPtr { + unsafe { + if second { + pair.b.add(index) + } else { + pair.a.add(index) + } + } +} + #[cuda_module] mod kernels { use super::*; @@ -109,6 +146,37 @@ mod kernels { *output.add(block * 2 + 1) = completed; } } + + /// `SharedPtr` shapes a helper boundary must keep: a two-field struct of + /// typed shared pointers passed by value, and a typed shared pointer + /// returned from a helper and used as a store target. + #[kernel] + #[launch_bounds(64)] + pub fn shared_ptr_shapes_probe(output: *mut f32, seed: f32) { + const LANES: usize = 64; + let lane = thread::threadIdx_x() as usize; + let smem = DynamicSharedArray::::shared_ptr(); + // SAFETY: the launch supplies 2048 shared bytes, enough for the two + // 64-slot regions, and every lane writes only its own index and one + // slot of its partner lane, after a barrier. + unsafe { + let pair = Pair { + a: smem, + b: smem.add(LANES), + }; + let value = seed + lane as f32; + write_pair(pair, lane, value); + thread::sync_threads(); + // Each lane overwrites one slot of its partner lane through the + // pointer a helper returns; odd lanes pick the second field. + let partner = LANES - 1 - lane; + let slot = select_slot(pair, partner, lane % 2 == 1); + slot.as_ptr().write(-value); + thread::sync_threads(); + *output.add(lane) = *pair.a.add(lane).as_ptr(); + *output.add(LANES + lane) = *pair.b.add(lane).as_ptr(); + } + } } fn main() -> Result<(), Box> { @@ -175,5 +243,46 @@ fn main() -> Result<(), Box> { cases * BLOCKS ); } + + const LANES: usize = 64; + let shapes_output = DeviceBuffer::::zeroed(&stream, LANES * 2)?; + let shapes_config = LaunchConfig { + grid_dim: (1, 1, 1), + block_dim: (LANES as u32, 1, 1), + shared_mem_bytes: 2048, + }; + let seeds = [0.0f32, 17.0, 1024.0]; + for seed in seeds { + let out = shapes_output.cu_deviceptr() as *mut f32; + // SAFETY: one CTA of 64 lanes; output holds two f32 per lane, and the + // kernel uses 512 of the 2048 shared bytes supplied. + unsafe { module.shared_ptr_shapes_probe(&stream, shapes_config, out, seed)? }; + let got = shapes_output.to_host_vec(&stream)?; + for lane in 0..LANES { + let partner = LANES - 1 - lane; + // Lane `partner` overwrote one of this lane's slots: the second + // field when `partner` is odd, the first field otherwise. + let overwritten = -(seed + partner as f32); + let (want_a, want_b) = if partner % 2 == 1 { + (seed + lane as f32, overwritten) + } else { + (overwritten, seed + lane as f32 + 1000.0) + }; + assert_eq!( + got[lane], want_a, + "shared-ptr-shapes: field a, seed={seed}, lane={lane}" + ); + assert_eq!( + got[LANES + lane], + want_b, + "shared-ptr-shapes: field b, seed={seed}, lane={lane}" + ); + } + } + println!( + "PASS shared-ptr-shapes: {} launches, {} slot checks", + seeds.len(), + seeds.len() * LANES * 2 + ); Ok(()) } diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh index 8ffe705a18..e12ab06047 100755 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh @@ -3,8 +3,10 @@ # SPDX-License-Identifier: Apache-2.0 # # `SharedPtr` keeps a CTA-shared pointer typed `addrspace(3)` across a call: -# the non-inlined `shared_ptr_probe` helper takes one as its parameter, and -# no PTX converts a generic address to a shared one (`cvta.to.shared`). +# the non-inlined `shared_ptr_probe` helper takes one as its parameter, the +# two fields of a by-value `Pair` reach `write_pair` as two such parameters, +# `select_slot` returns one, and no PTX converts a generic address to a +# shared one (`cvta.to.shared`). set -euo pipefail dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" @@ -21,10 +23,22 @@ if ! grep -Eq '^define .*shared_ptr_probe.*\(ptr addrspace\(3\) ' "${ll}"; then exit 1 fi +if ! grep -Eq '^define .*write_pair.*\(ptr addrspace\(3\) %[A-Za-z0-9_.]+, ptr addrspace\(3\) %[A-Za-z0-9_.]+, ' "${ll}"; then + echo "FAIL: write_pair does not take both Pair fields as ptr addrspace(3) parameters" >&2 + grep -E '^define .*write_pair' "${ll}" >&2 || true + exit 1 +fi + +if ! grep -Eq '^define ptr addrspace\(3\) @[^(]*select_slot' "${ll}"; then + echo "FAIL: select_slot does not return ptr addrspace(3)" >&2 + grep -E '^define .*select_slot' "${ll}" >&2 || true + exit 1 +fi + if grep -q 'cvta\.to\.shared' "${ptx}"; then echo "FAIL: PTX converts a generic address to a shared one" >&2 grep -n 'cvta\.to\.shared' "${ptx}" >&2 exit 1 fi -echo "PASS: SharedPtr stays addrspace(3) across shared_ptr_probe, no cvta.to.shared" +echo "PASS: SharedPtr stays addrspace(3) across shared_ptr_probe, write_pair and select_slot, no cvta.to.shared" From d7a3e3ad0c7b090a7a73ce48f82cb2cbb4897b0b Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 20:12:03 +0530 Subject: [PATCH 07/17] fix(intrinsics): move the CTA-local TMA entries to ABI ids i1034 to i1038 Main assigned i1030 to i1033 to the bf16x2 to FP8x2 pack conversions after this branch merged main, so the five cp_async_bulk_tensor_{1..5}d_g2s_cta entries move to the next free ids above main's ledger tail. The overlay shard, the generator recipes, the ordering test and the ledger carry the new ids; the ledger stays ascending, with i1029 followed by i1034 until the merge with current main brings in i1030 to i1033. The generated outputs are regenerated in this same commit (catalog SHA-256 fa1e911b, counts unchanged: 581 operations, 555 generated, 1034 entries, tma 116) and the book stamp is refreshed at both sites. Signed-off-by: nihalpasham --- .../cuda-device/src/generated/async_copy.rs | 2 +- .../cuda-device/src/generated/atomic.rs | 2 +- .../crates/cuda-device/src/generated/bf16.rs | 2 +- .../cuda-device/src/generated/bf16x2.rs | 2 +- .../crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- .../cuda-device/src/generated/convert.rs | 2 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- .../cuda-device/src/generated/dotprod.rs | 2 +- .../crates/cuda-device/src/generated/f16.rs | 2 +- .../crates/cuda-device/src/generated/f16x2.rs | 2 +- .../crates/cuda-device/src/generated/f32x2.rs | 2 +- .../crates/cuda-device/src/generated/fence.rs | 2 +- .../crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- .../cuda-device/src/generated/grid_sreg.rs | 2 +- .../crates/cuda-device/src/generated/i16x2.rs | 2 +- .../crates/cuda-device/src/generated/int.rs | 2 +- .../cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- .../cuda-device/src/generated/movmatrix.rs | 2 +- .../crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- .../crates/cuda-device/src/generated/sreg.rs | 2 +- .../cuda-device/src/generated/stmatrix.rs | 2 +- .../cuda-device/src/generated/tcgen05.rs | 2 +- .../crates/cuda-device/src/generated/tma.rs | 2 +- .../cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- .../src/resolve/families/tma.rs | 10 +-- .../src/resolve/tests/tma_clc_cluster.rs | 10 +-- .../cuda-intrinsics/src/generated/abi_v1.rs | 22 +++--- .../cuda-intrinsics/src/generated/mod.rs | 32 ++++----- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 2 +- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 22 +++--- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 22 +++--- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- .../src/ops/generated/active_mask.rs | 2 +- .../dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- .../dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 2 +- .../dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- .../dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- .../dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- .../dialect-nvvm/src/ops/generated/tma.rs | 2 +- .../dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 42 +++++------ .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 2 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 2 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 72 +++++++++---------- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 22 +++--- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 2 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 2 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../src/generated_intrinsics.rs | 22 +++--- .../cuda-oxide-book/compiler/mlir-dialects.md | 4 +- cuda-oxide/intrinsics/abi-v1.toml | 10 +-- cuda-oxide/intrinsics/catalog.json | 24 +++---- cuda-oxide/intrinsics/generated-reference.md | 22 +++--- cuda-oxide/intrinsics/overlay/tma.toml | 10 +-- cuda-oxide/intrinsics/probes/abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/active_mask.ll | 2 +- cuda-oxide/intrinsics/probes/add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f64.ll | 2 +- .../intrinsics/probes/add_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f64.ll | 2 +- .../intrinsics/probes/add_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f64.ll | 2 +- .../intrinsics/probes/add_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f64.ll | 2 +- .../intrinsics/probes/add_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/ballot_sync.ll | 2 +- .../probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- .../intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_sync.ll | 2 +- .../probes/barrier_cta_sync_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- .../probes/clc_query_is_canceled.ll | 2 +- .../intrinsics/probes/clc_try_cancel.ll | 2 +- .../probes/clc_try_cancel_multicast.ll | 2 +- .../intrinsics/probes/clock.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock.llvm_nvptx.ll | 2 +- .../intrinsics/probes/clock64.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- .../intrinsics/probes/cluster_block_count.ll | 2 +- .../intrinsics/probes/cluster_block_idx_x.ll | 2 +- .../intrinsics/probes/cluster_block_idx_y.ll | 2 +- .../intrinsics/probes/cluster_block_idx_z.ll | 2 +- .../intrinsics/probes/cluster_block_rank.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_z.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_x.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_y.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_z.ll | 2 +- .../intrinsics/probes/cos_approx_f32.ll | 2 +- .../intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_cta.ll | 2 +- .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- .../probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- .../intrinsics/probes/cp_async_ca_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_8.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- .../intrinsics/probes/cp_async_cg_16.ll | 2 +- .../intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- .../probes/cp_async_commit_group.ll | 2 +- .../probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- .../intrinsics/probes/cp_async_wait_all.ll | 2 +- .../intrinsics/probes/cp_async_wait_group.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f64.ll | 2 +- .../intrinsics/probes/div_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f64.ll | 2 +- .../intrinsics/probes/div_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f64.ll | 2 +- .../intrinsics/probes/div_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f64.ll | 2 +- .../intrinsics/probes/div_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_u32.ll | 2 +- .../intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- .../intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- .../probes/elect_sync.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- .../intrinsics/probes/ex2_approx_f16.ll | 2 +- .../intrinsics/probes/ex2_approx_f32.ll | 2 +- .../intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- cuda-oxide/intrinsics/probes/fma_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll | 2 +- .../intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- .../intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- .../intrinsics/probes/fma_relu_bf16x2.ll | 2 +- .../intrinsics/probes/fma_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f64.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f64.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f64.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f64.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll | 2 +- .../intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- .../probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- .../intrinsics/probes/grid_dependency_wait.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_z.ll | 2 +- .../intrinsics/probes/gridid.lib_nvvm.ll | 2 +- .../intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/lane_id.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_eq.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_ge.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_gt.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_le.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- .../intrinsics/probes/lg2_approx_f32.ll | 2 +- .../intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/map_shared_rank.ll | 2 +- .../intrinsics/probes/match_all_i64_sync.ll | 2 +- .../intrinsics/probes/match_all_sync.ll | 2 +- .../intrinsics/probes/match_any_i64_sync.ll | 2 +- .../intrinsics/probes/match_any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16.ll | 2 +- .../intrinsics/probes/max_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/max_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/max_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_u16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- .../probes/max_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- .../probes/max_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/mbarrier_arrive.ll | 2 +- .../probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_init.ll | 2 +- .../intrinsics/probes/mbarrier_inval.ll | 2 +- .../intrinsics/probes/mbarrier_test_wait.ll | 2 +- .../intrinsics/probes/mbarrier_try_wait.ll | 2 +- .../probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16.ll | 2 +- .../intrinsics/probes/min_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/min_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/min_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_u16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- .../probes/min_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- .../probes/min_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- .../probes/mma_m16n8k16_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- .../intrinsics/probes/movmatrix_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/mul_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f64.ll | 2 +- .../intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f64.ll | 2 +- .../intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f64.ll | 2 +- .../intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f64.ll | 2 +- .../intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/nanosleep.ll | 2 +- cuda-oxide/intrinsics/probes/neg_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/neg_f16x2.ll | 2 +- .../intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- .../probes/packed_atomic_add_bf16x2.ll | 2 +- .../probes/packed_atomic_add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/pmevent.ll | 2 +- .../probes/prefetch_tma_descriptor.ll | 2 +- cuda-oxide/intrinsics/probes/prmt.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_b4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecl.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecr.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_f4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc16.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc8.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f64.ll | 2 +- .../intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f64.ll | 2 +- .../intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f64.ll | 2 +- .../intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f64.ll | 2 +- .../intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/redux_sync_add.ll | 2 +- .../intrinsics/probes/redux_sync_and.ll | 2 +- .../probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_i32.ll | 2 +- .../probes/redux_sync_max_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_u32.ll | 2 +- .../probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_i32.ll | 2 +- .../probes/redux_sync_min_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_or.ll | 2 +- .../intrinsics/probes/redux_sync_xor.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f32.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f64.ll | 2 +- .../intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/setmaxnreg_dec.ll | 2 +- .../intrinsics/probes/setmaxnreg_inc.ll | 2 +- .../probes/shuffle_down_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_down_sync.ll | 2 +- .../probes/shuffle_down_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_sync.ll | 2 +- .../intrinsics/probes/shuffle_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- .../intrinsics/probes/sin_approx_f32.ll | 2 +- .../intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll | 2 +- .../intrinsics/probes/smid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/sqrt_approx_f32.ll | 2 +- .../intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll | 2 +- .../intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll | 2 +- .../intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll | 2 +- .../intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll | 2 +- .../intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/sub_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sync_mask.ll | 2 +- cuda-oxide/intrinsics/probes/sync_threads.ll | 2 +- .../intrinsics/probes/tanh_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc.ll | 2 +- .../intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_commit.ll | 2 +- .../intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- .../probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_load_wait.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_shared.ll | 2 +- .../probes/tcgen05_mma_sp_shared.ll | 2 +- .../probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- .../probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_shift_down.ll | 2 +- .../probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- .../intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- .../probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence.ll | 2 +- .../intrinsics/probes/threadfence_block.ll | 2 +- .../intrinsics/probes/threadfence_system.ll | 2 +- .../probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/trap.ll | 2 +- cuda-oxide/intrinsics/probes/uni_sync.ll | 2 +- .../intrinsics/probes/warpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/wgmma_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_fence.ll | 2 +- .../intrinsics/probes/wgmma_wait_group.ll | 2 +- 1326 files changed, 1484 insertions(+), 1484 deletions(-) diff --git a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs index 318879e1f1..0d8056f0d9 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs index 2c06168531..28a9159c6e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs index 1962d80438..952ce614b3 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs index 3aae5a6142..5af0e44280 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/clc.rs b/cuda-oxide/crates/cuda-device/src/generated/clc.rs index 4918b2d05a..d2e5c87f2e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/clc.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs index ee7530ce4f..2498401d34 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs index 677a31bfd7..e6334ff9eb 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs index 273a8d2253..a8a8d62b16 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::cluster`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/convert.rs b/cuda-oxide/crates/cuda-device/src/generated/convert.rs index 17fddd7e93..3755d638c7 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/convert.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::convert`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs index 5c6a2ef0f0..d0aa02b068 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs index 76e006e07d..d77544cc03 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs index d58c92b45d..b111e230fd 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs index 18963fed82..b0d108710f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16.rs b/cuda-oxide/crates/cuda-device/src/generated/f16.rs index d568cb01ac..5313f1a3f7 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs index 9a1591173b..fc5e8b4b52 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs index 11664b325e..344c5042ec 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/fence.rs b/cuda-oxide/crates/cuda-device/src/generated/fence.rs index 34cc985181..be5a1f6d8e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/fence.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/float.rs b/cuda-oxide/crates/cuda-device/src/generated/float.rs index 60db1c80e9..7fcfd12962 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/float.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::float`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs index 7457da7e6c..d69892da44 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs index 21db68db5d..589f1a5561 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs index 3aa25b8e8a..3aadb615fe 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/int.rs b/cuda-oxide/crates/cuda-device/src/generated/int.rs index b56a33ff0a..be4e71cf5c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/int.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs index 1ce99c08cf..092aa51e01 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs index b1d777d9fe..1ecce0855e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs index 3e4104456a..f1c079df9d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs index 10c1f3fdc8..7e1851ac5a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs index 7bc30d4521..a08fd6c9b6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs index 1501078c06..66c2cd11cc 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs index cef0fb4d14..6d00f8c772 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs index eb485fe174..25c232aa36 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs index 31a91a170e..c613dcd028 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs index 748a2c0495..837aa98636 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs index 6594ee5105..6ecbea3dbe 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs index 5a34a464ff..f934c45f0a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tma.rs b/cuda-oxide/crates/cuda-device/src/generated/tma.rs index 7da6e40ca6..744d841543 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs index b9bc56377f..fa3f38b417 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs index 9c6207060b..4476fcf397 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs index b782f5003c..c0172af859 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/tma.rs @@ -415,7 +415,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "nvvm.cp_async_bulk_tensor_g2s_tile_1d", ), TmaOperation::G2sCtaTile1d => ( - "i1030", + "i1034", "cp_async_bulk_tensor_1d_g2s_cta", "memory.copy.async.bulk.tensor.g2s.cta.tile.1d", "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d", @@ -433,7 +433,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "nvvm.cp_async_bulk_tensor_g2s_tile_2d", ), TmaOperation::G2sCtaTile2d => ( - "i1031", + "i1035", "cp_async_bulk_tensor_2d_g2s_cta", "memory.copy.async.bulk.tensor.g2s.cta.tile.2d", "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d", @@ -469,7 +469,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "nvvm.cp_async_bulk_tensor_g2s_tile_3d", ), TmaOperation::G2sCtaTile3d => ( - "i1032", + "i1036", "cp_async_bulk_tensor_3d_g2s_cta", "memory.copy.async.bulk.tensor.g2s.cta.tile.3d", "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d", @@ -487,7 +487,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "nvvm.cp_async_bulk_tensor_g2s_tile_4d", ), TmaOperation::G2sCtaTile4d => ( - "i1033", + "i1037", "cp_async_bulk_tensor_4d_g2s_cta", "memory.copy.async.bulk.tensor.g2s.cta.tile.4d", "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d", @@ -505,7 +505,7 @@ pub(in crate::resolve) fn tma_recipe(operation: TmaOperation) -> TmaRecipe { "nvvm.cp_async_bulk_tensor_g2s_tile_5d", ), TmaOperation::G2sCtaTile5d => ( - "i1034", + "i1038", "cp_async_bulk_tensor_5d_g2s_cta", "memory.copy.async.bulk.tensor.g2s.cta.tile.5d", "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d", diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs index 8a0d197e03..42022d35c9 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs @@ -190,17 +190,17 @@ fn compact_tma_admission_matches_llvm_and_fails_closed() { .collect::>(), [ ("i0328", "cp_async_bulk_tensor_1d_g2s"), - ("i1030", "cp_async_bulk_tensor_1d_g2s_cta"), + ("i1034", "cp_async_bulk_tensor_1d_g2s_cta"), ("i0329", "cp_async_bulk_tensor_2d_g2s"), - ("i1031", "cp_async_bulk_tensor_2d_g2s_cta"), + ("i1035", "cp_async_bulk_tensor_2d_g2s_cta"), ("i0330", "cp_async_bulk_tensor_2d_g2s_multicast"), ("i0331", "cp_async_bulk_tensor_2d_g2s_multicast_cg2"), ("i0332", "cp_async_bulk_tensor_3d_g2s"), - ("i1032", "cp_async_bulk_tensor_3d_g2s_cta"), + ("i1036", "cp_async_bulk_tensor_3d_g2s_cta"), ("i0333", "cp_async_bulk_tensor_4d_g2s"), - ("i1033", "cp_async_bulk_tensor_4d_g2s_cta"), + ("i1037", "cp_async_bulk_tensor_4d_g2s_cta"), ("i0334", "cp_async_bulk_tensor_5d_g2s"), - ("i1034", "cp_async_bulk_tensor_5d_g2s_cta"), + ("i1038", "cp_async_bulk_tensor_5d_g2s_cta"), ("i0335", "cp_async_bulk_tensor_1d_s2g"), ("i0336", "cp_async_bulk_tensor_2d_s2g"), ("i0337", "cp_async_bulk_tensor_3d_s2g"), diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 7a928dc5b9..5cc3d8d944 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Raw ABI functions recognized by cuda-oxide. @@ -1191,9 +1191,9 @@ pub unsafe fn i0328(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u6 /// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. #[inline(never)] -pub unsafe fn i1030(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u64) -> () { +pub unsafe fn i1034(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u64) -> () { unreachable!( - "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` executed outside device compilation" + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` executed outside device compilation" ) } @@ -1243,7 +1243,7 @@ pub unsafe fn i0329( /// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. #[inline(never)] -pub unsafe fn i1031( +pub unsafe fn i1035( _arg0: *mut u8, _arg1: *const u8, _arg2: i32, @@ -1251,7 +1251,7 @@ pub unsafe fn i1031( _arg4: *mut u64, ) -> () { unreachable!( - "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` executed outside device compilation" + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035` executed outside device compilation" ) } @@ -1346,7 +1346,7 @@ pub unsafe fn i0332( /// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. #[inline(never)] -pub unsafe fn i1032( +pub unsafe fn i1036( _arg0: *mut u8, _arg1: *const u8, _arg2: i32, @@ -1355,7 +1355,7 @@ pub unsafe fn i1032( _arg5: *mut u64, ) -> () { unreachable!( - "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` executed outside device compilation" + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036` executed outside device compilation" ) } @@ -1407,7 +1407,7 @@ pub unsafe fn i0333( /// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. #[inline(never)] -pub unsafe fn i1033( +pub unsafe fn i1037( _arg0: *mut u8, _arg1: *const u8, _arg2: i32, @@ -1417,7 +1417,7 @@ pub unsafe fn i1033( _arg6: *mut u64, ) -> () { unreachable!( - "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` executed outside device compilation" + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037` executed outside device compilation" ) } @@ -1479,7 +1479,7 @@ pub unsafe fn i0334( /// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. #[allow(clippy::too_many_arguments)] #[inline(never)] -pub unsafe fn i1034( +pub unsafe fn i1038( _arg0: *mut u8, _arg1: *const u8, _arg2: i32, @@ -1490,7 +1490,7 @@ pub unsafe fn i1034( _arg7: *mut u64, ) -> () { unreachable!( - "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` executed outside device compilation" + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038` executed outside device compilation" ) } diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs index aa7ca75dd5..18e5581bf4 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. #[doc(hidden)] @@ -1094,11 +1094,11 @@ pub mod tma { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0984 as cp_async_bulk_tensor_reduce_xor_im2col_3d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0985 as cp_async_bulk_tensor_reduce_xor_im2col_4d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0986 as cp_async_bulk_tensor_reduce_xor_im2col_5d; - pub use crate::__cuda_oxide_intrinsic_abi_v1::i1030 as cp_async_bulk_tensor_1d_g2s_cta; - pub use crate::__cuda_oxide_intrinsic_abi_v1::i1031 as cp_async_bulk_tensor_2d_g2s_cta; - pub use crate::__cuda_oxide_intrinsic_abi_v1::i1032 as cp_async_bulk_tensor_3d_g2s_cta; - pub use crate::__cuda_oxide_intrinsic_abi_v1::i1033 as cp_async_bulk_tensor_4d_g2s_cta; - pub use crate::__cuda_oxide_intrinsic_abi_v1::i1034 as cp_async_bulk_tensor_5d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1034 as cp_async_bulk_tensor_1d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1035 as cp_async_bulk_tensor_2d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1036 as cp_async_bulk_tensor_3d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1037 as cp_async_bulk_tensor_4d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1038 as cp_async_bulk_tensor_5d_g2s_cta; } /// Generated `warp` intrinsic source API. @@ -1695,11 +1695,11 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] - fn public_cp_async_bulk_tensor_1d_g2s_cta_reexports_abi_i1030() { + fn public_cp_async_bulk_tensor_1d_g2s_cta_reexports_abi_i1034() { let public: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = super::tma::cp_async_bulk_tensor_1d_g2s_cta; let canonical: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = - super::__cuda_oxide_intrinsic_abi_v1::i1030; + super::__cuda_oxide_intrinsic_abi_v1::i1034; assert_eq!(public as usize, canonical as usize); } #[test] @@ -1719,11 +1719,11 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] - fn public_cp_async_bulk_tensor_2d_g2s_cta_reexports_abi_i1031() { + fn public_cp_async_bulk_tensor_2d_g2s_cta_reexports_abi_i1035() { let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = super::tma::cp_async_bulk_tensor_2d_g2s_cta; let canonical: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = - super::__cuda_oxide_intrinsic_abi_v1::i1031; + super::__cuda_oxide_intrinsic_abi_v1::i1035; assert_eq!(public as usize, canonical as usize); } #[test] @@ -1759,11 +1759,11 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] - fn public_cp_async_bulk_tensor_3d_g2s_cta_reexports_abi_i1032() { + fn public_cp_async_bulk_tensor_3d_g2s_cta_reexports_abi_i1036() { let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = super::tma::cp_async_bulk_tensor_3d_g2s_cta; let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = - super::__cuda_oxide_intrinsic_abi_v1::i1032; + super::__cuda_oxide_intrinsic_abi_v1::i1036; assert_eq!(public as usize, canonical as usize); } #[test] @@ -1783,11 +1783,11 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] - fn public_cp_async_bulk_tensor_4d_g2s_cta_reexports_abi_i1033() { + fn public_cp_async_bulk_tensor_4d_g2s_cta_reexports_abi_i1037() { let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = super::tma::cp_async_bulk_tensor_4d_g2s_cta; let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = - super::__cuda_oxide_intrinsic_abi_v1::i1033; + super::__cuda_oxide_intrinsic_abi_v1::i1037; assert_eq!(public as usize, canonical as usize); } #[test] @@ -1807,11 +1807,11 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] - fn public_cp_async_bulk_tensor_5d_g2s_cta_reexports_abi_i1034() { + fn public_cp_async_bulk_tensor_5d_g2s_cta_reexports_abi_i1038() { let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = super::tma::cp_async_bulk_tensor_5d_g2s_cta; let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = - super::__cuda_oxide_intrinsic_abi_v1::i1034; + super::__cuda_oxide_intrinsic_abi_v1::i1038; assert_eq!(public as usize, canonical as usize); } #[test] diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index e85efe8e33..9047c9399a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index 442a332616..9dd276df21 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index 143cea136f..be8d3bc297 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index ed83253c5d..41ea7c5470 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index a7af5344ae..0b6b9225cb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index 6e6cd6834a..3269e52f38 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index 132e42f57d..7796b61cec 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 19de0d5a88..8db7ff7060 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index c03ac79fbb..9ecb66da27 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index 23841b8439..00f35465cf 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index 6fb2e2ef85..da3e6cb9e0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index d99779ccdc..ab7c55e397 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index 548860aefb..7a99a17c37 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index 5c3eb52a1f..c46d6e8411 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index 67d3dece72..e45a256356 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index 4ba5f99824..81df262d59 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 7820e75597..0ef6f93356 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index 0605e8cc67..ede1706123 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index 597007a6a0..2a99ad8ec0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 3146ae0269..decac74602 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index 6919d0b0dd..d7f24a11bb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index f133c9c2a0..0be64b277b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index b191571ea6..4af0ad8d37 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 54094d2c3b..ae014afa2b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index 48e9e8e2d7..d20da7a716 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 3b658ccf7b..c1166effde 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index dc889eb5cd..c4e0264442 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index 14864ab794..ccaf7a28a5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index 78e53e6363..3c11bd67b0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index 63b3fbc03d..c9ba0fb037 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index c568a6a44e..65458d0e8c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index 3562de4fd2..35e0e37a55 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index db644a1558..e5363bbc45 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index 1ca21e6ee8..2881ca70d6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index 4204ab5120..8d063983fc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index baef4fc35d..450f349fe3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index b593775d06..ed91f72e7c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index 7f632bddd7..024a040339 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index c0f768cf0b..2ca30dd196 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index f6ca650831..dd64c00701 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index 5bf34eec50..1009b66d14 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index d89b227403..8d84be9a26 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index 914de29661..6c3dc12ca6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 069c13135a..6ecaef951a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index 125d097eb0..6a779da5cf 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 59175e4a3d..70b9062b06 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index 947e6bd8b3..2a82ee8953 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 070bcdb18e..fc9dcc1cd9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index 9942389730..babe36eb63 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index d9a27b2b3c..3a6f428497 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests, grouped per family shard. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index 08811cb41b..3db0559841 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 72c32ac38f..1e522d764e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index 50c6718100..7d00f904ae 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index c94dbe9d6d..edfb55d6c2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index 2e035456c9..4d13fdabe1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index 5438d3374d..7b62a154c0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 07d5a64bf5..683cd4e3a5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index aaa6d1adfc..548bab68f1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index 266c54837e..3947d7de47 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index 2b704752c4..244aaab7c2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index 42bd2becf8..18f0b6fefc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index 67c9b93ad7..6f894d4809 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index 8038b4d5fb..c67df314aa 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index 971a086586..3b0d63b54f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index 5d4ed2a92c..a96f358483 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index 61f2a0edae..acdb91b8f3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index f0f98ff0df..5d2f0f56ba 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 3831e729f1..6b5d1346b9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index b44991f20e..c2634d93fb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index 068adac85d..cad89ec2c7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index 4de3cb9a94..33ffbb770c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index 2a9631b1b9..85afacb3f2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tma` intrinsics. @@ -6646,9 +6646,9 @@ fn cp_async_bulk_tensor_reduce_xor_im2col_5d_target_matches_the_catalog() { #[test] fn cp_async_bulk_tensor_1d_g2s_cta_target_matches_the_catalog() { - let target = generated_intrinsic_target_by_marker("v1:i1030").unwrap(); + let target = generated_intrinsic_target_by_marker("v1:i1034").unwrap(); assert_eq!(target.id, "cp_async_bulk_tensor_1d_g2s_cta"); - assert_eq!(target.abi_id, "i1030"); + assert_eq!(target.abi_id, "i1034"); assert_eq!( target.dialect_op, "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d" @@ -6715,9 +6715,9 @@ fn cp_async_bulk_tensor_1d_g2s_cta_target_matches_the_catalog() { #[test] fn cp_async_bulk_tensor_2d_g2s_cta_target_matches_the_catalog() { - let target = generated_intrinsic_target_by_marker("v1:i1031").unwrap(); + let target = generated_intrinsic_target_by_marker("v1:i1035").unwrap(); assert_eq!(target.id, "cp_async_bulk_tensor_2d_g2s_cta"); - assert_eq!(target.abi_id, "i1031"); + assert_eq!(target.abi_id, "i1035"); assert_eq!( target.dialect_op, "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d" @@ -6784,9 +6784,9 @@ fn cp_async_bulk_tensor_2d_g2s_cta_target_matches_the_catalog() { #[test] fn cp_async_bulk_tensor_3d_g2s_cta_target_matches_the_catalog() { - let target = generated_intrinsic_target_by_marker("v1:i1032").unwrap(); + let target = generated_intrinsic_target_by_marker("v1:i1036").unwrap(); assert_eq!(target.id, "cp_async_bulk_tensor_3d_g2s_cta"); - assert_eq!(target.abi_id, "i1032"); + assert_eq!(target.abi_id, "i1036"); assert_eq!( target.dialect_op, "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d" @@ -6853,9 +6853,9 @@ fn cp_async_bulk_tensor_3d_g2s_cta_target_matches_the_catalog() { #[test] fn cp_async_bulk_tensor_4d_g2s_cta_target_matches_the_catalog() { - let target = generated_intrinsic_target_by_marker("v1:i1033").unwrap(); + let target = generated_intrinsic_target_by_marker("v1:i1037").unwrap(); assert_eq!(target.id, "cp_async_bulk_tensor_4d_g2s_cta"); - assert_eq!(target.abi_id, "i1033"); + assert_eq!(target.abi_id, "i1037"); assert_eq!( target.dialect_op, "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d" @@ -6922,9 +6922,9 @@ fn cp_async_bulk_tensor_4d_g2s_cta_target_matches_the_catalog() { #[test] fn cp_async_bulk_tensor_5d_g2s_cta_target_matches_the_catalog() { - let target = generated_intrinsic_target_by_marker("v1:i1034").unwrap(); + let target = generated_intrinsic_target_by_marker("v1:i1038").unwrap(); assert_eq!(target.id, "cp_async_bulk_tensor_5d_g2s_cta"); - assert_eq!(target.abi_id, "i1034"); + assert_eq!(target.abi_id, "i1038"); assert_eq!( target.dialect_op, "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d" diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 07de9882e0..94bffc21f5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index c6d9df42ec..3414a6d3c9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index 1fbc940cb7..96c2d29bee 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index b08c5491bd..7932441b5a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 14caa5adcb..4ba75736c5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index 1ace05c9e8..33c1a28230 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tma` intrinsics. @@ -5671,9 +5671,9 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ }), }, GeneratedIntrinsicTarget { - marker: "v1:i1030", + marker: "v1:i1034", id: "cp_async_bulk_tensor_1d_g2s_cta", - abi_id: "i1030", + abi_id: "i1034", dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { @@ -5729,9 +5729,9 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ }), }, GeneratedIntrinsicTarget { - marker: "v1:i1031", + marker: "v1:i1035", id: "cp_async_bulk_tensor_2d_g2s_cta", - abi_id: "i1031", + abi_id: "i1035", dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { @@ -5787,9 +5787,9 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ }), }, GeneratedIntrinsicTarget { - marker: "v1:i1032", + marker: "v1:i1036", id: "cp_async_bulk_tensor_3d_g2s_cta", - abi_id: "i1032", + abi_id: "i1036", dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { @@ -5845,9 +5845,9 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ }), }, GeneratedIntrinsicTarget { - marker: "v1:i1033", + marker: "v1:i1037", id: "cp_async_bulk_tensor_4d_g2s_cta", - abi_id: "i1033", + abi_id: "i1037", dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { @@ -5903,9 +5903,9 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ }), }, GeneratedIntrinsicTarget { - marker: "v1:i1034", + marker: "v1:i1038", id: "cp_async_bulk_tensor_5d_g2s_cta", - abi_id: "i1034", + abi_id: "i1038", dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index ccdd65c2fa..de35350db1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index 2559ea2ca8..2c3734b1be 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index d977330505..3008328d03 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index 4bfaba20e1..ee606a127e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index 4199dbf817..e34deb5eb0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 2ca1a95d6d..9cb178f2bc 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for the generated active warp mask. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs index 9773039800..4337fbe52a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Cluster Launch Control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index 1ec533f00b..dda0373b39 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated operations for the closed cluster-barrier family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index c1ad4b86f5..ec782c5778 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs index 25758a6591..09f6d23edc 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for classic `cp.async` instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs index 68869c4b6a..6679070112 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated PTX debug controls. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs index 41b5a740b5..5a26a3362a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed integer dot products. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs index a42f65a7a9..345bf58d1b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated warp leader-election operation. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs index 4efec68715..d62779f92d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index b30f35f1be..7d8f8004f2 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for extended floating-point min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index dc17124a6d..8d448385b1 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated extended integer min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index b575b3e8f1..7ec40842d8 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index 66e29afc37..98b95d2ecd 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index 1e629a1179..74ab4c6169 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs index 38d74b564c..5001c9a3d2 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. mod active_mask; diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index 3256807492..246ed7bc24 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated in-register matrix transpose. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index c616df1189..05047afb28 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed floating-point arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index e6095e28e0..2c319ab427 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated packed-atomic family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index 686f008188..79e2a85fd4 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated packed conversion. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs index 4ee1b29577..22a9e54297 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated `prmt` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs index b210840413..306a39934c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 97a01da4e0..b12aa971f8 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index 178890b226..7da01fa0af 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index f086e0cefb..6dc41def28 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index efbd87b1c8..ec4cb026c5 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar math. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 84b10189e9..421822712d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated sparse MMA variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs index d516b35ed9..94498fd183 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for generated special-register reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index 75e73d66fb..fae13903b2 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the four generated `stmatrix` stores. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs index dbbc71bf58..de51abf49d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated CTA synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index febb5d5d3e..2f7dc47afc 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Core Generation 5 operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs index 1e2cfc266c..f7fe98742b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Memory Accelerator operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs index 9eee8930ac..0c90388d36 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `vote.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 15cc9dfb5f..7e8ac13295 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated warp synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs index 5a9e9a05eb..1df3847aad 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `match.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 918913cd6c..7efe79a065 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `shfl.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index be46657dbf..8acf45328a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated WGMMA controls. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index 6b5617476c..85fc979bc8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index e8725fc71b..e67ee6defb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index 192d3b3cfe..60d9a2e0a8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index 37472270b6..9a1005454c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index ec5e550f53..af658bee76 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index 67937e6e70..b209869e35 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 7959b71a2a..7f8e081948 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 0425331510..9fd88ca3e4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 61726de0c4..532aca6fe4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 6de2f7b601..a22d61d266 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index 3ca948b7ef..114150b7a4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 7e1c55badd..5f67948437 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index 67dfb8596b..660b998ad8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index 1cb6e084fd..b1364973c8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index c1a3afbdbb..3c0a14254d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index a5b9506522..88fc211f45 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -313,13 +313,13 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" @@ -329,19 +329,19 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" @@ -2311,21 +2311,21 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3446,14 +3446,14 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" => Some("v1:i0328"), - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" - | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => Some("v1:i1030"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => Some("v1:i1034"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => Some("v1:i0335"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" => Some("v1:i0329"), - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" - | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => Some("v1:i1031"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => Some("v1:i1035"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => Some("v1:i0330"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -3462,20 +3462,20 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" => Some("v1:i0336"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" => Some("v1:i0332"), - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" - | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => Some("v1:i1032"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => Some("v1:i1036"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => Some("v1:i0337"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" => Some("v1:i0333"), - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" - | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => Some("v1:i1033"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => Some("v1:i1037"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => Some("v1:i0338"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" => Some("v1:i0334"), - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" - | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => Some("v1:i1034"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => Some("v1:i1038"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => Some("v1:i0339"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 826ec0a0c0..9093bc63a9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index f80f9ae917..0fcc0379f9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index 954d37ba58..c7a996ee03 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index f49208b6e6..d7af8ab630 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index d6157b608f..9069be4e1f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index 0e0210e837..3cba44a811 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index b3c9deec18..f3c2d7d9d0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 2f30a26ef7..6fba4e05ce 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index 44afc5b403..e658ce8abd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index 75a6ffec1c..62dc82340d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 85972feb8f..f0569cd607 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 951e3223db..9b4871de57 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index d8401d06c8..1548cb4ac2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index 9e39168315..a544678be6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 3650d4129e..5096f808d4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index db4c304654..7cb7e86708 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 7dd180761f..5cd8f86d68 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index 268d20d8b9..b86cf188d7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index 97cd1350cc..0c2b505e34 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index 2a0776d80f..50bf4c88fb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index b16e300858..42126b9eae 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index f96cf5828c..04d040e270 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index ebb760e4a8..bff4c2e96a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index c5ed120dc5..bebfda252b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index 81cd65f3bf..882cf767b4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index 0c3554e3e0..b4305c63cc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index ad267da768..c97da28a37 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index 36590deb45..cb6ced8be6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index 6f4948b624..705cf12556 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index b42f2eb906..33c8aaaa02 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index c89000a507..7dbae34529 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index fc3841f6ed..9e0f0a1aad 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index bf37827147..919f2424de 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index 5fbb5c81ee..375ff039b3 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index 5e99c193ee..1d072656dd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index e217cc5837..a4ca387167 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index e662e47869..dfab20c1fd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 3ea7569a39..ba0cf8afeb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index ada5b75ba3..1e34307aeb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index 5cdcc95c39..017a29ae6e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index bf1a26c17b..c1909822da 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index b3184a9cdf..bf85307f02 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index 103f192168..209f04878c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index 02c6581dc2..2915a71a50 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 35068db47b..4837d00497 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 9cc3214c21..c77aaf4649 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index cd573bda93..337d058116 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 187557dafe..2807f3a509 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 0336e432ad..8a31ef2d48 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index ddd6c8c816..4d14203a73 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tma` intrinsics. @@ -531,19 +531,19 @@ fn cp_async_bulk_tensor_1d_g2s_uses_only_canonical_or_compatibility_defpaths() { #[test] fn cp_async_bulk_tensor_1d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" )); assert!(is_raw_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" )); assert_eq!( - generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030"), - Some("v1:i1030") + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034"), + Some("v1:i1034") ); assert!(matches!( classify_raw_intrinsic_path( "cuda_intrinsics", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030".into() + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034".into() ), RawIntrinsicIdentity::Known(_) )); @@ -558,10 +558,10 @@ fn cp_async_bulk_tensor_1d_g2s_cta_uses_only_canonical_or_compatibility_defpaths )); assert_eq!( generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta"), - Some("v1:i1030") + Some("v1:i1034") ); assert!(!is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1030" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1034" )); } #[test] @@ -639,19 +639,19 @@ fn cp_async_bulk_tensor_2d_g2s_uses_only_canonical_or_compatibility_defpaths() { #[test] fn cp_async_bulk_tensor_2d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" )); assert!(is_raw_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" )); assert_eq!( - generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031"), - Some("v1:i1031") + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035"), + Some("v1:i1035") ); assert!(matches!( classify_raw_intrinsic_path( "cuda_intrinsics", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031".into() + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035".into() ), RawIntrinsicIdentity::Known(_) )); @@ -666,10 +666,10 @@ fn cp_async_bulk_tensor_2d_g2s_cta_uses_only_canonical_or_compatibility_defpaths )); assert_eq!( generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta"), - Some("v1:i1031") + Some("v1:i1035") ); assert!(!is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1031" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1035" )); } #[test] @@ -819,19 +819,19 @@ fn cp_async_bulk_tensor_3d_g2s_uses_only_canonical_or_compatibility_defpaths() { #[test] fn cp_async_bulk_tensor_3d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" )); assert!(is_raw_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" )); assert_eq!( - generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032"), - Some("v1:i1032") + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036"), + Some("v1:i1036") ); assert!(matches!( classify_raw_intrinsic_path( "cuda_intrinsics", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032".into() + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036".into() ), RawIntrinsicIdentity::Known(_) )); @@ -846,10 +846,10 @@ fn cp_async_bulk_tensor_3d_g2s_cta_uses_only_canonical_or_compatibility_defpaths )); assert_eq!( generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta"), - Some("v1:i1032") + Some("v1:i1036") ); assert!(!is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1032" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1036" )); } #[test] @@ -927,19 +927,19 @@ fn cp_async_bulk_tensor_4d_g2s_uses_only_canonical_or_compatibility_defpaths() { #[test] fn cp_async_bulk_tensor_4d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" )); assert!(is_raw_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" )); assert_eq!( - generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033"), - Some("v1:i1033") + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037"), + Some("v1:i1037") ); assert!(matches!( classify_raw_intrinsic_path( "cuda_intrinsics", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033".into() + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037".into() ), RawIntrinsicIdentity::Known(_) )); @@ -954,10 +954,10 @@ fn cp_async_bulk_tensor_4d_g2s_cta_uses_only_canonical_or_compatibility_defpaths )); assert_eq!( generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta"), - Some("v1:i1033") + Some("v1:i1037") ); assert!(!is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1033" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1037" )); } #[test] @@ -1035,19 +1035,19 @@ fn cp_async_bulk_tensor_5d_g2s_uses_only_canonical_or_compatibility_defpaths() { #[test] fn cp_async_bulk_tensor_5d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" )); assert!(is_raw_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" )); assert_eq!( - generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034"), - Some("v1:i1034") + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038"), + Some("v1:i1038") ); assert!(matches!( classify_raw_intrinsic_path( "cuda_intrinsics", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034".into() + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038".into() ), RawIntrinsicIdentity::Known(_) )); @@ -1062,10 +1062,10 @@ fn cp_async_bulk_tensor_5d_g2s_cta_uses_only_canonical_or_compatibility_defpaths )); assert_eq!( generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta"), - Some("v1:i1034") + Some("v1:i1038") ); assert!(!is_generated_intrinsic_path( - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1034" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1038" )); } #[test] diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index afbfce5f2c..6a93ea34fa 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 5e3b17d8df..8cf560a47b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index bf74c33b4a..f2a3360730 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index 95eacbc9b2..e4b460ee0c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index c4e9373c17..a3aec23796 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 38708c024c..ac03649215 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. @@ -587,11 +587,11 @@ pub(super) fn try_dispatch( "v1:i0328", )?)) } - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => { Ok(Some(super::super::tma::emit_tma_g2s_cta( ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 1, - "v1:i1030", + "v1:i1034", )?)) } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" @@ -608,11 +608,11 @@ pub(super) fn try_dispatch( "v1:i0329", )?)) } - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => { Ok(Some(super::super::tma::emit_tma_g2s_cta( ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 2, - "v1:i1031", + "v1:i1035", )?)) } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" @@ -641,11 +641,11 @@ pub(super) fn try_dispatch( "v1:i0332", )?)) } - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => { Ok(Some(super::super::tma::emit_tma_g2s_cta( ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 3, - "v1:i1032", + "v1:i1036", )?)) } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" @@ -662,11 +662,11 @@ pub(super) fn try_dispatch( "v1:i0333", )?)) } - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => { Ok(Some(super::super::tma::emit_tma_g2s_cta( ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 4, - "v1:i1033", + "v1:i1037", )?)) } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" @@ -683,11 +683,11 @@ pub(super) fn try_dispatch( "v1:i0334", )?)) } - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => { Ok(Some(super::super::tma::emit_tma_g2s_cta( ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 5, - "v1:i1034", + "v1:i1038", )?)) } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 742fd66b6d..a374fa35ac 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index a0f4db2d57..cb1980288e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 1d5832299a..141afdc483 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index fd4f4683ca..4be241e560 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 9e88dd3302..21f69424ac 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 4eb027ff4d..562ee4de1a 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index 62f363da29..f14c885308 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index ec6caee406..94d3860469 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index b5574be1bc..c4f3c0e770 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index b93d87380c..2436e611b8 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index 76ce652f5f..67cdd9d14b 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 76dd9fcdc1..cb0c7d52e0 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index 2e84715962..3c3991d43d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index a0c461a455..f2c1953ae1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index 92751f538d..d2c6c069e2 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index 7c46d1f3b2..5e1668f170 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index c464f61053..5a6566fee2 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index a8fe431b0f..d0f4b9f4b1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 5155e53e3c..9da63e4d10 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index b038ef4d33..902dc671b5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index c86adea87f..76b6e28ed1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index 3662af1e3c..f8a872d8da 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index f58780d93f..4c78d7e8c8 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index c0ad431bd5..63fb716911 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 7354f3e94a..4b065119a7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index b1a042093a..2c23386dc5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index 67972807d0..030b311c24 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index f0cfa2ce46..55a40fea13 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index a0ffcd8439..78c2f1cf63 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 630a6f2a48..1165428740 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index 2b0cdbe397..eb3aaef250 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index bb8d589b59..a6e609c865 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 00da470c06..97cff65929 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 4386a0a0e7..ef6733f102 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index ee45bab7f6..3b30263f75 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index 879420d1c9..c43f692a8e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index bda450a06a..19e40c0443 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index 71e6b5f00f..01b4649fde 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index 1efa2de713..56c1353537 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index 90063d0f7c..eedb80a2c9 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index caf17366b3..262027dcf1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index 36399b6c9e..d0ccef5ceb 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated collector predicates for intrinsic placeholder functions. @@ -99,21 +99,21 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", - "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929", @@ -2179,21 +2179,21 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" - | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" diff --git a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md index 30f8dc7882..88756583a1 100644 --- a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `27a44de2` (the stamp in every `ops/generated/` file +At catalog SHA-256 `fa1e911b` (the stamp in every `ops/generated/` file header), the dialect holds 581 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `27a44de2`, the counts on this page predate the +stamp no longer starts with `fa1e911b`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `cuda-oxide/crates/dialect-nvvm/src/ops/`. These are the diff --git a/cuda-oxide/intrinsics/abi-v1.toml b/cuda-oxide/intrinsics/abi-v1.toml index 3a06cd74fe..3cfaa4707a 100644 --- a/cuda-oxide/intrinsics/abi-v1.toml +++ b/cuda-oxide/intrinsics/abi-v1.toml @@ -7175,35 +7175,35 @@ operation_key = "matrix.mma.sp.m16n8k64.row.col.f32.e5m2.e5m2.f32.not_applicable raw_rust_signature = { safe = false, arguments = ["[f32; 4]", "[u32; 4]", "[u32; 4]", "u32", "u32"], result = "[f32; 4]" } [[entry]] -abi_id = "i1030" +abi_id = "i1034" status = "active" catalog_id = "cp_async_bulk_tensor_1d_g2s_cta" operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.1d" raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "*mut u64"], result = "()" } [[entry]] -abi_id = "i1031" +abi_id = "i1035" status = "active" catalog_id = "cp_async_bulk_tensor_2d_g2s_cta" operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.2d" raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "*mut u64"], result = "()" } [[entry]] -abi_id = "i1032" +abi_id = "i1036" status = "active" catalog_id = "cp_async_bulk_tensor_3d_g2s_cta" operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.3d" raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "i32", "*mut u64"], result = "()" } [[entry]] -abi_id = "i1033" +abi_id = "i1037" status = "active" catalog_id = "cp_async_bulk_tensor_4d_g2s_cta" operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.4d" raw_rust_signature = { safe = false, arguments = ["*mut u8", "*const u8", "i32", "i32", "i32", "i32", "*mut u64"], result = "()" } [[entry]] -abi_id = "i1034" +abi_id = "i1038" status = "active" catalog_id = "cp_async_bulk_tensor_5d_g2s_cta" operation_key = "memory.copy.async.bulk.tensor.g2s.cta.tile.5d" diff --git a/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index dc3dadac9e..a92ec7a225 100644 --- a/cuda-oxide/intrinsics/catalog.json +++ b/cuda-oxide/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "54d47b8e281dd953dc95fe16f834ff38164d9809171c9a1af097231b0bc45cab", - "overlay_sha256": "1c92502edd54de3fca6e7ab332654262ef9ecd27b60210550f7c856e51e31fa9", - "abi_ledger_sha256": "ce40abc882afa0a313cc657758fd07aef21dd98338d8015f9cccf4edf85eb83f", + "overlay_sha256": "0281537302b94701ca292cc476ad6ed2fbc8e74f4350381382ebb0cfa1256f8c", + "abi_ledger_sha256": "dd4ce9543a61fbb9d2d80b48fb26447716be5bf05440713d6868c082606ee9d1", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -18578,7 +18578,7 @@ } ], "rust": { - "abi_id": "i1030", + "abi_id": "i1034", "module": "tma", "name": "cp_async_bulk_tensor_1d_g2s_cta", "arguments": [ @@ -18591,7 +18591,7 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", "compatibility_paths": [ "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" @@ -19389,7 +19389,7 @@ } ], "rust": { - "abi_id": "i1031", + "abi_id": "i1035", "module": "tma", "name": "cp_async_bulk_tensor_2d_g2s_cta", "arguments": [ @@ -19403,7 +19403,7 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035", "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", "compatibility_paths": [ "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" @@ -20819,7 +20819,7 @@ } ], "rust": { - "abi_id": "i1032", + "abi_id": "i1036", "module": "tma", "name": "cp_async_bulk_tensor_3d_g2s_cta", "arguments": [ @@ -20834,7 +20834,7 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036", "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "compatibility_paths": [ "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" @@ -21648,7 +21648,7 @@ } ], "rust": { - "abi_id": "i1033", + "abi_id": "i1037", "module": "tma", "name": "cp_async_bulk_tensor_4d_g2s_cta", "arguments": [ @@ -21664,7 +21664,7 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037", "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "compatibility_paths": [ "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" @@ -22486,7 +22486,7 @@ } ], "rust": { - "abi_id": "i1034", + "abi_id": "i1038", "module": "tma", "name": "cp_async_bulk_tensor_5d_g2s_cta", "arguments": [ @@ -22503,7 +22503,7 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038", "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "compatibility_paths": [ "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" diff --git a/cuda-oxide/intrinsics/generated-reference.md b/cuda-oxide/intrinsics/generated-reference.md index 6b5f87f27d..7b5033e716 100644 --- a/cuda-oxide/intrinsics/generated-reference.md +++ b/cuda-oxide/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. --> @@ -1393,8 +1393,8 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s` -- `cp_async_bulk_tensor_1d_g2s_cta` (`i1030`) - - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` +- `cp_async_bulk_tensor_1d_g2s_cta` (`i1034`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta` - `cp_async_bulk_tensor_1d_s2g` (`i0335`) @@ -1405,8 +1405,8 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s` -- `cp_async_bulk_tensor_2d_g2s_cta` (`i1031`) - - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` +- `cp_async_bulk_tensor_2d_g2s_cta` (`i1035`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta` - `cp_async_bulk_tensor_2d_g2s_multicast` (`i0330`) @@ -1425,8 +1425,8 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s` -- `cp_async_bulk_tensor_3d_g2s_cta` (`i1032`) - - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` +- `cp_async_bulk_tensor_3d_g2s_cta` (`i1036`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta` - `cp_async_bulk_tensor_3d_s2g` (`i0337`) @@ -1437,8 +1437,8 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s` -- `cp_async_bulk_tensor_4d_g2s_cta` (`i1033`) - - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` +- `cp_async_bulk_tensor_4d_g2s_cta` (`i1037`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta` - `cp_async_bulk_tensor_4d_s2g` (`i0338`) @@ -1449,8 +1449,8 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s` -- `cp_async_bulk_tensor_5d_g2s_cta` (`i1034`) - - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` +- `cp_async_bulk_tensor_5d_g2s_cta` (`i1038`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta` - `cp_async_bulk_tensor_5d_s2g` (`i0339`) diff --git a/cuda-oxide/intrinsics/overlay/tma.toml b/cuda-oxide/intrinsics/overlay/tma.toml index dbb2a826ca..2e0115a0a8 100644 --- a/cuda-oxide/intrinsics/overlay/tma.toml +++ b/cuda-oxide/intrinsics/overlay/tma.toml @@ -17,7 +17,7 @@ abi_id = "i0328" operation = "g2s_tile1d" [[tma.variant]] -abi_id = "i1030" +abi_id = "i1034" operation = "g2s_cta_tile1d" [[tma.variant]] @@ -25,7 +25,7 @@ abi_id = "i0329" operation = "g2s_tile2d" [[tma.variant]] -abi_id = "i1031" +abi_id = "i1035" operation = "g2s_cta_tile2d" [[tma.variant]] @@ -41,7 +41,7 @@ abi_id = "i0332" operation = "g2s_tile3d" [[tma.variant]] -abi_id = "i1032" +abi_id = "i1036" operation = "g2s_cta_tile3d" [[tma.variant]] @@ -49,7 +49,7 @@ abi_id = "i0333" operation = "g2s_tile4d" [[tma.variant]] -abi_id = "i1033" +abi_id = "i1037" operation = "g2s_cta_tile4d" [[tma.variant]] @@ -57,7 +57,7 @@ abi_id = "i0334" operation = "g2s_tile5d" [[tma.variant]] -abi_id = "i1034" +abi_id = "i1038" operation = "g2s_cta_tile5d" [[tma.variant]] diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index 22c75c57d5..48821d7099 100644 --- a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/abs_f16x2.ll b/cuda-oxide/intrinsics/probes/abs_f16x2.ll index f05702a307..30294fdfb0 100644 --- a/cuda-oxide/intrinsics/probes/abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/active_mask.ll b/cuda-oxide/intrinsics/probes/active_mask.ll index 42dd121c29..431b14a720 100644 --- a/cuda-oxide/intrinsics/probes/active_mask.ll +++ b/cuda-oxide/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_bf16x2.ll b/cuda-oxide/intrinsics/probes/add_bf16x2.ll index 6bc104a082..cc98f0c5d5 100644 --- a/cuda-oxide/intrinsics/probes/add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f16x2.ll b/cuda-oxide/intrinsics/probes/add_f16x2.ll index db1855991f..8af71d922c 100644 --- a/cuda-oxide/intrinsics/probes/add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f32x2.ll b/cuda-oxide/intrinsics/probes/add_f32x2.ll index 81850b270c..fb15add5ef 100644 --- a/cuda-oxide/intrinsics/probes/add_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll index 660aaf8dfa..2b511d0cdf 100644 --- a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_f32.ll index 557fea27f6..f87addfd8f 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f64.ll b/cuda-oxide/intrinsics/probes/add_rm_f64.ll index fe9b35b4d3..bef775686c 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll index bfa507a089..b2d1b1753f 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll index 4d8160a13c..b6fb95986d 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll index 89c1657c54..e3dafc59ef 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_f32.ll index dc6097fd89..068dd9d741 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f64.ll b/cuda-oxide/intrinsics/probes/add_rn_f64.ll index ef25a8e97c..64bb2522eb 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll index 15db69bf9e..5f69a54ece 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll index 55083142a1..713afb8b4a 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll index 6bca2816d5..df18ce7343 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_f32.ll index c99c790038..f22e5e3567 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f64.ll b/cuda-oxide/intrinsics/probes/add_rp_f64.ll index 4423f406cf..67c3b3bb7a 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll index 94226648c5..0bbddf35cf 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll index 8cd1f677a2..55b7fb84f6 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll index 142e1ff569..e4402242d0 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_f32.ll index 89efd5c6fd..89ac8d6106 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f64.ll b/cuda-oxide/intrinsics/probes/add_rz_f64.ll index 7d4309fdd6..14e8c63594 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll index 8fa33e1082..094acd711d 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll index 01ba575d9c..ae161e1c4d 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll index a26f5b0998..3ca88b3192 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/all_sync.ll b/cuda-oxide/intrinsics/probes/all_sync.ll index 448caf2240..41d6f5d4de 100644 --- a/cuda-oxide/intrinsics/probes/all_sync.ll +++ b/cuda-oxide/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/any_sync.ll b/cuda-oxide/intrinsics/probes/any_sync.ll index fde7a50a98..5c94810ba0 100644 --- a/cuda-oxide/intrinsics/probes/any_sync.ll +++ b/cuda-oxide/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ballot_sync.ll b/cuda-oxide/intrinsics/probes/ballot_sync.ll index c004732aa7..9894e7532a 100644 --- a/cuda-oxide/intrinsics/probes/ballot_sync.ll +++ b/cuda-oxide/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll index f24bde55cd..1c463ea722 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll index 124194fc83..a7c9af45e7 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index d000befa61..8f1155e079 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 542c7f34b1..e3d13407b3 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll index 06bb231758..50fa5f8956 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll index 2f74988842..c40df70490 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll index 20afac7dc8..d27f4b61a2 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll index 04afa560ab..4ec9d3ee3f 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll index 2cf36f959b..ba3f6c1ec1 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll index 4a2aca105c..617ff7f8ab 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_x.ll b/cuda-oxide/intrinsics/probes/block_dim_x.ll index 5ab6f284ca..8f86fe9fa8 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_y.ll b/cuda-oxide/intrinsics/probes/block_dim_y.ll index 72655eb3bc..bd9cf97d85 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_z.ll b/cuda-oxide/intrinsics/probes/block_dim_z.ll index a070dc03f4..c753626dfe 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_x.ll b/cuda-oxide/intrinsics/probes/block_idx_x.ll index fbffad8ccc..a271ef55e3 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_y.ll b/cuda-oxide/intrinsics/probes/block_idx_y.ll index 4df5a14364..8553f986c5 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_z.ll b/cuda-oxide/intrinsics/probes/block_idx_z.ll index 09d56bca9e..b3d248e9c6 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/breakpoint.ll b/cuda-oxide/intrinsics/probes/breakpoint.ll index 5c5467928f..253b94f66f 100644 --- a/cuda-oxide/intrinsics/probes/breakpoint.ll +++ b/cuda-oxide/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll index bf7c783ce3..f290a0dc96 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll index 91433e0a1e..42fba01d63 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll index eba3308558..771e9898a6 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll index fa7566b9f2..40b97105cc 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll index de912d3cab..f97f9118ca 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll index c7ce3a2e50..f3d49c676f 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll index 1d6fa7f9c4..ce10b4dba6 100644 --- a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll index 19d0d341c7..376cf04bb5 100644 --- a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll index 940294df54..5f35948dcb 100644 --- a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll index 73121c473a..1c9322bad0 100644 --- a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_count.ll b/cuda-oxide/intrinsics/probes/cluster_block_count.ll index d90a8b42ab..a9214d08d4 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_count.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll index 30beb12cfb..8fb14cabad 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll index 0de6c3fd98..744c142bd5 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll index 5c9b53d002..60578663b2 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll index d603980e31..c6f4f64b7f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll index 33c350224f..f64782ef5f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll index 320431d4d3..6ae54a1ecc 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll index cdbd4e1d36..798f456c08 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll index d0d53a9eb2..06cf2c98e1 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll index c2272581d4..e0cc294325 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll index 031bd8bacb..c05ae3b5c2 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll index 9454a1719b..a79a821c0c 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll index 70b2bd6f0d..467bae018d 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll index 61ce34a055..ed2bff3d09 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll index 2606c44604..960c52eb38 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll index a56bbecfc4..fe99c7266d 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll index d2a5850eb8..77d49874bb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index d4caf5921c..1578510abb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index afd44930b4..d14eb4ac61 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index fbe7eba277..abc72715fd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index 52e5d9ed12..ebdd826064 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 78e3c1117e..da8c255602 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index 3304feb633..6bfa78afef 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index 732ea47b3a..b6cbf1811d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index a12c241f4e..354909bb10 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index fa0d528aa2..42de94a262 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index 3a3bc93e7b..8e9f67e038 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index ede7a1bf5c..74b7504f22 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index 7a7237c6cb..a7762bb55d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index a089a31afa..74809390d5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll index 9b75d5a086..3bd29e7065 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 13fe996e4b..b36df85214 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 994e80d1ee..e09ff0c0bb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll index fae02897dc..d7aca9ba6e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index c60553dbb3..0ecf4b975f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index 1b3145914d..4ff973575c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 96823baabe..1dd137d769 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 41d378e0d4..3f9c4c7c8e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll index afa8455a52..63925d7bdd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index 1c5ebdf882..fa394cbd73 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index 7353d629c6..829417210c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll index eff7592546..4e748a7d03 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index 52699a5158..b9a583f397 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index ffae735162..ecfe14793f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll index 438093f702..164b4ecbaa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 3a20538fdc..870e80067e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index 1428b5fb32..99528e1d99 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index 251dc5ef6a..f1cb2d18e4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 312505000c..fe232797d6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 2b7520302a..f597b80e6c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index 3a3e914ce5..53d860b52c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index 6f40fec34b..733f476a80 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index a445dae82f..fb503094ff 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index 257dbe4f56..80a6a53b87 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index cc29493543..5518aabd94 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index 888c4926e6..8d5beceb2b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index 5e01d90ea2..a867735561 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 20eb1bc1ac..73a533acc8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index 6235a46125..620a7c1303 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index 585bcf9d69..975622a40c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index ad5841ee4b..ef06005d75 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index f37883a4ec..b4c8b2c62b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index 2a26025a46..72a59c8ec8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index cfe092807a..ed5cfa25ab 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 5f8d9e35b9..fdcd023bb8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index 58a7dfb768..a6e7d7bef1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 074be49ad0..89e329a6e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index 51f3e5b913..24c9169cb8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index 266a905d8e..ac74010d74 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index 1c27053af6..41f8beecd1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index f5ce586791..313a3a8bcb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index 439ff09386..5de0b66649 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index 555ad5baec..53a57c4b75 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index 3e0b615d9c..509431b06f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index d3f07485db..e9ce45d6d8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index 73522d3826..693b881eb9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index 5d6e87c7a6..b6f25450d7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index c78abd7393..ad47b0b6d0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index 56cca2d412..9365cbf206 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index 0ab6ad5cf6..01ddad00f5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index b522066a2c..3a4958bccc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index fd1a20a8bf..fa47ad3e73 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index 1e7159e43c..97a423a43d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index fe602942f9..9825e5457c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 487072d180..8f4ae70970 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index 8b7ae43ec4..b6aafd6b01 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index be1c7e96a5..3baa817dbb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index 85c1eb72cd..e233df3d35 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index 1232f397a6..f7c7c3ca05 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index f7d49c2382..c7e6172fc4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index 4e8ecbc060..cc37437ef7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index 75f8e732f5..798d3f7d05 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index 22b4993b9a..462a76a125 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index d46eed99cd..d709498109 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index 5710ac2b7f..88054c0309 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index 58778f1522..269b4306ce 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 91bfe4eaa1..68696dfd0d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index 186c812e21..a538fb168f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index 5e4900563b..431a10e0ae 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index fee9d07818..a97d4a1a32 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index c7f33e3fee..0c668062f1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 6072e03954..ec02d89de4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index c49ad65204..76e34839d5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index 078ae8fd13..9989544f25 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 1418decf8e..f0e9d057bd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index ffdfe59af0..2b67f386a6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index 2d1734a339..70a4a3bfd6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index 20571b8eb2..171af39e8e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 32c641fe24..470285fa88 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index 46dd5f540b..eb23160890 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll index 100d1f9e65..053db2afb4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll index ce36c37f48..473cb77a28 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll index e9c68a1daa..86604d2d5b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll index 8048442a7b..daf9970eba 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll index 6f7c12524b..ff865c60b0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll index 25b22f0fef..2864b60a6e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll index e10f101783..84b0f9c839 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll index cc0c6470a2..a2faf55af2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll index 72c3855b99..c7f24c5149 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll index ff8cd71b86..ed11f6c6c6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll index 0e217411fa..7f796702ad 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll index 64773c4221..3079b7f2ff 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index b172541c99..99828303b3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index b3df3c2e79..78926881e7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index 60f4ce600d..49cac04d4a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll index 5b59eb64fc..39cc9ce06b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll index 7ae12d53f4..f2b385d1bc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll index 210505c4f1..0f393b9f9b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll index f5923f5fef..5eba0f7295 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 1f03339fbb..8e69b3a32e 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index a1fcd253ba..4e210a8222 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index f9aa05c9dc..97646f9f23 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 6a68f8bbc8..1231c86ef8 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index d6f6d2444a..a16b400f1b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index dadd77c9a3..b8f9df3b7f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index d09ec7a29f..3235e188b1 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index 0b208cdaea..b25df34071 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index 77279877cf..fe8ffb2e28 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index 201563e358..c68c38e039 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index f541a66b16..eb59c34613 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index 5fdcd37e90..a5da5e916f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index 45882334fe..48968bb1c4 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index cb18c1355f..f563d9676f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index c47d68f479..bae0c283d9 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index f12c4175e4..b1b62bf591 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index 2b88617837..46f6fa840e 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll index 434240fcf9..16f83dc29c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index f93950a912..169fc6c9c2 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll index 59d2e34a80..dc6b1d79e2 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 15c80e7bff..995225bc5b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll index bc70035b0d..0c93a4dc99 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index 37dcac43b4..2a81650203 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 83d01fe815..379f63a77a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 0ba17e1ce7..78e8e1baa5 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll index 08f1d73053..960180e64a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_f32.ll index d0c10b1566..6577d76648 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f64.ll b/cuda-oxide/intrinsics/probes/div_rm_f64.ll index 8e5dd2663c..1d372c550b 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll index b669ec53c6..e53a44bbd6 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_f32.ll index a505743c46..29813b4c44 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f64.ll b/cuda-oxide/intrinsics/probes/div_rn_f64.ll index 4f51c9f531..87a9cdd527 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll index 63f90518e0..1a3436a519 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_f32.ll index 9c885c48a3..2f17617a8b 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f64.ll b/cuda-oxide/intrinsics/probes/div_rp_f64.ll index 5a0ebd2416..a36a9a6985 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll index 0fded723ca..83a279ae46 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_f32.ll index c7a5e7c086..fb39cafcef 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f64.ll b/cuda-oxide/intrinsics/probes/div_rz_f64.ll index 20af939ae4..aa1541f08a 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll index f6565a1147..2959961c46 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_s32.ll b/cuda-oxide/intrinsics/probes/dp2a_s32.ll index 2552e3c735..5412e896f5 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_u32.ll b/cuda-oxide/intrinsics/probes/dp2a_u32.ll index 9072debb72..02b8ae00d7 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_s32.ll b/cuda-oxide/intrinsics/probes/dp4a_s32.ll index bcb82f9d1f..cc19d7a8ae 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_u32.ll b/cuda-oxide/intrinsics/probes/dp4a_u32.ll index 9abe1882cd..6f4de34d11 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll index 9ed3b9eec0..1ab6329d11 100644 --- a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll +++ b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index 75f90a8b0c..266024d8f5 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index 3bdfd77862..dc2e95b519 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll index 8058afafcb..686dac90c1 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll index edf8632c4f..777df16fb9 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll index 9db9e66331..2f70e9c2f7 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll index c85a6c1453..1aba52cc07 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll index ab8fe7aac7..f4a3d83b1b 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll index 9da1e74dab..b423d2515e 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll index 8cf07f8cbe..a602321ad1 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll index d52bc8f975..fe3638a847 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll index c325cceb21..2eafce7098 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index c314f1fde2..ad0b853a14 100644 --- a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index caa9468382..e11b5d0fcb 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index 2668eeabcd..fa42876646 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll index 826bdc8bb7..975610bd49 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 2f4f46971e..3a83635bbf 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index e9790caf9e..af5992c320 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index 2bbf20e3f6..efa02d49fd 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index d09629dccc..07a455d135 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index 7fc8b5e11f..ed7a679d44 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index a986d4cb39..ef3eaf2d35 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index 03e56c88a4..40bc0d3530 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index 2274f93cc2..6f62f1e386 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll index ebb0ecd5f8..b460a6b5ee 100644 --- a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_f16x2.ll index 69e1f76e38..da267a6e0a 100644 --- a/cuda-oxide/intrinsics/probes/fma_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_f32x2.ll index aee367833b..edeec0de64 100644 --- a/cuda-oxide/intrinsics/probes/fma_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll index 3cac204bca..5c1433389f 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll index 526c77c6c9..8e089cee94 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll index f4925604b5..3ce53b5ec2 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll index bf198406c7..6fa27102ae 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll index 7cbe757e91..e6c967d664 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll index a4d097c36c..6040075500 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll index 96e41dab8f..7b01448c04 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll index 9eab5d1401..6795f3dd8f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll index 818b9a706a..a2038de19b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll index f4ee34bbfc..a558f779de 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll index df944e33e0..22dd9f186b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll index b61350ba89..5d854878fa 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll index fb975cd360..2ef6c01dd8 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll index 4d98ebc5e7..3eefc4f129 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll index c449d8edca..c8b7e879a2 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll index 754f683cf2..f7f5c4ec2f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll index 44f57ce5cc..9b6235f7d9 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll index dc084850d6..f0bdafb0ac 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll index df62539eb1..301fa04de9 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll index c35a86676f..5d4b3a31be 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll index d8bfb0d310..2f156d674d 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll index 2d8bec1780..26c564542b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll index f76da2feea..d3ba872079 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll index 2a431bd332..e795ff9275 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 4e603e3fb7..9fda92a736 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll index 49e77a67cc..bff88b77ae 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll index 8b19eb9c43..e0796152c1 100644 --- a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll index 237fc3f613..49e9f7848a 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll index 45fbb4bf5f..d6e2d60ae9 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll index be91431981..77ba8c6a93 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll index 1ced567037..93bde6b5cf 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_x.ll b/cuda-oxide/intrinsics/probes/grid_dim_x.ll index dcce8d5b74..f0f4836e15 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_y.ll b/cuda-oxide/intrinsics/probes/grid_dim_y.ll index a1c25a1d24..55b53e1d21 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_z.ll b/cuda-oxide/intrinsics/probes/grid_dim_z.ll index 02f49fdff5..14b7d870cc 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll index 3e47d5521d..339873531a 100644 --- a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll index bd406bd15d..57142bb648 100644 --- a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lane_id.ll b/cuda-oxide/intrinsics/probes/lane_id.ll index 1d4b745281..9b87194124 100644 --- a/cuda-oxide/intrinsics/probes/lane_id.ll +++ b/cuda-oxide/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_eq.ll b/cuda-oxide/intrinsics/probes/lanemask_eq.ll index 5047ef6ad0..1bd4686664 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_eq.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_ge.ll b/cuda-oxide/intrinsics/probes/lanemask_ge.ll index 04f1297f08..d8c035af03 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_ge.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_gt.ll b/cuda-oxide/intrinsics/probes/lanemask_gt.ll index 455c1aadae..2fdd9c9195 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_gt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_le.ll b/cuda-oxide/intrinsics/probes/lanemask_le.ll index 5ab9b4fdd9..9ef8f45aeb 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_le.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_lt.ll b/cuda-oxide/intrinsics/probes/lanemask_lt.ll index 6d90bbae05..a41b907bae 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_lt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index 5f408b68fd..3a9d07fe3b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 803918f385..5fba531de0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 05f665c244..d73f5b81f2 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index f0975ec155..a0fbadcaee 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index b86fb0b5a1..a83a382938 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index dcad6d4265..20f49a1a29 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index dfb1b3af01..576c4f657d 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index e8d4814b95..fe680162dc 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index 5db68b7995..d9e35d0b2c 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 7d198d1f58..e3a9a61645 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index 10e166855c..800d979a18 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index bb7d6929c0..110e6e771a 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 134f2609dd..8b576a29a5 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 6271b8abc6..ee47d21530 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index 0aadec29e9..6e8f417bc4 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index 28c14e69e8..e55ea438b0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index 239fe1faa0..0cf0506b0b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index 9ead3e37c9..31382416cd 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll index de7861d7ce..ff2f214b9b 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll index c459a1039e..ccc0de1ab9 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/map_shared_rank.ll b/cuda-oxide/intrinsics/probes/map_shared_rank.ll index 6c3be6cf3c..2a522f3b3e 100644 --- a/cuda-oxide/intrinsics/probes/map_shared_rank.ll +++ b/cuda-oxide/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll index e45b6954dd..85e2895860 100644 --- a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_sync.ll b/cuda-oxide/intrinsics/probes/match_all_sync.ll index 64d6424904..004fa2bceb 100644 --- a/cuda-oxide/intrinsics/probes/match_all_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll index 4dd9cbb891..e787ac97bf 100644 --- a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_sync.ll b/cuda-oxide/intrinsics/probes/match_any_sync.ll index 1f34f2e120..aa5fa01fcf 100644 --- a/cuda-oxide/intrinsics/probes/match_any_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16.ll b/cuda-oxide/intrinsics/probes/max_bf16.ll index db05412337..3915bf4c5e 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_bf16x2.ll index bf1c1867ce..5f0c08acbc 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16.ll b/cuda-oxide/intrinsics/probes/max_f16.ll index e082a8b657..19b902801c 100644 --- a/cuda-oxide/intrinsics/probes/max_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16x2.ll b/cuda-oxide/intrinsics/probes/max_f16x2.ll index 5f1e374ab6..34c2d3b799 100644 --- a/cuda-oxide/intrinsics/probes/max_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll index 5d6e777d78..8707b2a028 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll index 1b09f6813e..cba898aca5 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll index ef07624785..55aa161c6f 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll index e38ca068f9..2dbceeab30 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index fdcd13f9a8..f8c2129ff5 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index 58edbdd1b2..9eb8cf9b1b 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 8e15a2f8ce..767f4857e4 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 5fd11083eb..a3b9678f5c 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 9c54f200c9..390cd753fd 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index 7109c240a9..759bb8b173 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll index 43cfbd8382..3c66555707 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll index a692839b5c..0083aef67d 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_f16.ll index 6da1f602fb..5e616609d7 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll index 70cd3ded78..7b6fd62ceb 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index 9cf88a5331..b8ee7e86f8 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index e563f2d82d..a84daf63a3 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 65da1f724c..24b56227c4 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index 5c2061cb9f..a3de9052fa 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll index f6f393403e..07327c9aeb 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll index 33bb7babea..3d1f233927 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s32.ll b/cuda-oxide/intrinsics/probes/max_relu_s32.ll index 1c208a2c5f..11a569da31 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_s16x2.ll b/cuda-oxide/intrinsics/probes/max_s16x2.ll index f210671cf8..18b31fde20 100644 --- a/cuda-oxide/intrinsics/probes/max_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_u16x2.ll b/cuda-oxide/intrinsics/probes/max_u16x2.ll index 6807d26946..4d6774202b 100644 --- a/cuda-oxide/intrinsics/probes/max_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll index 7110a3443f..759ab13cc8 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll index 2af08f61ba..d4b370e723 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll index be1be97ccf..965aedef26 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll index 41b35d900d..2dfaa48f1f 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll index 756af2f216..d6794bf001 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll index 2305c066ff..43ca8491c1 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll index 21c70f5f46..aec9011ac4 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll index e43a23c267..5e082f8b86 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index 4fe1b80dd0..f2bc4b124d 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll index 6704a787f1..a9005d79df 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_init.ll b/cuda-oxide/intrinsics/probes/mbarrier_init.ll index b7483c4e59..470335c69a 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_init.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll index ce543bcaad..d359eea934 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll index f86daed164..85ac820225 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll index 72631f9c34..596e2e53d4 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll index 7bcdfcdeab..b9e946fa74 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 4c7304058e..d17db44a9e 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16.ll b/cuda-oxide/intrinsics/probes/min_bf16.ll index cbb4690428..73bc6b81e1 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_bf16x2.ll index 5c3658c6b6..b1135f9996 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16.ll b/cuda-oxide/intrinsics/probes/min_f16.ll index 0ab84d0b4e..fdffab9651 100644 --- a/cuda-oxide/intrinsics/probes/min_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16x2.ll b/cuda-oxide/intrinsics/probes/min_f16x2.ll index 0f546aa5c4..8c54c83a7d 100644 --- a/cuda-oxide/intrinsics/probes/min_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll index dfd984cae3..ea1e411fe7 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll index 9b93732ff1..c9d7fe43f8 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll index d967d8adbf..3d2a36ca24 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll index b9aa82f5e6..c8473ae935 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index 4a0f1fd17e..9e42160b0d 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index c197a01a64..cb46b6e3b1 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 3968b53633..9c0541e20f 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index 04fec34cf2..fc857cf1fc 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index db0f19956f..4930db1765 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index 3c091f240e..2c77d330d3 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll index 1a657290a9..fb0678ca97 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll index 9cd78c860a..094b6e2978 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_f16.ll index 24c77655a8..7bdac65303 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll index d77ddc12a1..c33ad303f1 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 50672b4b2f..099e5b4681 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index d94710799b..da3b30ab63 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 207e88d559..598aec9d31 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index ef17e34a36..c5f30485b4 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll index e9227ac768..99bbabf91d 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll index 6948ccec01..432d5b6a33 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s32.ll b/cuda-oxide/intrinsics/probes/min_relu_s32.ll index 1219f67785..b0b447f143 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_s16x2.ll b/cuda-oxide/intrinsics/probes/min_s16x2.ll index d0dbe7c39a..7a39dc092a 100644 --- a/cuda-oxide/intrinsics/probes/min_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_u16x2.ll b/cuda-oxide/intrinsics/probes/min_u16x2.ll index 49c37c57c9..d5f6fd3049 100644 --- a/cuda-oxide/intrinsics/probes/min_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll index 0077b89619..fa82bc5ec9 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll index fd80aa42a3..7049780f10 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll index 5b88aae4e0..670354cd31 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll index 46dc722397..8a10e2a118 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll index 0bcbbc2c15..2e4677b54f 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index 6f5ee9ae51..7fced439a3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index fc6f947505..8e7bbb9e16 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll index dda5519043..ea1a2aae65 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index 10658b63eb..02cbc8ab4e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 19dd5ae3ec..7be1a23d5a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index 5d358fada9..b4129bf4c2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 95c4e5e682..d1382e0a09 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 261978dc8e..8d511b7cbf 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index 2a6c11c1ca..6e209237a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index 9d1bb41ffb..9261af70c6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index eb4d3913b2..f66c517553 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 8dbf49f5b6..175c2c38eb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index c848e0c9df..3d8e2ac307 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 2eee1dd31a..e26c898a58 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 4649851dba..1929620c96 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 55c786e4d7..86eef77313 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index f271ff4cf5..f6986ef8d9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll index c3ad6feffd..b083ebe2f3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index 50fcbc9846..9bd7d04643 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index 23788b4c5c..181868aafc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index 2071535076..0fc039375a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index 964389d77c..4d10d2a61b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 2b8a4f7e01..fde94adda9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index f569a2fda6..3c2d556ebb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index cf4d3fbdd9..35f822ddba 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 8e03cd6e53..181aa20f03 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index d7bda206b6..f29c0093aa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 6cd7e352ef..fd8a533ecd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index 0d3b3ecd55..b30fdc56dc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index 1d39a71df2..a3c0143e11 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index 881d5fc1b8..70b309832b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index 1ea82e15d8..bd3bfd636a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 2b8c996052..1619042062 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index 8cb87a89b8..17a95720a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 9d0f00e9e2..99672d9ce7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index aa95263375..bb4e77bb5e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index 199e2aaf64..018cc6bc1a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index 93e717b25d..44ec697671 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 4f8cfdd3f4..d5a354976a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index b39bbd31f0..2e033ed2e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index 7b3133fcdb..9e4ffbb65b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 8de2266327..b016a18da6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index 29deee89de..610d938181 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index bd6f5d1fc1..50b349fc52 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index 14a14a2993..170210adbb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index 660e4f85da..237b1555ec 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index 46e2b7ccea..e0e5cc29f0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index 89194d3674..f098f1f448 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index c2e764e77b..1760670572 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index 8f7cdd2974..5633664a4a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index fe321897b7..ee78875414 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index 6ccac258cd..0c5dc9ae1a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index e7d6cdcd3c..3a7fb1dcd0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index 6d1fada2c9..82b732ead3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index 481da75e52..254b30b497 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index 4d42704022..5afdc0bc77 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index 025c429574..7355d26b8b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index 1522651fb2..b8ee605a54 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index c400dea514..e1793bb822 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index dba9b759e1..68ed65dbec 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index be0019ac64..624c2870ea 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index c0e4301aa1..07e4a9837f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index 243617e60c..23f53c3ed9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index 44b454f6ed..08eb63f4ff 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 67c43b1f2b..3a093fccb2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index c68bc1f5ff..bae03339bd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index a6f3e439b6..d3ea2d0e67 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index 2e7f838334..e1e89d2099 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index 45a9d81eb1..59604eaafc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index 0155176864..f1a2398c00 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index e219694466..5e9dd64c36 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index 411c8a8bcc..26eb73ce59 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index 4a22d1584d..1fc8a39e4a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index faec6219a1..676de69b4a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index ed89625875..7bb9204c88 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 261eb4a4eb..23a1294d19 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 59d833b503..94bab0edee 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index fa2c2c433b..f95acde0aa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index ef9c385462..e806a0c5e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index b7ae243457..c1794a1523 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index c8e31165b3..bcdc93402e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index 7356926c5f..2ce88b80d6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index e706da145e..92ecef714e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 2484931724..826fed3528 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index 0aa111bb26..45c103971b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index 2b07d65e1f..1d76f705fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index d95ef9dc93..f46a73643f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 22d1d2918b..cd8b44d639 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 2067f3c98e..2b405ee513 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index 35cc3aa8ed..cd5d3aef50 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index bc7bb21ec6..af9100ac35 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index 8443b6dd66..f0529d67f5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index c73edae0bc..6fe3c356e1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index 7db34fcea0..f60a7dcc81 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index 7d62c8bc52..83760cb36d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 7942ba545d..7b475b1d50 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index 1cea7459a3..b0f42a9914 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index d59cc377d7..9dc0f914e4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index ca23ddcde9..036ba68dd4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 5b64dbb711..fe7e7ad762 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index 7b7cce682d..35b8440b43 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index 574fb486dc..be8d33f63f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 895269fc56..b93c771d1a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index feb9590963..fb5f0cd66a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index dd4ca4a42e..1e3e4ca26f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index 152f4f02c0..a0d00cd238 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll index 274cb21c55..ae85ae5358 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index 1021d5bcbb..2542cee152 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index bbb55c1f02..7bd2508ea0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 9806945eff..77701339af 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll index d7b2cd90b7..6ebea0d7f8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index 6492380e40..8f96efe1c4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index 68d814d7c3..cec7b3c639 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index df1c9cddfd..e6041f5e5b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll index fed6ef6127..2d817647c1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index 6c6b47052f..03317e81f0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 091017fc77..31c1206b49 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index 21e9e7a5a0..59ef5ceee5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll index 264b567276..4106b9dd51 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index 9806ca7a66..18add54959 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index b102d27bdf..4b2b7a8693 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index e35b05fe13..d3da5dbd34 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index e06cbeaa83..696018c64d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll index 73fb61c7b6..386770bb86 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index 9a357be363..897ab8c80c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index d9b582f55f..cd6f7fa600 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index 1be1a63244..c8799b0b26 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll index cadd11116c..276de13741 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index d13841e3bf..5a0dac81e5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index 2f58e6ff99..d449c99ed0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index e7717999e4..db5555348b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll index 353abfb5b5..b244edcf96 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index 076e37fd64..2b7fce4acb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll index c7fac51337..2ed07bfa40 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index 12ecca3203..642af63935 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index c1bc8df8d2..eab4e1df10 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 5304775a25..5b3e4809d5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll index 54527b33c0..4e3907501f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index d349bc9d25..cbadc60c86 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index b865520e57..ed328c6bc6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 4d0b6c40b9..826f6247d8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll index 79db5d385d..1b406929dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index fc3eb2907f..777c59c03f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index bbc1e54413..baa01591d5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index 10038b2094..63dab80344 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll index b0e0ddd869..841a895993 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index 4c611e8f8f..53725bbe49 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index e347edd3c2..4262d7cc41 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 494c3640d6..3b80348e98 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll index ec13655a77..5bbb80d3e4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 102ee68860..944fd6925b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index 91cc530325..7bb1d14515 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index e5a0d710a2..435ec0475a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll index 9a0972aa4c..f48a02b01d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index b8af8a4f4d..d3a37e2d83 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index aedff8fe82..3dd916f1fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index 3b90323bc3..4ada6cb3bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index a81e346a1b..ab9419b5cb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 8e5d43894b..bd2e56bb00 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index 7e47db81c5..279b6af18c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index a043e62730..0c6c46df73 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index 234953e449..b65e696209 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index ba660953d2..fbbb09be27 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 7db0f082f9..5a77bd6c9c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index 4f14f5dbce..5d76390439 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index bfada891bf..826002701d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index cf8400d86e..7ec5abdcdc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index 0f9baa8a5e..729e43225d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index baf7438f5b..9fa96f842d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index 23939b8515..6034307b6a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 919c9152fa..4b0e4d3bbd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index bb5fb3798f..66c91a8ed1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index 9422c377d1..be1a3c7a5c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 3ed16d0737..516f59dd9a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 1e7b5cf0de..1f4352d5f0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index fc068dd512..81b448f9d0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index c77801ff82..e874a6a77e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index dfa24c9725..8f8139a8fc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index 59cb522573..b01e64c66d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index 731e5ac301..8098042fff 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index d4e98bfc41..4a1deaef32 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index 31aa2569a7..4056848f62 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index 15be9a1895..93a98c8bb8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index 634b4de537..ecddd681bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index b2298b4786..03e590cf3d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index 8840e0d960..a039640e76 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index efb7e85706..7a81b89eb3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 557edc105f..b646961498 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 74dbcf3733..2707943de6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 95a6910cfc..f5c5495fd5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index 479a2b5845..8a22419282 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index 1f52bdc1cc..7bbef8e2f9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index 99aaaee2de..3e4888b987 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index ada0ffc173..de0a73381a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index fade8cf58a..479b612aa5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index e39f6abfe5..ee155a2835 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index 996dbe6940..ac7c11afe0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index 565aee28e9..785da4962b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index 9f040d901b..5ab68e60f9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index dc89194270..f80581e88e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index 904deac0c4..533a20a12b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index f954ff465c..e3d294809b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index 37db99db47..c6dd533c90 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index 861963f194..0d059cc8b0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index ff876190ee..63392cfd56 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index b75b90b980..5441e3301b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index 4d108a2f71..15c498f00c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index 59ffa3d3e4..37f84392ae 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index bf5969ea56..34d7fdf4bd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index 7e5adfe19e..98cdaa5788 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index d78adeb8ac..a9ce87169f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index 865ed77c10..11b8a80354 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 3a143163e1..14e1538aca 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 1e95077965..8b5e58c1d7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index 51302dfa38..e5e3f006b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 43252d5679..95972da0c1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index c51b7672f3..87be6dbeb8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 0c8d5fee4e..e952e2202f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 40e222a3ad..b69b0e17a4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 726d51d545..45f71f114b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 961c8067ba..79133af8dc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index 3a01802e62..8c970ae19c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index cd9fe6286e..56681e7372 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index 47cc75bf1e..9cdd727326 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index 704620d22c..1d6985ee0d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 1f777d0f24..855d8494fb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 175aac78ca..32a8bbf7bb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index d2e902677d..6e1a6daa04 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index 533854175a..3eb87b3c8f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index acaa0fb446..a8a12cb9bd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index 6990466bcb..26662aaab6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 8662c2d00e..a146a7740b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index f9ff8eb12c..d31e91be9d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index 25851abca0..976e930259 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index 8bf84ebbd2..6ef1b978fb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index f8c13e1af2..c97e3a1f6b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index 977056eb6a..1b40fafaed 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index 03a20a933b..f310c09bd5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index 838cc7cf74..b10ae69bb6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index ca9df15bac..ff25572567 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index 0a2ebb5c6d..de3b84e011 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 85f9b9678e..480d7a9707 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 3aa5e8d7ea..1eaf18bc80 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 17d3acf829..d8a630e6c7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index da6687a9d4..b02716dcfe 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 08c9116ab8..e6c9e38b97 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 6cc54042cc..62273cd6f0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index 6a01c4971f..6ce2a30e32 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index 03b28035f9..86bc5028b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index de7e2b3232..53351a5fe0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index 7bc8dbca4d..ff1f10fb2d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index 3489016362..ed4c5e28d0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 8703acd89b..970f921b73 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index 94f5c8cd77..0155cb1b82 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index 3c344b62ad..924049efaf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index a19a847946..fb2f58208a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index d5aaaa8491..c7f4ae8b5a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index 3f546f610b..099d5020b5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index e6d644d332..42153ae6fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index 6ba9934e4f..9a3655942a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index fb92595a29..d154adaa0a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index 5bec047651..1e1e3d1338 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 5f1d6176ac..1e24adb1be 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index 96a0c57fe2..86e1306abe 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index aa72f4833e..c9fa98194b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index e0f69dfa9e..1233589aeb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 8ea2df6939..c1e38dd0e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 0686cb1967..610fbfc8d3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index e23e51179a..393e80e333 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 0c33ea62a9..ef70834c8d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 70668b88aa..2e08fbe67c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index 261c676925..9b186b8733 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index 83b0603519..d068ae07e7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index 8c72eee160..9aa315e7dc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index 84ac5400be..1ac0a2c95a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index efd541f6dd..bb812e6382 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index 25741d3ae2..2f0b34d0ec 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index 887af17cc2..8f3729b4d6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index 552d047b6c..e4c5d17d7b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index 628572fbbe..75027e3b06 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll index 93c710e4b7..6b2b8c9764 100644 --- a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll index ccb6479f46..9f5adab448 100644 --- a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f16x2.ll b/cuda-oxide/intrinsics/probes/mul_f16x2.ll index 887e069f78..e2f634a6da 100644 --- a/cuda-oxide/intrinsics/probes/mul_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_f32x2.ll index d8bd454149..93f7653620 100644 --- a/cuda-oxide/intrinsics/probes/mul_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll index 96c9eabc25..27bf56da39 100644 --- a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll index 30816998c2..767c9588a1 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll index 5d25b27bd2..a1a40ca223 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll index 654ef5c4d0..6f2e4bc649 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll index 8ddb2deada..081a7acdbe 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll index ded2a2550a..289f5dac59 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll index 851a9b7723..ca930103e8 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll index c6496c2dfc..2d07a17112 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll index 7f2e7256f2..164b54a6a4 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll index d7f0800313..20c33af950 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll index 8e8f68927f..1876974f6f 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll index a2342939eb..5a61dbd8d3 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll index 1d6b2645d8..845cfb8d46 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nanosleep.ll b/cuda-oxide/intrinsics/probes/nanosleep.ll index cfbef4a781..5c5742795d 100644 --- a/cuda-oxide/intrinsics/probes/nanosleep.ll +++ b/cuda-oxide/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll index 0ffa3e0736..1bfb55c5f9 100644 --- a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_f16x2.ll b/cuda-oxide/intrinsics/probes/neg_f16x2.ll index cadb800301..f5e802f9fc 100644 --- a/cuda-oxide/intrinsics/probes/neg_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll index 72afddc32c..495ea86902 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll index bb4cee13f4..052d83cb08 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll index f707f5154b..d228b3b822 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll index c294e0a568..8b238858b2 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll index d32ec4c9df..bec7b08860 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll index b100e2bbd6..927183ce75 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/pmevent.ll b/cuda-oxide/intrinsics/probes/pmevent.ll index 2688caca3a..537547869e 100644 --- a/cuda-oxide/intrinsics/probes/pmevent.ll +++ b/cuda-oxide/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll index f8fe3e88c2..340f1ffb13 100644 --- a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt.ll b/cuda-oxide/intrinsics/probes/prmt.ll index 6089a48da6..338bc0375b 100644 --- a/cuda-oxide/intrinsics/probes/prmt.ll +++ b/cuda-oxide/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_b4e.ll b/cuda-oxide/intrinsics/probes/prmt_b4e.ll index e3f84c4991..8264cb1049 100644 --- a/cuda-oxide/intrinsics/probes/prmt_b4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecl.ll b/cuda-oxide/intrinsics/probes/prmt_ecl.ll index 6f81144609..bb65dd77a6 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecl.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecr.ll b/cuda-oxide/intrinsics/probes/prmt_ecr.ll index 27b3245aee..62f2119d5d 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecr.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_f4e.ll b/cuda-oxide/intrinsics/probes/prmt_f4e.ll index 91d513b69d..6f7be9b38b 100644 --- a/cuda-oxide/intrinsics/probes/prmt_f4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc16.ll b/cuda-oxide/intrinsics/probes/prmt_rc16.ll index b8953a6771..6e54594ca5 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc16.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc8.ll b/cuda-oxide/intrinsics/probes/prmt_rc8.ll index 023cb5aa25..cfd03586f5 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc8.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll index 07ff28f22a..e6374b6400 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll index b21318cddc..5aca8b1bd3 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll index 839b1d26ef..4f50162c99 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll index 85be84f8b1..901690b487 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll index c5df1f1172..eff7887b5f 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll index 7db89bd26d..8dba5abc6b 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll index 285c55022c..76a224d96d 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll index 9c43f1c8f4..3526c85336 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll index 15e5bfcc34..7b16be3727 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll index f6a53d62ba..c1b8f8ba09 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll index c4ad38e6c0..07939287f4 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll index 923fd5bef0..02362a83f1 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll index faa83b9b4c..e0d65a91ea 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll index 6b6ab37629..e761f2665f 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_add.ll b/cuda-oxide/intrinsics/probes/redux_sync_add.ll index e6a43e9c8a..34e84b5ea7 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_add.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_and.ll b/cuda-oxide/intrinsics/probes/redux_sync_and.ll index 0a1756b9e7..3b34f20d00 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_and.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll index 51c36d8914..feb70e5bf6 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 518743e908..758740dd27 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll index d9e9c8d9ee..2d065b6a3e 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll index e34284de22..46b821ecb4 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll index 13004abce0..59c5a88f66 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll index 3d1f655cbf..70b93e0a36 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll index 34afa8d841..9e60798895 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 4435152d52..96a9cc4adb 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll index 251b1e8fdb..dd96152da5 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll index b98fc2784e..dafd07069d 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll index 8dca4e944a..53940f52e3 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll index 1414367cf3..92aa91f845 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_or.ll b/cuda-oxide/intrinsics/probes/redux_sync_or.ll index f41f3f9209..998209071f 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_or.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll index f3d621bcf9..cfc83dd98c 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll index 8e14421257..bc2f96de16 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll index c497f14225..f0e37cad3c 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll index bcf761b484..39051d2333 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll index c55691a78b..f7cd027867 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll index 474925191e..7921cfc111 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll index 81ebd917e3..a3f455d794 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll index 04c0c5ca57..84b32f1b2f 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll index a4ecb95a64..9ffa61a7d6 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll index 0263f8f204..5de3003225 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_sync.ll index 8feab8c6ba..c1c7e4962d 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll index 46ba3a1f2b..d9837487b4 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll index 06927e75cb..7f1f40c357 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll index fe5ed42e22..55378b037e 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll index abfe57c1df..02337e412f 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll index e845d30470..15f5d098d3 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll index 9a89c9db6b..d8f15708cb 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll index 3429d65598..832d1f87ae 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll index 9ef2269cdf..71ee800291 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll index 9de516e3f1..310077aae4 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll index 1aeb3808a2..198e744f4f 100644 --- a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll index 42a3ae72e3..48f6284a8b 100644 --- a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll index 3dee05eb88..47a9a47e1c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll index 4efe659c3d..4fea2a3faa 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll index 693a8621dd..0b56788d6c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll index 035c300b0a..cedadf6913 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll index 73c7495ee5..a881021afb 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll index f5e0fcafea..5417992ec3 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll index 61ebdf30c4..297e517c86 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll index e9ee3e5195..8cc3aff55c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll index 453ab3634c..79a8fa3fbd 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll index a5c09b6ef6..849b49a4d2 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll index 960fd3dfed..7c896e1b19 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll index ef80ae6040..79519b960b 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll index 973f251ded..32109e9fe8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll index 0548c4b775..4fec7078e8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index d9b0a35832..406373c1fc 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index 995874eb30..f8ec76d7f4 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 83deea6de7..df88383709 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 2a43e14ca8..b0e0476360 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll index 0d3df158de..5903229905 100644 --- a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f16x2.ll b/cuda-oxide/intrinsics/probes/sub_f16x2.ll index abc5a2bbb6..b9ae761fe2 100644 --- a/cuda-oxide/intrinsics/probes/sub_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_f32x2.ll index 48dedd923d..692af94584 100644 --- a/cuda-oxide/intrinsics/probes/sub_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll index fba9396876..76acc7d49c 100644 --- a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_mask.ll b/cuda-oxide/intrinsics/probes/sync_mask.ll index 7d39ac816d..5bc286dac6 100644 --- a/cuda-oxide/intrinsics/probes/sync_mask.ll +++ b/cuda-oxide/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_threads.ll b/cuda-oxide/intrinsics/probes/sync_threads.ll index 782acc0e7c..7b505f8d1b 100644 --- a/cuda-oxide/intrinsics/probes/sync_threads.ll +++ b/cuda-oxide/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll index 2bc026ef17..162cbeae61 100644 --- a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll index 9e426ace52..d6943698e3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll index ecfb457c13..2837dcbb03 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll index d757224555..beacca1007 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll index 31908a726a..58fb41f6f3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll index c725fb7e5b..f36df523f1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index dd393d6993..72f1383e37 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll index 5603bb550e..73fba13603 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index 55e47e9c86..7af2c1d3f2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll index fce3746ee6..e85cb347ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index 21d2b928ba..f149c33ca8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index cf5480440d..e4fe3f095e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index 6ad52ec2a3..57dbde9665 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 33d066f9e5..84c7b15c1d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 69a8be76f3..6d2bd5561d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index 79b136506f..b39499e088 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index 402206f242..7228f47dad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index c8abd2de4c..ab7727afd7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 0af9e40376..238e817552 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index bbc0451a8f..95c4475585 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index 3c4d174ed4..d767c12a37 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index 489d1e088b..6fef130535 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index e1fede9e1d..ffc41d76e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index d477f22827..0afdfa9ca9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index 8c9209308d..ed408670cf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll index e650245bf0..c613f766b8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index 7bc1b9cad0..5a98865c2e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index 8037a5e6b7..d3e253d043 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index d7735bf157..6fd99291d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index 1c9bee8b8d..4004f0418e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index be9cbf8077..b5b79ddd1c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 7cc251b960..11a5371925 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 61ab14a67a..d29d7dd933 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index a1559f4107..fa31bf9270 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index 11400ef08b..6584de44d3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index a5f40b211c..4f9add26c5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index d89e33b39b..0dbbb6ab60 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index 5f386d01ab..3722c651d1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index 804faa5dea..76fa7b7bed 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index 2756e25059..26f9570fbe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index d97f8e4c75..59173a1603 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index 4cabb5222c..e2dfb22882 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index f900f76cea..e135b319e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index b0ce19f2a9..67ee306a58 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index 6a8f751c55..9e53814bfa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll index f85928e972..03c0ae1698 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll index 24ba197fdc..ca58804d61 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 4df227d8cc..8bd3647e42 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index e4f50ce3df..e0edb3136a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index fd5abaa846..a160c6b45b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index 1e85632e7d..891d849b57 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index e6b75d13c5..fbe7b206a3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index 737ccc389d..124a5cd7d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index b8a5638417..9ede8b9b04 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index 764e135816..983be806f5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index ac267979cf..1121c9fab0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index 51dd5eaf39..200e95acaf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index 31f425756e..17ebdf94e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index b10a37a164..118a3629d4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index 4148f67e4a..7db7ebb301 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index 7e06f46c65..9384c5dcc2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index e4633628d1..25a8200b94 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index 0e1eb56461..b72a062822 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index aec768bc94..f9636c21eb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index 88acff2bc9..dd883fb6f5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index edbb0849cb..9885c43019 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index 6a0af12de8..cb78a6f05d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index 0f2629f72a..d247a37d6f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index 2e865247f4..35dd18b483 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index b50ed4ddca..c777f79190 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index 07abd729d3..12f81127d3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 01dd2998d8..68ccbdeaa7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index ba20cf8a28..844f205e44 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 95bfd49fb2..b90bf2f35e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 14f6b3a1ce..bf56b8db17 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index 6c440250a7..495e010b37 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index c5cb8baf7a..4179e9461d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index b1df59ac7f..f7e28a9867 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index 5bffe4aebe..6aea175b72 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index 5522b69b5c..a973d3e18c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index 9c33a63b82..456db010e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index c1588cb3f6..8f4b19f8b6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 9cb4b58d5c..99923b6154 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index d6b411d814..ca6561cddf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 2eb66024af..d4ddd0cf2f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index 38ed797246..973a45004c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 13c7606ccb..6057102ab3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index f4ce146c59..b5ebefe766 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 9f2519ff32..80f8f5ecfc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index 666ca70830..a6170da5a1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index b079fa0c34..32711ea927 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index 599f83de66..cce87662b3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index 2fb81fb674..1934b30cbf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 3139f585e0..259e5d6797 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index e04d72393e..f92bb6e698 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index 8469b8ab44..1960cdddd1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 82669815a7..6db8d17fee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index 84dd7fbdaf..bee818a3f0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index a5b9920669..dcd0c87501 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index a6fc0a6a95..d70299024a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index e2785354f8..08139272ab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 9eba9bb84b..45aa888971 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index ebc358fe30..76bd18d709 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index 53e6e8f82e..45b5777f96 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index c428cf74db..a60d01ecea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 9582d2da89..d9f9b18628 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index 93e1f17a8b..5c99d06ab8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index 7d8e6b45bb..dc69d64b17 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 50e554379f..dd3cda3be6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index 249c70b23d..62b67ccbbb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index fd719613f3..d082e1389d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 61e0134fe0..f0d988e403 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 4d3b637caa..1f8d233dec 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index 060aa01112..3e0b5468eb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index 8491197808..db87e0555a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index fed611b36e..6b40750836 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index c2e1ef6832..764b1000d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index 9ee5e83e39..95c49c3c52 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index 9529df77c6..5bd294fa02 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index b46f73307c..34920f11f8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index ebbed8efb3..446714b6bc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index 381c7762da..94777e6486 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index 0f3416adc3..5bd72fb077 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 2cb5c40793..36c6d8d6e0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index f31f2f0632..987ff59e58 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll index d01db69f1f..adeb70f7fe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll index 2502e6114e..6c26a4c2a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll index 3653aba278..ba4a16889b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll index 0eaf5cefa8..cfe42d5387 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll index 4c79b78dab..d6ebf29cb7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll index 8b55a03290..f7de1c7c48 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll index 48a0025ed3..738bf9409a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll index d71d368b81..db8b78264a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll index c674ca9109..e75ad4cea6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll index 8d6523dd25..5e0dc4ea2e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll index 71e2929e83..674a3c3e2e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index 26065d41c7..c15358ed65 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll index 66ba1ea8d7..c82984c821 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index de696158d7..64350f9196 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll index e227cca3bd..39161987de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index 6e75c79401..99cd6d9aa2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index 3280019bf6..abb335fc7a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index efda75b2a9..cbf21fbe93 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index a07c0f525a..ec6ad26254 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 251eaa1315..f639bac706 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll index 821bd62781..0e3b89fc5b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll index 0d851cb5f9..440799189d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 9ab2388ad1..a3def66e21 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index dac0aa39aa..d4be006a79 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index bc486b1c3f..f6d81b1310 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index 083f52c4d4..940ddc7270 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index b8deac351d..d657c36950 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 607db6e054..4c1c905000 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index 6c492622f7..d664279e7b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll index b7e67809ef..c4dc80f1e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index c02c3c8abe..a2b1951111 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index 062364fb49..c0dfb331a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll index b6e5f90be5..4ac1d62ab5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll index ef1f15226a..77f58bb984 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index 0388b6262d..1670df36e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index bef6918b9d..a7ea337f4e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 645e9d5bf4..3c08768bd1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index a115cdeab0..2eb2114f3d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 09527c4636..fc272716cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 08b9728159..0c5e14cadd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index 4566d85d7a..2cc73d5a01 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index bcea3b049a..130b55314d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index 3a4407f174..2f86d6c38d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index 9f922b78f2..33c8075adb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index e0457e8ad5..db127c4f42 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index 011c4186b5..15a816298c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index fcfdd5386b..ddfa93d994 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 6cd6472ad0..bc19b0bd8e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index 102ee1c3f4..b06eafbbeb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 66afeb2014..f9a2bd896d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index 5fef1eeee0..a514c46c9f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 4b5f30eeb5..600c6caba2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index f188ef695d..dc2d298f50 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index cc48c50c1e..a499f9a95e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index e6ba731c0c..61648ea7b5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index e1e653ec61..eb3712638c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index d87779c277..059c7df624 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 48e4683eaa..fb3a65cfec 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index 3a7b3b4c93..fd4483b727 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index 0fd1258151..0a2ec1cc23 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index 3e90947285..314b253209 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index f14b322892..d8dc9f7226 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index 6e8236d176..c788c2ed5f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index b125d4fb6f..78917ba10c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index 6455b2b272..05a62819f2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index fb4d5268c8..c033fff8e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index cbd301cd65..57d9202b02 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index e8e5843ec0..30d004a5c0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index f7745999ce..d9dc47604c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index cdb2de9c6d..4a6d6936b5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 8943d86cb7..12f3bca5b9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index d14208da2f..91a5c22e13 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index 6ecadfef8f..bd53c0d797 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index abd195d2d0..56e1fa810a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index def46cdb6b..f6f1f43644 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index 9303d27bfa..732b503ed8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index a40ace1810..c0503471ca 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index 692958f0d4..b145dfe4d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 9123e8d811..136bd82d56 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 58775f57ab..9a02a9c510 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index a8d59e3580..dfee46f9d7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index 72f728955b..e76c2bec33 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index 8af53799b5..30f34ed4cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 6265704980..927ea6619f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index bd25989d24..0c1de5a09e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 90527b7716..ce93cb99db 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index 4927ad9af9..57fd1803d8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index 8ac21fe1d4..d114e6ea28 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index e86dbe0b34..8ff57ae088 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index d673714d2d..b05e5f5740 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index f7159562ee..6c166b0bcb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index 226884b03c..ad77e0345a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index f79457cd7a..12a67d9fbe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index 4432602fe4..b03da5daa3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index 12e49a7878..5e60cefd38 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index feb7d5bbc6..1f930c93ae 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 25423f664b..c4a719efe7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index 05cc163a49..f114a5c589 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index 12feccf87f..45aa0a2fe8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index c536310f02..c98203cdec 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index 61c4d3a8de..d53b708620 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 2768ccc918..27047b70a0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index 4ae2194a9e..0e1c0e1177 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index 67aa3025b4..fa7f1e4ccd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index 36c4b2a2b1..efac59087b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 1387967b63..622ef6eeba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index 2c4c88bb40..ab8568a484 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index eb54aeb09a..3d945bd566 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll index 55dc3933b4..77b23f869d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll index 3f9bc48017..7966e4c053 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll index 5f4269e53f..2b8ee3ef55 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll index 95a4a45603..3d2ee3f4a7 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll index 9bc1cc0423..8ecc38eb59 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll index 72d0906084..72fe5ac04d 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll index d61623d712..ea05110031 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll index 8ee24ea81b..b198fcaef7 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll index 543fb4b31b..ce4492853d 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll index c11170ecc2..4f5b445493 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index b91c145697..9172905973 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll index ccec4214ac..03c7e418c9 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_x.ll b/cuda-oxide/intrinsics/probes/thread_idx_x.ll index 1d94d11e56..0120318154 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_y.ll b/cuda-oxide/intrinsics/probes/thread_idx_y.ll index 7e65898e09..9e0150b697 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_z.ll b/cuda-oxide/intrinsics/probes/thread_idx_z.ll index abce05daa7..8da2aea7b1 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence.ll b/cuda-oxide/intrinsics/probes/threadfence.ll index 86d8442c25..f5b0070374 100644 --- a/cuda-oxide/intrinsics/probes/threadfence.ll +++ b/cuda-oxide/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_block.ll b/cuda-oxide/intrinsics/probes/threadfence_block.ll index be7273dce8..67b8785a11 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_block.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_system.ll b/cuda-oxide/intrinsics/probes/threadfence_system.ll index 0d614edec0..a8d252831f 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_system.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll index 1a48217b60..463d6e9eed 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll index f59485847e..361e2e45a2 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/trap.ll b/cuda-oxide/intrinsics/probes/trap.ll index 2e09de8887..9bf29f424a 100644 --- a/cuda-oxide/intrinsics/probes/trap.ll +++ b/cuda-oxide/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/uni_sync.ll b/cuda-oxide/intrinsics/probes/uni_sync.ll index df1c4d05c7..b2b6e9368c 100644 --- a/cuda-oxide/intrinsics/probes/uni_sync.ll +++ b/cuda-oxide/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll index c2e3f4f1f0..1040df1f9a 100644 --- a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll index bdab62f7f4..1c4b89d6bb 100644 --- a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll index 651e3114aa..26c835e46e 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_fence.ll b/cuda-oxide/intrinsics/probes/wgmma_fence.ll index ed11cb5c62..14e94865cf 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_fence.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll index 96d32d7b59..9c97434c56 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 27a44de2ff0c6443e82dfe9a9a11f5a4a9d09eb92c8696e2f6bc02469bb866a9. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" From e20462a1c5f4429df4f6b9df0ed3ca56a0c40602 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 20:44:22 +0530 Subject: [PATCH 08/17] chore(intrinsics): regenerate the catalog outputs after the second merge The merged overlay holds main's bf16x2 to FP8 pack admission (i1030 to i1033) next to the CTA-local TMA entries (i1034 to i1038), so the catalog hash and every generated, probe and reference file change. Generated by `cargo run -p cuda-intrinsics-gen -- generate` with no hand edits. Signed-off-by: nihalpasham --- .../cuda-device/src/generated/async_copy.rs | 2 +- .../cuda-device/src/generated/atomic.rs | 2 +- .../crates/cuda-device/src/generated/bf16.rs | 2 +- .../cuda-device/src/generated/bf16x2.rs | 2 +- .../crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- .../cuda-device/src/generated/convert.rs | 2 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- .../cuda-device/src/generated/dotprod.rs | 2 +- .../crates/cuda-device/src/generated/f16.rs | 2 +- .../crates/cuda-device/src/generated/f16x2.rs | 2 +- .../crates/cuda-device/src/generated/f32x2.rs | 2 +- .../crates/cuda-device/src/generated/fence.rs | 2 +- .../crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- .../cuda-device/src/generated/grid_sreg.rs | 2 +- .../crates/cuda-device/src/generated/i16x2.rs | 2 +- .../crates/cuda-device/src/generated/int.rs | 2 +- .../cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- .../cuda-device/src/generated/movmatrix.rs | 2 +- .../crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- .../crates/cuda-device/src/generated/sreg.rs | 2 +- .../cuda-device/src/generated/stmatrix.rs | 2 +- .../cuda-device/src/generated/tcgen05.rs | 2 +- .../crates/cuda-device/src/generated/tma.rs | 100 +- .../cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- .../cuda-intrinsics/src/generated/abi_v1.rs | 113 +- .../cuda-intrinsics/src/generated/mod.rs | 47 +- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 2 +- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 347 ++- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 292 ++- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- .../src/ops/generated/active_mask.rs | 2 +- .../dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- .../dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 2 +- .../dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- .../dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- .../dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- .../dialect-nvvm/src/ops/generated/tma.rs | 82 +- .../dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 27 +- .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 2 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 2 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 182 +- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 37 +- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 2 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 72 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../src/generated_intrinsics.rs | 22 +- cuda-oxide/intrinsics/catalog.json | 2042 ++++++++++++++--- cuda-oxide/intrinsics/generated-reference.md | 72 +- cuda-oxide/intrinsics/probes/abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/active_mask.ll | 2 +- cuda-oxide/intrinsics/probes/add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f64.ll | 2 +- .../intrinsics/probes/add_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f64.ll | 2 +- .../intrinsics/probes/add_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f64.ll | 2 +- .../intrinsics/probes/add_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f64.ll | 2 +- .../intrinsics/probes/add_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/ballot_sync.ll | 2 +- .../probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- .../intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_sync.ll | 2 +- .../probes/barrier_cta_sync_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- .../probes/clc_query_is_canceled.ll | 2 +- .../intrinsics/probes/clc_try_cancel.ll | 2 +- .../probes/clc_try_cancel_multicast.ll | 2 +- .../intrinsics/probes/clock.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock.llvm_nvptx.ll | 2 +- .../intrinsics/probes/clock64.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- .../intrinsics/probes/cluster_block_count.ll | 2 +- .../intrinsics/probes/cluster_block_idx_x.ll | 2 +- .../intrinsics/probes/cluster_block_idx_y.ll | 2 +- .../intrinsics/probes/cluster_block_idx_z.ll | 2 +- .../intrinsics/probes/cluster_block_rank.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_z.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_x.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_y.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_z.ll | 2 +- .../intrinsics/probes/cos_approx_f32.ll | 2 +- .../intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_cta.ll | 2 +- .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s_cta.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- .../probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- .../intrinsics/probes/cp_async_ca_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_8.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- .../intrinsics/probes/cp_async_cg_16.ll | 2 +- .../intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- .../probes/cp_async_commit_group.ll | 2 +- .../probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- .../intrinsics/probes/cp_async_wait_all.ll | 2 +- .../intrinsics/probes/cp_async_wait_group.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f64.ll | 2 +- .../intrinsics/probes/div_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f64.ll | 2 +- .../intrinsics/probes/div_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f64.ll | 2 +- .../intrinsics/probes/div_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f64.ll | 2 +- .../intrinsics/probes/div_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_u32.ll | 2 +- .../intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- .../intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- .../probes/elect_sync.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- .../intrinsics/probes/ex2_approx_f16.ll | 2 +- .../intrinsics/probes/ex2_approx_f32.ll | 2 +- .../intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- cuda-oxide/intrinsics/probes/fma_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll | 2 +- .../intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- .../intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- .../intrinsics/probes/fma_relu_bf16x2.ll | 2 +- .../intrinsics/probes/fma_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f64.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f64.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f64.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f64.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll | 2 +- .../intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- .../probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- .../intrinsics/probes/grid_dependency_wait.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_z.ll | 2 +- .../intrinsics/probes/gridid.lib_nvvm.ll | 2 +- .../intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/lane_id.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_eq.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_ge.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_gt.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_le.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- .../intrinsics/probes/lg2_approx_f32.ll | 2 +- .../intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/map_shared_rank.ll | 2 +- .../intrinsics/probes/match_all_i64_sync.ll | 2 +- .../intrinsics/probes/match_all_sync.ll | 2 +- .../intrinsics/probes/match_any_i64_sync.ll | 2 +- .../intrinsics/probes/match_any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16.ll | 2 +- .../intrinsics/probes/max_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/max_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/max_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_u16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- .../probes/max_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- .../probes/max_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/mbarrier_arrive.ll | 2 +- .../probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_init.ll | 2 +- .../intrinsics/probes/mbarrier_inval.ll | 2 +- .../intrinsics/probes/mbarrier_test_wait.ll | 2 +- .../intrinsics/probes/mbarrier_try_wait.ll | 2 +- .../probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16.ll | 2 +- .../intrinsics/probes/min_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/min_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/min_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_u16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- .../probes/min_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- .../probes/min_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- .../probes/mma_m16n8k16_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- .../intrinsics/probes/movmatrix_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/mul_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f64.ll | 2 +- .../intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f64.ll | 2 +- .../intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f64.ll | 2 +- .../intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f64.ll | 2 +- .../intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/nanosleep.ll | 2 +- cuda-oxide/intrinsics/probes/neg_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/neg_f16x2.ll | 2 +- .../intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- .../probes/packed_atomic_add_bf16x2.ll | 2 +- .../probes/packed_atomic_add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/pmevent.ll | 2 +- .../probes/prefetch_tma_descriptor.ll | 2 +- cuda-oxide/intrinsics/probes/prmt.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_b4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecl.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecr.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_f4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc16.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc8.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f64.ll | 2 +- .../intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f64.ll | 2 +- .../intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f64.ll | 2 +- .../intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f64.ll | 2 +- .../intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/redux_sync_add.ll | 2 +- .../intrinsics/probes/redux_sync_and.ll | 2 +- .../probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_i32.ll | 2 +- .../probes/redux_sync_max_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_u32.ll | 2 +- .../probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_i32.ll | 2 +- .../probes/redux_sync_min_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_or.ll | 2 +- .../intrinsics/probes/redux_sync_xor.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f32.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f64.ll | 2 +- .../intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/setmaxnreg_dec.ll | 2 +- .../intrinsics/probes/setmaxnreg_inc.ll | 2 +- .../probes/shuffle_down_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_down_sync.ll | 2 +- .../probes/shuffle_down_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_sync.ll | 2 +- .../intrinsics/probes/shuffle_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- .../intrinsics/probes/sin_approx_f32.ll | 2 +- .../intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll | 2 +- .../intrinsics/probes/smid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/sqrt_approx_f32.ll | 2 +- .../intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll | 2 +- .../intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll | 2 +- .../intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll | 2 +- .../intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll | 2 +- .../intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/sub_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sync_mask.ll | 2 +- cuda-oxide/intrinsics/probes/sync_threads.ll | 2 +- .../intrinsics/probes/tanh_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc.ll | 2 +- .../intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_commit.ll | 2 +- .../intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- .../probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_load_wait.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_shared.ll | 2 +- .../probes/tcgen05_mma_sp_shared.ll | 2 +- .../probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- .../probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_shift_down.ll | 2 +- .../probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- .../intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- .../probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence.ll | 2 +- .../intrinsics/probes/threadfence_block.ll | 2 +- .../intrinsics/probes/threadfence_system.ll | 2 +- .../probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/trap.ll | 2 +- cuda-oxide/intrinsics/probes/uni_sync.ll | 2 +- .../intrinsics/probes/warpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/wgmma_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_fence.ll | 2 +- .../intrinsics/probes/wgmma_wait_group.ll | 2 +- 1325 files changed, 4396 insertions(+), 1663 deletions(-) diff --git a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs index 03e050e5c8..6005b2eec4 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs index 2dae6c0395..294731d29b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs index f452534a7a..f7e9a219bf 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs index 13faf64f3d..ec3d6daafd 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/clc.rs b/cuda-oxide/crates/cuda-device/src/generated/clc.rs index 70cf419843..600b958d7c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/clc.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs index e227ffaee8..40445e21bb 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs index fbff9c9153..94b5fa04d9 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs index 54ac3d92f6..400c6d0dee 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::cluster`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/convert.rs b/cuda-oxide/crates/cuda-device/src/generated/convert.rs index 883d367a2d..14d972971c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/convert.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::convert`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs index a51ac2a765..26128f3f27 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs index fc4f94874a..5ccf651049 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs index 1c38bef616..6271828e97 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs index 11bcebcedd..d36bd2e52f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16.rs b/cuda-oxide/crates/cuda-device/src/generated/f16.rs index 37bb5d2051..dd1f10a879 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs index e0ba7d1157..dbaa705cd1 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs index 82010ea7d7..6ba54ef4b3 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/fence.rs b/cuda-oxide/crates/cuda-device/src/generated/fence.rs index 7025a6e77a..1a932ef845 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/fence.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/float.rs b/cuda-oxide/crates/cuda-device/src/generated/float.rs index bb17949ced..16a109f312 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/float.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::float`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs index 7660a74197..32d74cbc0d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs index e80f63b6cc..aaf16a69b4 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs index 8079ed880c..e52f271074 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/int.rs b/cuda-oxide/crates/cuda-device/src/generated/int.rs index 91e59c67f2..2a130f9a7c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/int.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs index 15c90d0aea..7213f1a853 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs index 42420ffe43..ac64208cde 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs index 499be8e9ed..ab37209b13 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs index 958ac2da4d..6cafa5dff3 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs index ebbac73f06..7068c9f98a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs index f45cc3ae7f..dc19ccc2dd 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs index 654e3954ed..8b99762cc6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs index 8e2dd4cde0..7cc59247f0 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs index 80526e716b..5e69b8f2f5 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs index 6f178f6014..dfde583a27 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs index 25577a893e..433011ffcc 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs index 508f5eada4..e1a13e41b9 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tma.rs b/cuda-oxide/crates/cuda-device/src/generated/tma.rs index 8800c3a641..93cb9d720b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tma` to keep its public API stable. @@ -233,6 +233,23 @@ pub unsafe fn cp_async_bulk_tensor_1d_g2s( unreachable!("cp_async_bulk_tensor_1d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_1d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, barrier); + unreachable!("cp_async_bulk_tensor_1d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -263,6 +280,24 @@ pub unsafe fn cp_async_bulk_tensor_2d_g2s( unreachable!("cp_async_bulk_tensor_2d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_2d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, barrier); + unreachable!("cp_async_bulk_tensor_2d_g2s_cta called outside CUDA kernel context") +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// # Safety @@ -329,6 +364,25 @@ pub unsafe fn cp_async_bulk_tensor_3d_g2s( unreachable!("cp_async_bulk_tensor_3d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_3d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, barrier); + unreachable!("cp_async_bulk_tensor_3d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -363,6 +417,26 @@ pub unsafe fn cp_async_bulk_tensor_4d_g2s( unreachable!("cp_async_bulk_tensor_4d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +pub unsafe fn cp_async_bulk_tensor_4d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + barrier: *mut Barrier, +) { + let _ = (dst, tensor_map, coord0, coord1, coord2, coord3, barrier); + unreachable!("cp_async_bulk_tensor_4d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety @@ -402,6 +476,30 @@ pub unsafe fn cp_async_bulk_tensor_5d_g2s( unreachable!("cp_async_bulk_tensor_5d_g2s called outside CUDA kernel context") } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// # Safety +/// `dst`, `tensor_map`, and `barrier` must remain valid until the copy completes. +/// `barrier` must point into the issuing CTA's shared memory, as for the other +/// `mbarrier` operations. `dst` is a [`SharedPtr`], which points there by its type. +#[inline(never)] +#[allow(clippy::too_many_arguments)] +pub unsafe fn cp_async_bulk_tensor_5d_g2s_cta( + dst: SharedPtr, + tensor_map: *const TmaDescriptor, + coord0: i32, + coord1: i32, + coord2: i32, + coord3: i32, + coord4: i32, + barrier: *mut Barrier, +) { + let _ = ( + dst, tensor_map, coord0, coord1, coord2, coord3, coord4, barrier, + ); + unreachable!("cp_async_bulk_tensor_5d_g2s_cta called outside CUDA kernel context") +} + /// Starts a TMA tile copy from shared to global memory. /// /// # Safety diff --git a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs index 8f1aa708c4..557554a1af 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs index 750c0d2d44..ba8dd5c16c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 698a64daf9..f657dd4665 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Raw ABI functions recognized by cuda-oxide. @@ -1181,6 +1181,22 @@ pub unsafe fn i0328(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u6 ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_1d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1034(_arg0: *mut u8, _arg1: *const u8, _arg2: i32, _arg3: *mut u64) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_1d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d`; expects PTX `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1217,6 +1233,28 @@ pub unsafe fn i0329( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_2d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1035( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035` executed outside device compilation" + ) +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. /// /// Catalog ID: `cp_async_bulk_tensor_2d_g2s_multicast`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d`; expects PTX `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;`. @@ -1298,6 +1336,29 @@ pub unsafe fn i0332( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_3d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1036( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_3d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d`; expects PTX `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1336,6 +1397,30 @@ pub unsafe fn i0333( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_4d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[inline(never)] +pub unsafe fn i1037( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_4d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d`; expects PTX `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];`. @@ -1383,6 +1468,32 @@ pub unsafe fn i0334( ) } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +/// +/// Catalog ID: `cp_async_bulk_tensor_5d_g2s_cta`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];`. +/// Available on `sm_90+` targets from PTX 8.6. +/// +/// # Safety +/// The destination and barrier must lie in the issuing CTA's own shared memory, and the tensor map must be a live descriptor for this dimensionality. +/// A generic destination or barrier pointer is converted with `cvta.to.shared`, so a cluster-shared pointer from `map_shared_rank` would yield a wrong local offset; copies into another CTA's shared memory use the cluster form. +/// Keep every object alive until the asynchronous copy completes. Only the designated issuing thread may start this transfer. +#[allow(clippy::too_many_arguments)] +#[inline(never)] +pub unsafe fn i1038( + _arg0: *mut u8, + _arg1: *const u8, + _arg2: i32, + _arg3: i32, + _arg4: i32, + _arg5: i32, + _arg6: i32, + _arg7: *mut u64, +) -> () { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038` executed outside device compilation" + ) +} + /// Starts a TMA tile copy from shared to global memory. /// /// Catalog ID: `cp_async_bulk_tensor_5d_s2g`. Source: LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d`; expects PTX `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];`. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs index 814cc0e023..c6488d7acb 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. #[doc(hidden)] @@ -1098,6 +1098,11 @@ pub mod tma { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0984 as cp_async_bulk_tensor_reduce_xor_im2col_3d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0985 as cp_async_bulk_tensor_reduce_xor_im2col_4d; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0986 as cp_async_bulk_tensor_reduce_xor_im2col_5d; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1034 as cp_async_bulk_tensor_1d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1035 as cp_async_bulk_tensor_2d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1036 as cp_async_bulk_tensor_3d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1037 as cp_async_bulk_tensor_4d_g2s_cta; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1038 as cp_async_bulk_tensor_5d_g2s_cta; } /// Generated `warp` intrinsic source API. @@ -1694,6 +1699,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_1d_g2s_cta_reexports_abi_i1034() { + let public: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_1d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1034; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_1d_s2g_reexports_abi_i0335() { let public: unsafe fn(*const u8, *const u8, i32) -> () = super::tma::cp_async_bulk_tensor_1d_s2g; @@ -1710,6 +1723,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_2d_g2s_cta_reexports_abi_i1035() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_2d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1035; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_2d_g2s_multicast_reexports_abi_i0330() { let public: unsafe fn(*mut u8, *const u8, i32, i32, *mut u64, u16) -> () = super::tma::cp_async_bulk_tensor_2d_g2s_multicast; @@ -1742,6 +1763,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_3d_g2s_cta_reexports_abi_i1036() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_3d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1036; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_3d_s2g_reexports_abi_i0337() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_3d_s2g; @@ -1758,6 +1787,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_4d_g2s_cta_reexports_abi_i1037() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_4d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1037; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_4d_s2g_reexports_abi_i0338() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_4d_s2g; @@ -1774,6 +1811,14 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cp_async_bulk_tensor_5d_g2s_cta_reexports_abi_i1038() { + let public: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::tma::cp_async_bulk_tensor_5d_g2s_cta; + let canonical: unsafe fn(*mut u8, *const u8, i32, i32, i32, i32, i32, *mut u64) -> () = + super::__cuda_oxide_intrinsic_abi_v1::i1038; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cp_async_bulk_tensor_5d_s2g_reexports_abi_i0339() { let public: unsafe fn(*const u8, *const u8, i32, i32, i32, i32, i32) -> () = super::tma::cp_async_bulk_tensor_5d_s2g; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index 093d3bf025..762b7fe476 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index 688a055fe9..fce5194341 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index 2c88800e44..6080a13962 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index 49d2527e37..35c3b5a4b8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index 7224a18db8..ef7573f553 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index 48b8883a5b..f1c34aea50 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index f451ca2cd1..295b13295f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 537884cd43..4b884cc74e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index e1b206d735..8b17848643 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index 58981a1e8d..c0dbaa2d27 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index 89807769de..988a33ef02 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index 352537c282..1109ff8fa8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index c8c5f39b6d..2cf4b179a8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index efc2c5af98..cca8493bee 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index a9f16d3a87..06501cfd4d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index 75742c9f56..21ccb310a2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 1d0c819e57..662cd82879 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index 47a457cd82..9a6180d16a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index d32c5ca869..429b6f3f2a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 6e2ea55160..b2fb0a2fbc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index cb61ea288e..15a6571b36 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index 0944315314..2d614270c8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index 22caa376f5..7c011622fe 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 45b6c0a6c0..1246212991 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index 12ee214154..95db81f709 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 3fad449c57..77aac0373f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index 702bb761a2..dce8c8e251 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index ee826a9198..e006284d67 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index 90a5a6f5cf..c545d4bc52 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index 2eaa0883ba..05ad33c613 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index 3ba5b6bc94..952755650a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index 5b33dcaab4..c0ea24b296 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index 966f5d1253..c996991f72 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index 4391014223..91934ba52d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index b26d920818..6b92fbf3d0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index a91888a9e9..dfb74bc785 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index 28ac52b10c..4ab2393cd0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index 8bfd36c97c..82ea1f7331 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index 25eaa798b7..075b1b73bb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index ebcf09356a..5346461528 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index 0fba32d40b..e9eb94805e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index bde2e5e030..41ba8a6149 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index e76d955ec1..b45dec5b80 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 75ddd78bd0..deacdfef77 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index 035edab564..6a4498941d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 8e53cab669..6d66affba9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index 02ebdb7bd2..261b4ac119 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 1dec957e1f..7b70285b88 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index 6af07e4e16..bb920b133a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index 8a1c0f797f..7508690da5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests, grouped per family shard. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index b4d35cc175..3a7c7de799 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 4f85f04af4..ddf741939f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index 49d50ea99c..2ecf439f62 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index e8995a5f9c..674c543f54 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index ab4efef6dd..7387fa58fa 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index fdbce3ea50..28a2cbc78f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index ff14c1f0f1..e568072fbf 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index a3142f7585..715449d0de 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index 5a905c322f..380646c942 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index a752dbd24f..45cac4183a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index 30bf7c6ad7..4c39a3e6dc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index 699539ac93..c8650d137b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index 1d3657b881..81e17610ba 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index 948959dcac..1e75444e5d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index f4730fecbe..d0582987c9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index dc2676f8b2..5d32dcefa2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index 3e29504925..0e30bd672d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 8ae2ee5ada..96d2c473bb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index 7ed5e540eb..28611a1740 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index 4c5fc7ce95..91e22ce0d7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index 155020b735..6db2425d65 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index 6468154b5c..4300e46d3a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tma` intrinsics. @@ -6643,3 +6643,348 @@ fn cp_async_bulk_tensor_reduce_xor_im2col_5d_target_matches_the_catalog() { ] as &[&str] ); } + +#[test] +fn cp_async_bulk_tensor_1d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1034").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_1d_g2s_cta"); + assert_eq!(target.abi_id, "i1034"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_2d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1035").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_2d_g2s_cta"); + assert_eq!(target.abi_id, "i1035"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_3d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1036").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_3d_g2s_cta"); + assert_eq!(target.abi_id, "i1036"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_4d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1037").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_4d_g2s_cta"); + assert_eq!(target.abi_id, "i1037"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} + +#[test] +fn cp_async_bulk_tensor_5d_g2s_cta_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1038").unwrap(); + assert_eq!(target.id, "cp_async_bulk_tensor_5d_g2s_cta"); + assert_eq!(target.abi_id, "i1038"); + assert_eq!( + target.dialect_op, + "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 86); + assert_eq!( + target.requirement.hardware, + const { GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm(90)]) } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ] + } + ); + assert_eq!( + target.selections, + &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0 + }] + } + },] + ); + assert_eq!( + target.llvm.unwrap().properties, + &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ] as &[&str] + ); +} diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 7aadebb14d..454968edb1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index d1faf65d46..8bdc1e7cd4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index 404e804d43..a82f032380 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index d69610fcb5..f9bda00d63 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 86ec25cef1..f553a689be 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index ac7d58f3ee..9beadfb3b0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tma` intrinsics. @@ -5670,4 +5670,294 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ result_range: None, }), }, + GeneratedIntrinsicTarget { + marker: "v1:i1034", + id: "cp_async_bulk_tensor_1d_g2s_cta", + abi_id: "i1034", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_1D", + asm: "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 5, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1035", + id: "cp_async_bulk_tensor_2d_g2s_cta", + abi_id: "i1035", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_2D", + asm: "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 6, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1036", + id: "cp_async_bulk_tensor_3d_g2s_cta", + abi_id: "i1036", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_3D", + asm: "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 7, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1037", + id: "cp_async_bulk_tensor_4d_g2s_cta", + abi_id: "i1037", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_4D", + asm: "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 8, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1038", + id: "cp_async_bulk_tensor_5d_g2s_cta", + abi_id: "i1038", + dialect_op: "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[GeneratedHardwareAlternative::MinimumSm( + 90, + )]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(86), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::MinimumSm(90), + ]), + }, + }, + ], + selections: &[GeneratedSelectionAlternative { + source_record: "TMA_G2S_CTA_TILE_5D", + asm: "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + predicates: &[ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86", + ], + constraints: GeneratedSelectionConstraints { + address_space: None, + immediate_bindings: &[GeneratedImmediateBinding { + argument_index: 9, + value: 0, + }], + }, + }], + llvm: Some(GeneratedLlvmFacts { + properties: &[ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly", + ], + result_no_undef: false, + result_range: None, + }), + }, ]; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 4bfa775523..e83ef298c3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index cccaf4dbd6..cc7a79ffc4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index 000d21b379..02d5833fc5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index 878925151f..1f5ded8fb5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index 30055018e9..bdb2b14135 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs index a515bdbbff..0285163635 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for the generated active warp mask. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs index 4d251c3589..f7f9f0bd6d 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Cluster Launch Control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index 73b38afd48..47116163ae 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated operations for the closed cluster-barrier family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index b6a6c99840..83490c3b23 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs index 553c161a53..6b381fdccc 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for classic `cp.async` instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs index 101fc85f8a..5842b7f177 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated PTX debug controls. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs index 4acdd93b37..e9be20852f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed integer dot products. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs index a8612264f5..80e4e0b729 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated warp leader-election operation. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs index f998bbb13f..e7f4eb8258 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index e9f503bb0b..83d99129d4 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for extended floating-point min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 7013c998e5..963fb56a73 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated extended integer min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index 74ec271c15..4e5bfb1ee6 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index ea844c8cf2..59068a3dba 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index d5de90d50f..fdf4c8fa5c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs index 2cb7aff92e..555ffa1256 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. mod active_mask; diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index 88980724ce..0e8a6e9ce0 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated in-register matrix transpose. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index 51e660245f..380726ac16 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed floating-point arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index 8c19c30ae9..d608bad233 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated packed-atomic family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index fa83a5345f..03eb96c079 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated packed conversion. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs index f147341d0b..2cd3e3b934 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated `prmt` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs index c6bcdbf3bc..c3914109d1 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 4767e487c6..7bb191879c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index fa498fbe9b..dc1d8c0d30 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index d6a767a80c..4495afa83e 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index 0cdce1daac..763ac4cb97 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar math. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 9aaa2f04c4..45386eda8b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated sparse MMA variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs index f63a3efacb..a6a8000b17 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for generated special-register reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index 35fad6b583..40e487b4c2 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the four generated `stmatrix` stores. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs index 212c36f829..f58708c173 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated CTA synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index 88b913dea1..be79474f4b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Core Generation 5 operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs index de5e56756b..4c100e828e 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Memory Accelerator operations. @@ -227,6 +227,21 @@ impl CpAsyncBulkTensorG2sTile1dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<5>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile1dOp; + +impl CpAsyncBulkTensorG2sCtaTile1dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_1d", @@ -257,6 +272,21 @@ impl CpAsyncBulkTensorG2sTile2dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<6>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile2dOp; + +impl CpAsyncBulkTensorG2sCtaTile2dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a multicast TMA tile copy from global to cluster shared memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast", @@ -317,6 +347,21 @@ impl CpAsyncBulkTensorG2sTile3dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<7>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile3dOp; + +impl CpAsyncBulkTensorG2sCtaTile3dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_3d", @@ -347,6 +392,21 @@ impl CpAsyncBulkTensorG2sTile4dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<8>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile4dOp; + +impl CpAsyncBulkTensorG2sCtaTile4dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_4d", @@ -377,6 +437,21 @@ impl CpAsyncBulkTensorG2sTile5dOp { } } +/// Starts a TMA tile copy from global to the issuing CTA's shared memory. +#[pliron_op( + name = "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", + format, + verifier = "succ", + interfaces = [NOpdsInterface<9>, NResultsInterface<0>], +)] +pub struct CpAsyncBulkTensorG2sCtaTile5dOp; + +impl CpAsyncBulkTensorG2sCtaTile5dOp { + pub fn new(op: Ptr) -> Self { + Self { op } + } +} + /// Starts a TMA tile copy from shared to global memory. #[pliron_op( name = "nvvm.cp_async_bulk_tensor_s2g_tile_5d", @@ -1697,16 +1772,21 @@ pub(super) fn register(ctx: &mut Context) { CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op::register(ctx); CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp::register(ctx); CpAsyncBulkTensorG2sTile1dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile1dOp::register(ctx); CpAsyncBulkTensorS2gTile1dOp::register(ctx); CpAsyncBulkTensorG2sTile2dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastOp::register(ctx); CpAsyncBulkTensorG2sTile2dMulticastCg2Op::register(ctx); CpAsyncBulkTensorS2gTile2dOp::register(ctx); CpAsyncBulkTensorG2sTile3dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile3dOp::register(ctx); CpAsyncBulkTensorS2gTile3dOp::register(ctx); CpAsyncBulkTensorG2sTile4dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile4dOp::register(ctx); CpAsyncBulkTensorS2gTile4dOp::register(ctx); CpAsyncBulkTensorG2sTile5dOp::register(ctx); + CpAsyncBulkTensorG2sCtaTile5dOp::register(ctx); CpAsyncBulkTensorS2gTile5dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col3dOp::register(ctx); CpAsyncBulkTensorReduceAddIm2col4dOp::register(ctx); diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs index 02e65111a9..a14d51fe26 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `vote.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 0698e06b23..196ebb1edd 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated warp synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs index 4472cf1ca4..8593a8d9d9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `match.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 611cf8e798..5b7121aaea 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `shfl.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index 5733dd83cb..d18eb1b157 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated WGMMA controls. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index ce0a280e2c..07bd68d3e0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index a1e5adb5a9..c8e4d5bae3 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index 05c3918a4a..8e0078ab18 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index c11ddf6108..28539f3b1b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index 3c840f9b5a..191abcfbc5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index a9e4e19ff4..e132908a90 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 3e7c7620da..69ef08528b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 59f961cb33..e6689bcb4c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index ee0c6b7551..5d882c857b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 967a901883..fbc2004a4a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index 9a1899a62f..1f0b35910e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 6f3fdd86f0..2f5d55c53a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index 7dfe1ccfb0..32efb100e5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index eb53af9b1e..20da6ede9b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index ad552194d0..cc0182a019 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index 9c980fca9e..6b65a34d53 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -313,10 +313,14 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -325,14 +329,20 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" @@ -2309,16 +2319,21 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3443,10 +3458,14 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" => Some("v1:i0328"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => Some("v1:i1034"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => Some("v1:i0335"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" => Some("v1:i0329"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => Some("v1:i1035"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => Some("v1:i0330"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" @@ -3455,14 +3474,20 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" => Some("v1:i0336"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" => Some("v1:i0332"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => Some("v1:i1036"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => Some("v1:i0337"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" => Some("v1:i0333"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => Some("v1:i1037"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => Some("v1:i0338"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" => Some("v1:i0334"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => Some("v1:i1038"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => Some("v1:i0339"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 2384dc6f85..58fbc4a634 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 480526765e..327b3af7c3 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index 32377c7d40..b171278603 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index b7cbacf93b..76f72e5213 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index a15e3196ad..b1bb0fc010 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index 9a630b3ed3..918c065730 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index 47ad59a700..084dc6b730 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index cbdf2bbda6..addefa3aa8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index 9e08cbdc77..2bb6ea0f92 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index 93fc0784ea..faac1b40c6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 993ac1a84f..aeb6f4692b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 97c70d8889..ecb4a00026 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index a0f41e97f6..0fab7612c9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index 484c1a6a74..883e6627df 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 9aa3a1c92e..859850ed4c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index d47e6abb1d..ea74b7936a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index eec859a000..73b27e24b7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index 223a7e2117..717c58f437 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index 599f8c5dfe..3feab3a35e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index c24ccc905f..e043060f75 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index da599ce05c..23dd6691e8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index a96747bd8c..722175dade 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index 6218b18f60..04ef58485f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index 5b14990157..4bba62c619 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index 160cdd2547..2fed06b33f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index 481dcfbe08..0c0e6867f7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index c73be2a520..d896c9f2c5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index 57351d2b92..3195737ca0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index b7ce30d142..375e65af30 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index 87f98b4dfd..65de75d12d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index da129938f1..1631dfdc23 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index a6235ba466..462d98f014 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 45f494fe2f..9a85710cc6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index 65795232e5..d6c89b19c7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index cb55540952..df2e3cc863 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 1f01e82de2..c3aabbec71 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index 05b1990903..6966a2e993 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 311c4f9703..55f8fe4474 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index 0928fdfbcc..c84bb61905 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index cbe3751ad6..0d242fcab5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index 38a8944891..ed8e0f6957 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index 11eebceed1..11f5adb322 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index 35a1416a22..3485763029 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index a6efe97fe2..2258c54f86 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 375f4cddd1..3607247775 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 62cb47f775..d9f6e3e460 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 7d2caa1ac9..484742f420 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 0dad5bf19f..1ea02524fb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 3eb2414b71..71ab9a6842 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 7cac99888a..f9fea098fe 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tma` intrinsics. @@ -529,6 +529,42 @@ fn cp_async_bulk_tensor_1d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_1d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034"), + Some("v1:i1034") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta"), + Some("v1:i1034") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1034" + )); +} +#[test] fn cp_async_bulk_tensor_1d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" @@ -601,6 +637,42 @@ fn cp_async_bulk_tensor_2d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_2d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035"), + Some("v1:i1035") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta"), + Some("v1:i1035") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1035" + )); +} +#[test] fn cp_async_bulk_tensor_2d_g2s_multicast_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" @@ -745,6 +817,42 @@ fn cp_async_bulk_tensor_3d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_3d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036"), + Some("v1:i1036") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta"), + Some("v1:i1036") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1036" + )); +} +#[test] fn cp_async_bulk_tensor_3d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" @@ -817,6 +925,42 @@ fn cp_async_bulk_tensor_4d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_4d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037"), + Some("v1:i1037") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta"), + Some("v1:i1037") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1037" + )); +} +#[test] fn cp_async_bulk_tensor_4d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" @@ -889,6 +1033,42 @@ fn cp_async_bulk_tensor_5d_g2s_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cp_async_bulk_tensor_5d_g2s_cta_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038"), + Some("v1:i1038") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta"), + Some("v1:i1038") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1038" + )); +} +#[test] fn cp_async_bulk_tensor_5d_s2g_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index 26513fe1f1..12870d14ff 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 1f5103a1e4..2c2b5aec29 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index 8db832c682..2ab93890d0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index f8f1fd992a..038ad862eb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index 97d4b2cacd..8347321374 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 8059ee6882..579000521c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. @@ -587,6 +587,13 @@ pub(super) fn try_dispatch( "v1:i0328", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 1, + "v1:i1034", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -601,6 +608,13 @@ pub(super) fn try_dispatch( "v1:i0329", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 2, + "v1:i1035", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" => { Ok(Some(super::super::tma::emit_tma_g2s_multicast( @@ -627,6 +641,13 @@ pub(super) fn try_dispatch( "v1:i0332", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 3, + "v1:i1036", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -641,6 +662,13 @@ pub(super) fn try_dispatch( "v1:i0333", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 4, + "v1:i1037", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( @@ -655,6 +683,13 @@ pub(super) fn try_dispatch( "v1:i0334", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" => { + Ok(Some(super::super::tma::emit_tma_g2s_cta( + ctx, body, args, target, block_ptr, prev_op, value_map, block_map, loc, 5, + "v1:i1038", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" => { Ok(Some(super::super::tma::emit_tma_s2g( diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 5cbe5f53af..68e89bfbc0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index c4e3307b00..7da22f606d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 8b562adb4b..ad74958552 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index bbb15ace98..a148cfb705 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index a9d7fadebc..b0b8cf5436 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 30c8776679..2ba93f5ee4 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index 8fa8586470..9b81219dbb 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index 4e2734e8d0..b72d3d677b 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 2829f83d49..eee115f3ce 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index 2bc5b88829..d336d75706 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index 5813b36f72..ab0700d55b 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 264a05534e..e613145a4d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index efc66dd8b1..018ae5b5cc 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index 73face4f76..ebf4ddb3ab 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index f128e318b5..1845fb6e49 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index a35c8be28e..fe90e07d48 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index 580155604d..e6b93926cc 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index f54e78bfa2..5bbc592460 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 916f51a07e..76cf328ca1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index bfb759ffac..8a8e14904c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index 49e86c6d9c..1a3bc2b87d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index a86511b63d..82c5b85ff3 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index 87d5e3e381..5bc05bfcd0 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index 5f778e31c9..94b3ebd97a 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 8e843de47f..1ffac3ad60 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index e8106051b9..bf6e7322cc 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index 719f6e4c06..16d330c5a7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index 8fe82b112f..2849836687 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index 278f2e00d4..af234e64e8 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index eeda1b07fb..bbb32996b7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index 300f1ed21d..7bbeff63d5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index 1989dcc8b6..6f6fb05c69 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 7b6f25b870..da17d4e5f7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index f6bea5c034..4f7cd0813f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index cdc12343b7..5c47b505a1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index d7306ca5bb..a34e313787 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tma` intrinsics. @@ -12,9 +12,9 @@ use crate::conversion_interface::MirToLlvmConversion; use crate::convert::intrinsics::{ tma::PrefetchTileConfig, tma::ReduceConfig, tma::convert_control, tma::convert_g2s, - tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, tma::convert_prefetch_tile, - tma::convert_reduce_s2g, tma::convert_s2g, tma::convert_tensormap_fence, - tma::convert_tensormap_replace, + tma::convert_g2s_cta, tma::convert_g2s_multicast_cg2, tma::convert_prefetch_tensormap, + tma::convert_prefetch_tile, tma::convert_reduce_s2g, tma::convert_s2g, + tma::convert_tensormap_fence, tma::convert_tensormap_replace, }; use dialect_nvvm::ops::{ CpAsyncBulkCommitGroupOp, CpAsyncBulkPrefetchTensor1dL2CacheHintOp, @@ -23,7 +23,9 @@ use dialect_nvvm::ops::{ CpAsyncBulkPrefetchTensor3dL2Op, CpAsyncBulkPrefetchTensor4dL2CacheHintOp, CpAsyncBulkPrefetchTensor4dL2Op, CpAsyncBulkPrefetchTensor5dL2CacheHintOp, CpAsyncBulkPrefetchTensor5dL2Op, CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2CacheHintOp, - CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sTile1dOp, + CpAsyncBulkPrefetchTensorGather4TwoDimensionalL2Op, CpAsyncBulkTensorG2sCtaTile1dOp, + CpAsyncBulkTensorG2sCtaTile2dOp, CpAsyncBulkTensorG2sCtaTile3dOp, + CpAsyncBulkTensorG2sCtaTile4dOp, CpAsyncBulkTensorG2sCtaTile5dOp, CpAsyncBulkTensorG2sTile1dOp, CpAsyncBulkTensorG2sTile2dMulticastCg2Op, CpAsyncBulkTensorG2sTile2dMulticastOp, CpAsyncBulkTensorG2sTile2dOp, CpAsyncBulkTensorG2sTile3dOp, CpAsyncBulkTensorG2sTile4dOp, CpAsyncBulkTensorG2sTile5dOp, CpAsyncBulkTensorReduceAddIm2col3dOp, @@ -383,6 +385,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile1dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile1dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 1) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile1dOp { fn convert( @@ -407,6 +421,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile2dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 2) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile2dMulticastOp { fn convert( @@ -455,6 +481,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile3dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile3dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 3) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile3dOp { fn convert( @@ -479,6 +517,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile4dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile4dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 4) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile4dOp { fn convert( @@ -503,6 +553,18 @@ impl MirToLlvmConversion for CpAsyncBulkTensorG2sTile5dOp { } } +#[op_interface_impl] +impl MirToLlvmConversion for CpAsyncBulkTensorG2sCtaTile5dOp { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + operands_info: &OperandsInfo, + ) -> Result<()> { + convert_g2s_cta(ctx, rewriter, self.get_operation(), operands_info, 5) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CpAsyncBulkTensorS2gTile5dOp { fn convert( diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index 8ec94c0b3d..219add4383 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index cfebdc1122..43a0524d5f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index 3516d587b3..7c6eb0c56c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index 2a8c876e71..05ab52282d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index 8fb47633da..dd8526e8dc 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index 9ee8fb794a..311200a1ae 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated collector predicates for intrinsic placeholder functions. @@ -99,16 +99,21 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929", @@ -1135,16 +1140,21 @@ pub(crate) const GENERATED_INTRINSIC_PUBLIC_PATHS: &[&str] = &[ "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2", "cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d", "cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d", @@ -2177,16 +2187,21 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0893" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0922" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0928" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0929" @@ -3220,16 +3235,21 @@ pub(crate) fn is_generated_intrinsic_compatibility_path(path: &str) -> bool { | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2" | "cuda_device::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint" | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_1d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast" | "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" | "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + | "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" | "cuda_device::tma::cp_async_bulk_tensor_5d_s2g" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_3d" | "cuda_device::tma::cp_async_bulk_tensor_reduce_add_im2col_4d" diff --git a/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index 52aea22d8d..a4918e40b1 100644 --- a/cuda-oxide/intrinsics/catalog.json +++ b/cuda-oxide/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "54d47b8e281dd953dc95fe16f834ff38164d9809171c9a1af097231b0bc45cab", - "overlay_sha256": "9a9ba81bc0e7a27a16ccf6a5254123c36ad98b47a8829bbc222cfb77173ff31e", - "abi_ledger_sha256": "0e8cec7fdd52ad28ff4365bea512101aef06f29ad7781dcde6b867ad93a0a5bc", + "overlay_sha256": "bdfaca84cf4ed51ee06ef7caf5ce1acc979849117f7fd7afcd65b8b8397bdb37", + "abi_ledger_sha256": "f4cf9e3a7cf9f40725d1e5a2645796167e9615c3c47d2998c67ab1c3952fb210", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -71,6 +71,7 @@ "cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json:7cc62aa25f9d696eaefcaedf104f53da9d08f24385f18d1539308f34d56d38f8", "cuda-13.3-libnvvm-13.3.73-register-mma-mxf8f6f4-f32.json:e437bf84570f4df0a28a743a132e0b93f359f9431d1af03b443f324f62b715ba", "cuda-13.3-libnvvm-13.3.73-sparse-mma-fp8.json:8a57d207ffd510b8c1bb40643bdb828d3d22e5d1c5ab24e1bc1b697538e9bc21", + "cuda-13.3-libnvvm-13.3.73-tma-cta.json:352f3e82f0f1c190d99e12e717ef5357a05729521a1fbab100b40964bc9526e7", "cuda-13.3-libnvvm-13.3.73-tma-descriptor-control.json:ac44f485122404f9a900cfb94bff658160a6c23d9e2f8ef74bb751f044b848fc", "cuda-13.3-libnvvm-13.3.73-tma-reduce.json:f5df5bb51068966594660cb01b1b28c3b72cb004c456fcd8dff377ff0bee4fc0", "cuda-13.3-libnvvm-13.3.73-tma.json:8607f93605f215858a98075a0ea1394f2718d090538bcaf0a795e11b4cb66d47", @@ -129,6 +130,7 @@ "rust-llvm-23.1.3-aab11ee0-tcgen05-st.json:e559aa775e3e964f9889483e9d9d8c45b46628931ff78050c46f33e448f69e86", "rust-llvm-23.1.3-aab11ee0-tcgen05.json:b6cf11d81c0880877ac5ad8958231222c728d5ab0243ddc84f1db901fbc8608b", "rust-llvm-23.1.3-aab11ee0-threadfence.json:087343e0247e7f9571c74fe763c0d48fbd26e246a21a9f10e5466a9a4457f83e", + "rust-llvm-23.1.3-aab11ee0-tma-cta.json:5aa68833705b5a98b393c4c752739a93cb3360d8d1c2c130bc28e09e475a208c", "rust-llvm-23.1.3-aab11ee0-tma-descriptor-control.json:7b6deaedfb8f8ac4b6551c47701cdefd04867431ac4dc5c3fbaf8f8bc604c9e9", "rust-llvm-23.1.3-aab11ee0-tma-reduce.json:e472b96476cad1d406fdac6dd8209645357f068cbd545669e2f09a8c6a232b79", "rust-llvm-23.1.3-aab11ee0-tma.json:97fb372aa7c6966bd01667da32d9d0857e385776131cb548f7d70f73b52718bc", @@ -18551,6 +18553,274 @@ }, "summary": "Starts a TMA tile copy from global to cluster shared memory." }, + { + "id": "cp_async_bulk_tensor_1d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.1d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d" + }, + "selections": [ + { + "source_record": "TMA_G2S_CTA_TILE_1D", + "asm": "cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0}}], [$mbar];", + "predicates": [ + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 5, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i1034", + "module": "tma", + "name": "cp_async_bulk_tensor_1d_g2s_cta", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sCtaTile1dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d", + "arguments": [ + "shared_ptr", + "shared_ptr", + "ptr", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probe for the 1D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_1d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,007-byte source had SHA-256 253d3ea39f9d26f5000f83febb9f72a15af2bed6ad0f668ab206976ce4c6d6ba; the 2,640-byte bitcode had SHA-256 9f42944557579c5c9c4e38716cce52607a07c51e5f2272393948fdc40cfb8961; the 1,234-byte canonical IR had SHA-256 835c2efc3b53d5333ec6a8260fcad29d5d0ce889194d0523feb57f7b22281fff.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "typed LLVM call for the 1D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,165-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 b0b4469f60fea3342f58e26eb4698834c6cb0714aac17fd3ffaed79cbd4f226a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the 1D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_cta_tile1d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "1d", + "shared::cta", + "global", + "tile", + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." + }, { "id": "cp_async_bulk_tensor_1d_s2g", "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.1d", @@ -19095,29 +19365,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "id": "cp_async_bulk_tensor_2d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.2d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_CTA_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 7, - "value": -1 - }, - { - "argument_index": 8, + "argument_index": 6, "value": 0 } ] @@ -19125,7 +19391,282 @@ } ], "rust": { - "abi_id": "i0330", + "abi_id": "i1035", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_cta", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sCtaTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d", + "arguments": [ + "shared_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probe for the 2D CTA-local TMA G2S operation", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_2d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,038-byte source had SHA-256 f0d47c1b6670854dfddce5bc45206d8f4ad16e9a05236dd4e1d4036b518666ea; the 2,648-byte bitcode had SHA-256 6ab53bfd2185eb8b6d7f7ffb3793d72cb7b48a07f61846a5317604700bdd3225; the 1,265-byte canonical IR had SHA-256 2c00613205683c093c5002f0f7347d3b2d6b82e46c5bc00072fb5a781f7d9d44.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "typed LLVM call for the 2D CTA-local TMA G2S operation", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,299-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 59fa9746f48c669cbff59faffea6039a796055f2721b8647a6331d78b5a26521.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx86" + ], + "representation": "LLVM-produced PTX retaining the 2D CTA-local TMA G2S instruction", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "linked cubin retaining five UTMALDG instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_cta_tile2d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cta", + "global", + "tile", + "mbarrier::complete_tx::bytes" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0330", "module": "tma", "name": "cp_async_bulk_tensor_2d_g2s_multicast", "arguments": [ @@ -19161,13 +19702,889 @@ "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_2D_MC", + "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "predicates": [ + "Subtarget->hasTMABlackwellSupport()" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 7, + "value": -1 + }, + { + "argument_index": 8, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0331", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "*mut u64", + "u16" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "arguments": [ + "shared_cluster_ptr", + "shared_ptr", + "ptr", + "i32", + "i32", + "i16", + "i64", + "i1", + "i1", + "i32" + ], + "results": [], + "properties": [ + "ImmArg", + "ImmArg", + "ImmArg", + "IntrConvergent", + "Range", + "ReadOnly", + "WriteOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + }, + "ptx_result": "()", + "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100a", + "ptx_feature": "+ptx86" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_100a" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 065ffbad7cc3f4a868f3de9e7f86d3a4fb1b0ea7047e1ddcdd578192e1fe101c.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 c6f52db1ccb491246b76348fb43a73fc961823d087d80432a2cec08dc36439f1.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx86" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 eea1fea3e3abbce7a6d377fafd8ad8e2c8fec998a990afbd664b7ce4f3856dff.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.6", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 101 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_100a", + "ptx93" + ], + "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_100a", + "ptx93" + ], + "representation": "linked cubin retaining the CTA-pair TMA instruction", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "g2s_tile2d_multicast_cg2", + "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "shared::cluster", + "global", + "tile", + "mbarrier::complete_tx::bytes", + "multicast::cluster", + "cta_group::2" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "address" + }, + { + "kind": "register" + } + ] + }, + "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + }, + { + "id": "cp_async_bulk_tensor_2d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + }, + "selections": [ + { + "source_record": "TMA_TENSOR_S2G_TILE_2D", + "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 5, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0336", + "module": "tma", + "name": "cp_async_bulk_tensor_2d_s2g", + "arguments": [ + "*const u8", + "*const u8", + "i32", + "i32" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorS2gTile2dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "operands": [ + "ptr", + "ptr", + "i32", + "i32" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "arguments": [ + "shared_ptr", + "ptr", + "i32", + "i32", + "i64", + "i1" + ], + "results": [], + "properties": [ + "ImmArg", + "IntrConvergent", + "ReadOnly", + "ReadOnly" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": false, + "memory": "read_write", + "convergent": true, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + }, + "ptx_result": "()", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_90", + "ptx_feature": "+ptx80" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_90" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx80" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "target": { + "minimum_ptx": "8.0", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 90 + } + ] + } + }, + "version": "CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0", + "sha256": "5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc", + "artifact_path": null, + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_90", + "ptx93" + ], + "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_90", + "ptx93" + ], + "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "artifact_kind": "cubin", + "tool_path": "libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa" + } + ] + } + ], + "tma": { + "operation": "s2g_tile2d", + "adapter": "s2g_pointers_coordinates_inject_defaults", + "runtime_validation": "unexecuted" + }, + "lowering": "generated_tma", + "expected_ptx": { + "mnemonic": "cp", + "modifiers": [ + "async", + "bulk", + "tensor", + "2d", + "global", + "shared::cta", + "tile", + "bulk_group" + ], + "operands": [ + { + "kind": "address" + }, + { + "kind": "address" + } + ] + }, + "summary": "Starts a TMA tile copy from shared to global memory." + }, + { + "id": "cp_async_bulk_tensor_3d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "family": "tma", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + }, + "selections": [ + { + "source_record": "TMA_G2S_TILE_CG0_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "predicates": [ + "Subtarget->getPTXVersion() >= 80", + "Subtarget->getSmVersion() >= 90" + ], + "constraints": { + "immediate_bindings": [ + { + "argument_index": 8, + "value": 0 + }, + { + "argument_index": 9, + "value": 0 + } + ] + } + } + ], + "rust": { + "abi_id": "i0332", + "module": "tma", + "name": "cp_async_bulk_tensor_3d_g2s", + "arguments": [ + "*mut u8", + "*const u8", + "i32", + "i32", + "i32", + "*mut u64" + ], + "result": "()", + "safe": false, + "must_use": false, + "safe_allowlist_reason": null, + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "compatibility_paths": [ + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + ] + }, + "dialect": { + "op_type": "CpAsyncBulkTensorG2sTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "operands": [ + "ptr", + "ptr", + "ptr", + "i32", + "i32", + "i32", + "i16", + "i64" + ], + "results": [] + }, + "llvm": { + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", "arguments": [ "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", + "i32", "i16", "i64", "i1", @@ -19176,11 +20593,11 @@ ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "ImmArg", "ImmArg", "IntrConvergent", - "Range", + "Range", "ReadOnly", "WriteOnly" ], @@ -19346,8 +20763,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19357,12 +20774,11 @@ "async", "bulk", "tensor", - "2d", + "3d", "shared::cluster", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster" + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19373,37 +20789,31 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.2d.multicast.cta_group_2", + "id": "cp_async_bulk_tensor_3d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_2D_MC", - "asm": "cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster$cg [$dst], [$tmap, {{$d0, $d1}}], [$mbar], $mc;", + "source_record": "TMA_G2S_CTA_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", "predicates": [ - "Subtarget->hasTMABlackwellSupport()" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { "argument_index": 7, - "value": -1 - }, - { - "argument_index": 8, "value": 0 } ] @@ -19411,62 +20821,57 @@ } ], "rust": { - "abi_id": "i0331", + "abi_id": "i1036", "module": "tma", - "name": "cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "name": "cp_async_bulk_tensor_3d_g2s_cta", "arguments": [ "*mut u8", "*const u8", "i32", "i32", - "*mut u64", - "u16" + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0331", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2" + "cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile2dMulticastCg2Op", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2", + "op_type": "CpAsyncBulkTensorG2sCtaTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d", "operands": [ "ptr", "ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", - "i16", + "i32", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ "ImmArg", - "ImmArg", - "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -19487,63 +20892,39 @@ "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] }, "ptx_result": "()", - "targets": "sm_100a|sm_101a|sm_103a|sm_110a", + "targets": "all", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.26.5 Asynchronous bulk tensor copy", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100a", + "gpu_target": "sm_90", "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19556,13 +20937,13 @@ "stages": [ { "targets": [ - "sm_100a" + "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 3D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 065ffbad7cc3f4a868f3de9e7f86d3a4fb1b0ea7047e1ddcdd578192e1fe101c.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_3d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,069-byte source had SHA-256 159ab50347ad4f81c8a110df7de35db35da9eb099938874523e5ec10ea867c8c; the 2,656-byte bitcode had SHA-256 3b5f19c93cffadf279efad33b501519f09cacbc660784f195ab3d007b5bca854; the 1,296-byte canonical IR had SHA-256 223dd701edae4ecf145ca18afc659dd5a50785cf8e492f6ba5c7b7f8299dc030.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19570,14 +20951,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 3D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 c6f52db1ccb491246b76348fb43a73fc961823d087d80432a2cec08dc36439f1.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,433-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 e9bab5f8e2c2a697ca435f86eaff468984a03470ae4bf28f975c2c89fc6676b9.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19585,14 +20966,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 3D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 eea1fea3e3abbce7a6d377fafd8ad8e2c8fec998a990afbd664b7ce4f3856dff.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -19603,27 +20984,15 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "exact_architecture", - "sm": 100 - }, - { - "kind": "exact_architecture", - "sm": 101 - }, - { - "kind": "exact_architecture", - "sm": 103 - }, - { - "kind": "exact_architecture", - "sm": 110 + "kind": "minimum_sm", + "sm": 90 } ] } @@ -19636,14 +21005,14 @@ "stages": [ { "targets": [ - "compute_100a", + "compute_90", "ptx93" ], - "representation": "one NVVM IR 2.0 module containing the exact CTA-pair inline-PTX form", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained the exact cta_group::2 instruction with no extern call. The source IR SHA-256 is be32e3ad9cf01c97078b9a6799a994afc5e1c1264fd0ad0831b2e98f6f2275ca, LTOIR SHA-256 is ebc556f17c97cd4da6eb87512373c3d93bd44ba3faf68db41206f934a22a60ac, and linked PTX SHA-256 is 8b48f700abf6f55e3dc13443bf1ffe7a45f639cd16f683443cfbf5c84feacc38.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -19651,14 +21020,14 @@ }, { "targets": [ - "sm_100a", + "sm_90", "ptx93" ], - "representation": "linked cubin retaining the CTA-pair TMA instruction", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a valid sm_100a CUDA ELF cubin with SHA-256 07a9b441996fc5ba0b8a05aebf0ce67218f0ea8e75f45e4c652a0778207a5343.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -19668,8 +21037,8 @@ } ], "tma": { - "operation": "g2s_tile2d_multicast_cg2", - "adapter": "g2s_pointers_coordinates_barrier_mask_inject_defaults", + "operation": "g2s_cta_tile3d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -19679,13 +21048,11 @@ "async", "bulk", "tensor", - "2d", - "shared::cluster", + "3d", + "shared::cta", "global", "tile", - "mbarrier::complete_tx::bytes", - "multicast::cluster", - "cta_group::2" + "mbarrier::complete_tx::bytes" ], "operands": [ { @@ -19696,26 +21063,23 @@ }, { "kind": "address" - }, - { - "kind": "register" } ] }, - "summary": "Starts a multicast TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_2d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.2d", + "id": "cp_async_bulk_tensor_3d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_2d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_2D", - "asm": "cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1}}], [$src];", + "source_record": "TMA_TENSOR_S2G_TILE_3D", + "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19723,7 +21087,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 5, + "argument_index": 6, "value": 0 } ] @@ -19731,49 +21095,52 @@ } ], "rust": { - "abi_id": "i0336", + "abi_id": "i0337", "module": "tma", - "name": "cp_async_bulk_tensor_2d_s2g", + "name": "cp_async_bulk_tensor_3d_s2g", "arguments": [ "*const u8", "*const u8", "i32", + "i32", "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0336", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_2d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile2dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_2d", + "op_type": "CpAsyncBulkTensorS2gTile3dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", "operands": [ "ptr", "ptr", "i32", + "i32", "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", "arguments": [ "shared_ptr", "ptr", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", "ReadOnly", "ReadOnly" @@ -19940,7 +21307,7 @@ } ], "tma": { - "operation": "s2g_tile2d", + "operation": "s2g_tile3d", "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, @@ -19951,7 +21318,7 @@ "async", "bulk", "tensor", - "2d", + "3d", "global", "shared::cta", "tile", @@ -19969,17 +21336,17 @@ "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_3d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_3D", - "asm": "cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2}}], [$mbar];", + "source_record": "TMA_G2S_TILE_CG0_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -19987,11 +21354,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 8, + "argument_index": 9, "value": 0 }, { - "argument_index": 9, + "argument_index": 10, "value": 0 } ] @@ -19999,30 +21366,31 @@ } ], "rust": { - "abi_id": "i0332", + "abi_id": "i0333", "module": "tma", - "name": "cp_async_bulk_tensor_3d_g2s", + "name": "cp_async_bulk_tensor_4d_g2s", "arguments": [ "*mut u8", "*const u8", "i32", "i32", "i32", + "i32", "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_3d", + "op_type": "CpAsyncBulkTensorG2sTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", "operands": [ "ptr", "ptr", @@ -20030,13 +21398,14 @@ "i32", "i32", "i32", + "i32", "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", "arguments": [ "shared_cluster_ptr", "shared_ptr", @@ -20044,6 +21413,7 @@ "i32", "i32", "i32", + "i32", "i16", "i64", "i1", @@ -20053,10 +21423,10 @@ "results": [], "properties": [ "ImmArg", - "ImmArg", + "ImmArg", "ImmArg", "IntrConvergent", - "Range", + "Range", "ReadOnly", "WriteOnly" ], @@ -20222,7 +21592,7 @@ } ], "tma": { - "operation": "g2s_tile3d", + "operation": "g2s_tile4d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -20233,7 +21603,7 @@ "async", "bulk", "tensor", - "3d", + "4d", "shared::cluster", "global", "tile", @@ -20254,25 +21624,25 @@ "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_3d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.3d", + "id": "cp_async_bulk_tensor_4d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_3d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_3D", - "asm": "cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2}}], [$src];", + "source_record": "TMA_G2S_CTA_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 6, + "argument_index": 8, "value": 0 } ] @@ -20280,55 +21650,62 @@ } ], "rust": { - "abi_id": "i0337", + "abi_id": "i1037", "module": "tma", - "name": "cp_async_bulk_tensor_3d_s2g", + "name": "cp_async_bulk_tensor_4d_g2s_cta", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_3d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile3dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_3d", + "op_type": "CpAsyncBulkTensorG2sCtaTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", - "i32" + "i32", + "i32", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d", "arguments": [ + "shared_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", + "i32", "i64", "i1" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -20342,7 +21719,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20359,21 +21736,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20394,11 +21771,11 @@ "targets": [ "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 4D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_4d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,100-byte source had SHA-256 becebd363b47d76e388fccf46646db4d8bba5225b4eb23cc7cae1673a5411ddd; the 2,664-byte bitcode had SHA-256 77f23ed44102d830fb8a576a8ff62f947e96b2b15871e2984083851d1e82f63e; the 1,327-byte canonical IR had SHA-256 6fd6a38abbb214d1e1fc5a4c8aa800ab963d1f53350afc3a7ec965501e467fb4.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20407,13 +21784,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 4D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,567-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 0607dfca39829f02dbda9367602cb27d28e6dc5dd8f5909a451fa4648442629a.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20422,13 +21799,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 4D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -20439,9 +21816,9 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -20463,11 +21840,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -20478,11 +21855,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -20492,8 +21869,8 @@ } ], "tma": { - "operation": "s2g_tile3d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_cta_tile4d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20503,35 +21880,38 @@ "async", "bulk", "tensor", - "3d", - "global", + "4d", "shared::cta", + "global", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { - "id": "cp_async_bulk_tensor_4d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.4d", + "id": "cp_async_bulk_tensor_4d_s2g", + "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_4D", - "asm": "cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3}}], [$mbar];", + "source_record": "TMA_TENSOR_S2G_TILE_4D", + "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20539,11 +21919,7 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 9, - "value": 0 - }, - { - "argument_index": 10, + "argument_index": 7, "value": 0 } ] @@ -20551,69 +21927,58 @@ } ], "rust": { - "abi_id": "i0333", + "abi_id": "i0338", "module": "tma", - "name": "cp_async_bulk_tensor_4d_g2s", + "name": "cp_async_bulk_tensor_4d_s2g", "arguments": [ - "*mut u8", + "*const u8", "*const u8", "i32", "i32", "i32", - "i32", - "*mut u64" + "i32" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_4d", + "op_type": "CpAsyncBulkTensorS2gTile4dOp", + "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", "operands": [ - "ptr", "ptr", "ptr", "i32", "i32", "i32", - "i32", - "i16", - "i64" + "i32" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", "arguments": [ - "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", - "ReadOnly", - "WriteOnly" + "ReadOnly", + "ReadOnly" ], "result_facts": { "no_undef": false, @@ -20777,8 +22142,8 @@ } ], "tma": { - "operation": "g2s_tile4d", - "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", + "operation": "s2g_tile4d", + "adapter": "s2g_pointers_coordinates_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -20789,37 +22154,34 @@ "bulk", "tensor", "4d", - "shared::cluster", "global", + "shared::cta", "tile", - "mbarrier::complete_tx::bytes" + "bulk_group" ], "operands": [ { "kind": "address" }, - { - "kind": "address" - }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from shared to global memory." }, { - "id": "cp_async_bulk_tensor_4d_s2g", - "operation_key": "memory.copy.async.bulk.tensor.s2g.tile.4d", + "id": "cp_async_bulk_tensor_5d_g2s", + "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_s2g_tile_4d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" }, "selections": [ { - "source_record": "TMA_TENSOR_S2G_TILE_4D", - "asm": "cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [$tmap, {{$d0, $d1, $d2, $d3}}], [$src];", + "source_record": "TMA_G2S_TILE_CG0_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ "Subtarget->getPTXVersion() >= 80", "Subtarget->getSmVersion() >= 90" @@ -20827,7 +22189,11 @@ "constraints": { "immediate_bindings": [ { - "argument_index": 7, + "argument_index": 10, + "value": 0 + }, + { + "argument_index": 11, "value": 0 } ] @@ -20835,58 +22201,72 @@ } ], "rust": { - "abi_id": "i0338", + "abi_id": "i0334", "module": "tma", - "name": "cp_async_bulk_tensor_4d_s2g", + "name": "cp_async_bulk_tensor_5d_g2s", "arguments": [ - "*const u8", + "*mut u8", "*const u8", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "*mut u64" ], "result": "()", "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_4d_s2g" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorS2gTile4dOp", - "op_name": "nvvm.cp_async_bulk_tensor_s2g_tile_4d", + "op_type": "CpAsyncBulkTensorG2sTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", "operands": [ + "ptr", "ptr", "ptr", "i32", "i32", "i32", - "i32" + "i32", + "i32", + "i16", + "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", "arguments": [ + "shared_cluster_ptr", "shared_ptr", "ptr", "i32", "i32", "i32", "i32", + "i32", + "i16", "i64", - "i1" + "i1", + "i1", + "i32" ], "results": [], "properties": [ - "ImmArg", + "ImmArg", + "ImmArg", + "ImmArg", "IntrConvergent", - "ReadOnly", - "ReadOnly" + "Range", + "ReadOnly", + "WriteOnly" ], "result_facts": { "no_undef": false, @@ -21050,8 +22430,8 @@ } ], "tma": { - "operation": "s2g_tile4d", - "adapter": "s2g_pointers_coordinates_inject_defaults", + "operation": "g2s_tile5d", + "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, "lowering": "generated_tma", @@ -21061,47 +22441,46 @@ "async", "bulk", "tensor", - "4d", + "5d", + "shared::cluster", "global", - "shared::cta", "tile", - "bulk_group" + "mbarrier::complete_tx::bytes" ], "operands": [ { "kind": "address" }, + { + "kind": "address" + }, { "kind": "address" } ] }, - "summary": "Starts a TMA tile copy from shared to global memory." + "summary": "Starts a TMA tile copy from global to cluster shared memory." }, { - "id": "cp_async_bulk_tensor_5d_g2s", - "operation_key": "memory.copy.async.bulk.tensor.g2s.tile.5d", + "id": "cp_async_bulk_tensor_5d_g2s_cta", + "operation_key": "memory.copy.async.bulk.tensor.g2s.cta.tile.5d", "family": "tma", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_tile_5d" + "source_record": "int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d" }, "selections": [ { - "source_record": "TMA_G2S_TILE_CG0_5D", - "asm": "cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes$cg [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", + "source_record": "TMA_G2S_CTA_TILE_5D", + "asm": "cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [$dst], [$tmap, {{$d0, $d1, $d2, $d3, $d4}}], [$mbar];", "predicates": [ - "Subtarget->getPTXVersion() >= 80", - "Subtarget->getSmVersion() >= 90" + "Subtarget->getSmVersion() >= 90", + "Subtarget->getPTXVersion() >= 86" ], "constraints": { "immediate_bindings": [ { - "argument_index": 10, - "value": 0 - }, - { - "argument_index": 11, + "argument_index": 9, "value": 0 } ] @@ -21109,9 +22488,9 @@ } ], "rust": { - "abi_id": "i0334", + "abi_id": "i1038", "module": "tma", - "name": "cp_async_bulk_tensor_5d_g2s", + "name": "cp_async_bulk_tensor_5d_g2s_cta", "arguments": [ "*mut u8", "*const u8", @@ -21126,15 +22505,15 @@ "safe": false, "must_use": false, "safe_allowlist_reason": null, - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334", - "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038", + "public_path": "cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta", "compatibility_paths": [ - "cuda_device::tma::cp_async_bulk_tensor_5d_g2s" + "cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta" ] }, "dialect": { - "op_type": "CpAsyncBulkTensorG2sTile5dOp", - "op_name": "nvvm.cp_async_bulk_tensor_g2s_tile_5d", + "op_type": "CpAsyncBulkTensorG2sCtaTile5dOp", + "op_name": "nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d", "operands": [ "ptr", "ptr", @@ -21144,15 +22523,14 @@ "i32", "i32", "i32", - "i16", "i64" ], "results": [] }, "llvm": { - "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d", + "symbol": "llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d", "arguments": [ - "shared_cluster_ptr", + "shared_ptr", "shared_ptr", "ptr", "i32", @@ -21160,19 +22538,13 @@ "i32", "i32", "i32", - "i16", "i64", - "i1", - "i1", - "i32" + "i1" ], "results": [], "properties": [ - "ImmArg", - "ImmArg", - "ImmArg", + "ImmArg", "IntrConvergent", - "Range", "ReadOnly", "WriteOnly" ], @@ -21188,7 +22560,7 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21205,21 +22577,21 @@ "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor" }, "backend": { - "profile": "rust-llvm-23.1.3-aab11ee0-tma", + "profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "version": "LLVM version 23.1.3-rust-1.101.0-nightly", "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", "gpu_target": "sm_90", - "ptx_feature": "+ptx80" + "ptx_feature": "+ptx86" }, "backend_lowerings": [ { "backend": "llvm_nvptx", "mechanism": "typed_nvvm", - "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21240,11 +22612,11 @@ "targets": [ "sm_90" ], - "representation": "separate typed LLVM declaration probes", + "representation": "separate typed LLVM declaration probe for the 5D CTA-local TMA G2S operation", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581.", + "detail": "Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_5d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,131-byte source had SHA-256 da8bf49e04c4250626e97150345c704f33802066582d2e80d1c17dd58c0fa0b0; the 2,668-byte bitcode had SHA-256 d1d535b89cc23b4983c8fb5b00562e06266454177d8eb3a81684859ac5430cb5; the 1,358-byte canonical IR had SHA-256 e4862fd8477461d79f65871ae6dc8750d273a0e50911602e0bfea91bae76df7c.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21253,13 +22625,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate probes for the recorded lowering mechanism", + "representation": "typed LLVM call for the 5D CTA-local TMA G2S operation", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 418eaf16785cbdd0e5b8fb5d886e2768f4347b8dda669cd924ad0cd857e7e08a.", + "detail": "The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,701-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 2f87ab7dc1ed68335e4ad29e4bcc7bd264779b33deb26f1d18e0c955f6bf8223.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21268,13 +22640,13 @@ { "targets": [ "sm_90", - "ptx80" + "ptx86" ], - "representation": "separate PTX outputs assembled for the recorded target", + "representation": "LLVM-produced PTX retaining the 5D CTA-local TMA G2S instruction", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 45075184f52be336a8f24e8232251aec1a5890d9a6f0d35cbef2602dca1dafaa.", + "detail": "CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -21285,9 +22657,9 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73-tma-cta", "target": { - "minimum_ptx": "8.0", + "minimum_ptx": "8.6", "hardware": { "kind": "any_of", "alternatives": [ @@ -21309,11 +22681,11 @@ "compute_90", "ptx93" ], - "representation": "14 per-operation NVVM IR 2.0 modules covering the non-CTA-pair inline-PTX forms", + "representation": "one NVVM IR 2.0 module containing the five CTA-local TMA G2S inline-PTX forms", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX.", + "detail": "CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -21324,11 +22696,11 @@ "sm_90", "ptx93" ], - "representation": "14 linked cubins covering the non-CTA-pair TMA instructions", + "representation": "linked cubin retaining five UTMALDG instructions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303.", + "detail": "CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions.", "artifact_kind": "cubin", "tool_path": "libnvJitLink.so.13.3.33", "tool_version": "CUDA 13.3 nvJitLink 13.3.33", @@ -21338,7 +22710,7 @@ } ], "tma": { - "operation": "g2s_tile5d", + "operation": "g2s_cta_tile5d", "adapter": "g2s_pointers_coordinates_barrier_inject_defaults", "runtime_validation": "unexecuted" }, @@ -21350,7 +22722,7 @@ "bulk", "tensor", "5d", - "shared::cluster", + "shared::cta", "global", "tile", "mbarrier::complete_tx::bytes" @@ -21367,7 +22739,7 @@ } ] }, - "summary": "Starts a TMA tile copy from global to cluster shared memory." + "summary": "Starts a TMA tile copy from global to the issuing CTA's shared memory." }, { "id": "cp_async_bulk_tensor_5d_s2g", diff --git a/cuda-oxide/intrinsics/generated-reference.md b/cuda-oxide/intrinsics/generated-reference.md index fbcbe1298e..444c16376b 100644 --- a/cuda-oxide/intrinsics/generated-reference.md +++ b/cuda-oxide/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. --> @@ -97,16 +97,21 @@ This table is generated from the resolved catalog. Evidence stages below disting | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4 [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | `nvvm.cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint` | LLVM `llvm.nvvm.cp.async.bulk.tensor.prefetch.tile.gather4.2d` from `int_nvvm_cp_async_bulk_tensor_prefetch_tile_gather4_2d` | unsafe; scope thread; impure; memory read; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, Tensor-map descriptor and asynchronous bulk tensor operations](https://docs.nvidia.com/cuda/parallel-thread-execution/#tensor-map)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.prefetch.tensor.2d.L2.global.tile::gather4.L2::cache_hint [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.1d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.1d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_1d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.1d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_1d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.1d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster [
], [
], [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast_cg2` | `nvvm.cp_async_bulk_tensor_g2s_tile_2d_multicast_cg2` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_100a or sm_101a or sm_103a or sm_110a ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.shared::cluster.global.tile.mbarrier::complete_tx::bytes.multicast::cluster.cta_group::2 [
], [
], [
], ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_2d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.2d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_2d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.2d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.3d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.3d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.3d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.3d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.4d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.4d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.4d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.4d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` | `nvvm.cp_async_bulk_tensor_g2s_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.5d.shared::cluster.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` | `nvvm.cp_async_bulk_tensor_g2s_cta_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.5d` from `int_nvvm_cp_async_bulk_tensor_g2s_cta_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.6 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-cta`; expects `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes [
], [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` | `nvvm.cp_async_bulk_tensor_s2g_tile_5d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.s2g.tile.5d` from `int_nvvm_cp_async_bulk_tensor_s2g_tile_5d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5 Asynchronous bulk tensor copy](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma`; expects `cp.async.bulk.tensor.5d.global.shared::cta.tile.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_3d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_3d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_3d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-reduce`; expects `cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::tma::cp_async_bulk_tensor_reduce_add_im2col_4d` | `nvvm.cp_async_bulk_tensor_reduce_add_im2col_4d` | LLVM `llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.4d` from `int_nvvm_cp_async_bulk_tensor_reduce_add_im2col_4d` | unsafe; scope thread; impure; memory read_write; convergent true | PTX 8.0 on sm_90+ ([PTX ISA 9.3, 9.7.9.26.5.3 Data Movement and Conversion Instructions: cp.reduce.async.bulk.tensor](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cp-reduce-async-bulk-tensor)) | validated on `rust-llvm-23.1.3-aab11ee0-tma-reduce`; expects `cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.im2col_no_offs.bulk_group [
], [
];` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | @@ -1392,6 +1397,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0328` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s` +- `cp_async_bulk_tensor_1d_g2s_cta` (`i1034`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1034` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_1d_g2s_cta` - `cp_async_bulk_tensor_1d_s2g` (`i0335`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0335` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_1d_s2g` @@ -1400,6 +1409,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0329` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s` +- `cp_async_bulk_tensor_2d_g2s_cta` (`i1035`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1035` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_2d_g2s_cta` - `cp_async_bulk_tensor_2d_g2s_multicast` (`i0330`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0330` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_2d_g2s_multicast` @@ -1416,6 +1429,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0332` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s` +- `cp_async_bulk_tensor_3d_g2s_cta` (`i1036`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1036` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_3d_g2s_cta` - `cp_async_bulk_tensor_3d_s2g` (`i0337`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0337` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_3d_s2g` @@ -1424,6 +1441,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0333` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s` +- `cp_async_bulk_tensor_4d_g2s_cta` (`i1037`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1037` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_4d_g2s_cta` - `cp_async_bulk_tensor_4d_s2g` (`i0338`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0338` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_4d_s2g` @@ -1432,6 +1453,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0334` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s` - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s` +- `cp_async_bulk_tensor_5d_g2s_cta` (`i1038`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1038` + - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_g2s_cta` + - compatibility compiler path: `cuda_device::tma::cp_async_bulk_tensor_5d_g2s_cta` - `cp_async_bulk_tensor_5d_s2g` (`i0339`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0339` - public source path: `cuda_intrinsics::tma::cp_async_bulk_tensor_5d_s2g` @@ -5827,16 +5852,21 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cp_async_bulk_prefetch_tensor_gather4_2d_l2`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_PF_TILE_GATHER4_2D_CH`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_1d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_1d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_1d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_1D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_2d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_2D_MC`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_2d_g2s_multicast_cg2`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_2D_MC`; selection predicates `Subtarget->hasTMABlackwellSupport()` - `cp_async_bulk_tensor_2d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_2D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_3d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_3d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_3D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_4d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_4d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_4D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_g2s`: properties `ImmArg`, `ImmArg`, `ImmArg`, `IntrConvergent`, `Range`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_TILE_CG0_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` +- `cp_async_bulk_tensor_5d_g2s_cta`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `WriteOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_G2S_CTA_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_5d_s2g`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records `TMA_TENSOR_S2G_TILE_5D`; selection predicates `Subtarget->getPTXVersion() >= 80`, `Subtarget->getSmVersion() >= 90` - `cp_async_bulk_tensor_reduce_add_im2col_3d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cp_async_bulk_tensor_reduce_add_im2col_4d`: properties `ImmArg`, `IntrConvergent`, `ReadOnly`, `ReadOnly`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none @@ -7308,6 +7338,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_1d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_1d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,007-byte source had SHA-256 253d3ea39f9d26f5000f83febb9f72a15af2bed6ad0f668ab206976ce4c6d6ba; the 2,640-byte bitcode had SHA-256 9f42944557579c5c9c4e38716cce52607a07c51e5f2272393948fdc40cfb8961; the 1,234-byte canonical IR had SHA-256 835c2efc3b53d5333ec6a8260fcad29d5d0ce889194d0523feb57f7b22281fff. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.1d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,165-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 b0b4469f60fea3342f58e26eb4698834c6cb0714aac17fd3ffaed79cbd4f226a. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_1d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7324,6 +7362,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_2d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_2d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,038-byte source had SHA-256 f0d47c1b6670854dfddce5bc45206d8f4ad16e9a05236dd4e1d4036b518666ea; the 2,648-byte bitcode had SHA-256 6ab53bfd2185eb8b6d7f7ffb3793d72cb7b48a07f61846a5317604700bdd3225; the 1,265-byte canonical IR had SHA-256 2c00613205683c093c5002f0f7347d3b2d6b82e46c5bc00072fb5a781f7d9d44. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.2d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,299-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 59fa9746f48c669cbff59faffea6039a796055f2721b8647a6331d78b5a26521. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_2d_g2s_multicast`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7356,6 +7402,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_3d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_3d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,069-byte source had SHA-256 159ab50347ad4f81c8a110df7de35db35da9eb099938874523e5ec10ea867c8c; the 2,656-byte bitcode had SHA-256 3b5f19c93cffadf279efad33b501519f09cacbc660784f195ab3d007b5bca854; the 1,296-byte canonical IR had SHA-256 223dd701edae4ecf145ca18afc659dd5a50785cf8e492f6ba5c7b7f8299dc030. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.3d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,433-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 e9bab5f8e2c2a697ca435f86eaff468984a03470ae4bf28f975c2c89fc6676b9. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_3d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7372,6 +7426,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_4d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_4d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,100-byte source had SHA-256 becebd363b47d76e388fccf46646db4d8bba5225b4eb23cc7cae1673a5411ddd; the 2,664-byte bitcode had SHA-256 77f23ed44102d830fb8a576a8ff62f947e96b2b15871e2984083851d1e82f63e; the 1,327-byte canonical IR had SHA-256 6fd6a38abbb214d1e1fc5a4c8aa800ab963d1f53350afc3a7ec965501e467fb4. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.4d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,567-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 0607dfca39829f02dbda9367602cb27d28e6dc5dd8f5909a451fa4648442629a. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_4d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. @@ -7388,6 +7450,14 @@ This table is generated from the resolved catalog. Evidence stages below disting - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma` at PTX 8.0 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained every exact instruction with no extern calls. SHA-256 over the sorted per-operation ID/hash lists is 0b5d1f4be62738f2ec3f7c2d58a78bf4a984c225f07190229a422d2befcb71ce for source IR, fbb3c6200d5aab32ef48e87fe38ef428e0d44a8d1c99190ad4291a0a3a4f9797 for LTOIR, and 0f8b2b96f9dc634cb2e8c4ffdebff62f8bcebfe00ae08ba7ab2f975ad5ee382f for linked PTX. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced valid CUDA ELF cubins; SHA-256 over the sorted per-operation ID/cubin-hash list is ea355e5985a293b10266691d71302aad755879854308203a8f9a3f3bd0ce1303. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. +- `cp_async_bulk_tensor_5d_g2s_cta`: runtime `not recorded` + - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_90`: `succeeded` — Candidate probe (`probe --candidate --intrinsic cp_async_bulk_tensor_5d_g2s_cta --gpu-target sm_90 --ptx-feature +ptx86`) with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as canonicalized the intrinsic attributes and the declaration facts matched the imported record. The 1,131-byte source had SHA-256 da8bf49e04c4250626e97150345c704f33802066582d2e80d1c17dd58c0fa0b0; the 2,668-byte bitcode had SHA-256 d1d535b89cc23b4983c8fb5b00562e06266454177d8eb3a81684859ac5430cb5; the 1,358-byte canonical IR had SHA-256 e4862fd8477461d79f65871ae6dc8750d273a0e50911602e0bfea91bae76df7c. + - backend codegen on `sm_90, ptx86`: `succeeded` — The pinned llc (-march=nvptx64 -mcpu=sm_90 -mattr=+ptx86) emitted the exact `cp.async.bulk.tensor.5d.shared::cta.global.tile.mbarrier::complete_tx::bytes` instruction with no extern declarations. The 1,701-byte PTX (`.version 8.6`, `.target sm_90`) had SHA-256 2f87ab7dc1ed68335e4ad29e4bcc7bd264779b33deb26f1d18e0c955f6bf8223. + - PTX assembly on `sm_90, ptx86`: `succeeded` — CUDA 13.3 ptxas (-arch=sm_90) produced a 1,720-byte cubin with SHA-256 9f49927d82f4aeb6e387fbbe19d851b010dd170a6a71ccf4b346d5cb49ecb480. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73-tma-cta` at PTX 8.6 / sm_90+: status `validated` (`CUDA 13.3 libNVVM package 13.3.73; libNVVM API 2.0`, SHA-256 `5317a9fb3c8b719198e89e27fd60e2b59f2702fe6b1eab357167c920b52be9cc`) + - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3.73 libNVVM retained all five exact 1D-5D instructions. Legacy NVVM IR SHA-256 fbde8ef7febd45940a6dbe906a5315dc8e90ad08d25d549bdfafa130e29071e6; LTOIR SHA-256 c08bf2d430ee33d995d17eaf7ab71fb419a9d23ed2db4423adb22230f2cce693; linked PTX SHA-256 26f53d690b7a281bc3cc5eedfb820858f7802f0428c8673d1012ff6fac18de60. + - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced cubin SHA-256 87ec36d9e25d1c954361a691e300867c8b0c0e8f212aa228dd63adf640419cae; nvdisasm retained five TMA load instructions. Tool `libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `f84abd240b428825c3ef16632deb1026b208252fe19cef225a9f4ad06a39d1aa`). Artifact `cubin`. - `cp_async_bulk_tensor_5d_s2g`: runtime `not recorded` - `llvm_nvptx` uses `typed_nvvm` from profile `rust-llvm-23.1.3-aab11ee0-tma` at PTX 8.0 / sm_90+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_90`: `succeeded` — Revalidated 14 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 efc480471cb427bf10e4f248a24454a98102aaef6bf200026684072b2f5f2581. diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index 067d4c64f9..e46f6f4069 100644 --- a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/abs_f16x2.ll b/cuda-oxide/intrinsics/probes/abs_f16x2.ll index 6dae85758d..8d053af58b 100644 --- a/cuda-oxide/intrinsics/probes/abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/active_mask.ll b/cuda-oxide/intrinsics/probes/active_mask.ll index 7fa13f210e..53bb01071b 100644 --- a/cuda-oxide/intrinsics/probes/active_mask.ll +++ b/cuda-oxide/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_bf16x2.ll b/cuda-oxide/intrinsics/probes/add_bf16x2.ll index 9a18789357..929185a050 100644 --- a/cuda-oxide/intrinsics/probes/add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f16x2.ll b/cuda-oxide/intrinsics/probes/add_f16x2.ll index fc61781575..1575fe0486 100644 --- a/cuda-oxide/intrinsics/probes/add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f32x2.ll b/cuda-oxide/intrinsics/probes/add_f32x2.ll index 37d3007c21..8e9261908e 100644 --- a/cuda-oxide/intrinsics/probes/add_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll index 57b115569d..d8fb761418 100644 --- a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_f32.ll index 7052963bf2..fd762bdaaa 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f64.ll b/cuda-oxide/intrinsics/probes/add_rm_f64.ll index 93e78e7647..d74c143bb1 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll index 0f4fb6b035..0ea7710560 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll index 90ee605e4d..570c874254 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll index d69fbd14af..931ce024dd 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_f32.ll index 73ed7cd3b8..638e681cec 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f64.ll b/cuda-oxide/intrinsics/probes/add_rn_f64.ll index 3a6786a3c5..ae039497a2 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll index 9973ab1f59..465c651bb3 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll index 4711626207..f3f552ca9a 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll index a8030adfb4..ff3963dd1c 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_f32.ll index 13a461bd4a..23fd8ddadb 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f64.ll b/cuda-oxide/intrinsics/probes/add_rp_f64.ll index dd49cf5236..59d7d70036 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll index 8cf544e9fc..01a2de0748 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll index 6c24875b2c..3c4177c838 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll index 4fc32f3f88..a488949e2d 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_f32.ll index 6ac13a58b0..cacec2e59d 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f64.ll b/cuda-oxide/intrinsics/probes/add_rz_f64.ll index 2af8a3023f..cdcbdb81fc 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll index 2bcdc125ce..ace477c68e 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll index 355e301756..0bce30539b 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll index 094f9f6deb..c5cbe3c4f4 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/all_sync.ll b/cuda-oxide/intrinsics/probes/all_sync.ll index 29d10663de..7446ffeb96 100644 --- a/cuda-oxide/intrinsics/probes/all_sync.ll +++ b/cuda-oxide/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/any_sync.ll b/cuda-oxide/intrinsics/probes/any_sync.ll index 576cf8dd34..19614abd1f 100644 --- a/cuda-oxide/intrinsics/probes/any_sync.ll +++ b/cuda-oxide/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ballot_sync.ll b/cuda-oxide/intrinsics/probes/ballot_sync.ll index 8befb7faae..c44d1b877d 100644 --- a/cuda-oxide/intrinsics/probes/ballot_sync.ll +++ b/cuda-oxide/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll index cd1c453396..e1eef9d7ae 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll index 5199ada72c..b4c112c163 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index 2c4ba7b688..93b7a5a5b2 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 2665f273a9..5e5aa39471 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll index e0304517aa..8183f455a2 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll index fb0f319061..e0adeef3e0 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll index f527f256a7..6ae0c40b82 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll index a79bf1924d..413749a19f 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll index b53080cc4d..d59eb301fd 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll index 8cfb36aa36..10c6dea369 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_x.ll b/cuda-oxide/intrinsics/probes/block_dim_x.ll index e27224c43b..8c1ca292c2 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_y.ll b/cuda-oxide/intrinsics/probes/block_dim_y.ll index 078e0407fb..e766323f65 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_z.ll b/cuda-oxide/intrinsics/probes/block_dim_z.ll index 328f912a0f..51ac9520e7 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_x.ll b/cuda-oxide/intrinsics/probes/block_idx_x.ll index d3f10ee9b5..ac4a5f8588 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_y.ll b/cuda-oxide/intrinsics/probes/block_idx_y.ll index 393af52146..16f2ef82f4 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_z.ll b/cuda-oxide/intrinsics/probes/block_idx_z.ll index 7e46684f4d..bc5668cc30 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/breakpoint.ll b/cuda-oxide/intrinsics/probes/breakpoint.ll index 06a2c469b1..4e802e2810 100644 --- a/cuda-oxide/intrinsics/probes/breakpoint.ll +++ b/cuda-oxide/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll index f743449a72..9d5559f08a 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll index 03b30c409f..521403f12f 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll index f7e0cdfe75..3f1b745429 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll index 3fb1cef4c0..0ba56f6f22 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll index c4c08feb4d..6a6e971323 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll index 68c4df50e9..17b9884504 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll index 386bf7683c..269ea615a4 100644 --- a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll index 4c04d476de..58eee26aa5 100644 --- a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll index f29a9214a6..814e8c1f12 100644 --- a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll index 0c7b38daa7..0601318628 100644 --- a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_count.ll b/cuda-oxide/intrinsics/probes/cluster_block_count.ll index d244f044ae..4096cec178 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_count.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll index 7488d448ac..d2103b8914 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll index 7376e4d983..1e63183e96 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll index 1a2fc34785..987c3da4f8 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll index 440c88061e..7613a6debb 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll index 7230c4358b..0939650295 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll index 7c4e5810ce..8d4cfdd628 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll index e7c273060f..7ba5086956 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll index 1dd35af076..4a3244946e 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll index 9508406af5..77e47ab58f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll index 910ee9bc3c..ac60470cf3 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll index 8e799974cb..e1703a1e62 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll index a445fe11c2..a61c264498 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll index 64bab3b053..65a453dc55 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll index 17c45671ba..18ed784b24 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll index c00c249aff..988d73e769 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll index aaec000da3..e9978bd657 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index 46e442df90..a547fb53f1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 9f65921817..44ed1e3610 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index 454680f011..3a662ba40f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index 68cde3cd85..22fc3d8080 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 3a129e2a15..89b6304b21 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index 2b171015e8..202d5c91e5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index 739d1e76ae..b142222fa5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index a7443c9ee5..dcedc60938 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index 6f1160d85b..0b1031554b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index bf99c97424..bc32f8c21f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index c1adbd1564..140271b48f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index f499d5ff07..d7f4619d5f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index 5f6cd7d1a6..2448c8e8b1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll index 3bd29e7065..1c2365a73f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 26c100136e..33459bfe66 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index e5287efe79..4bba5ecfa0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll index d7aca9ba6e..d1fa8e2898 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index 8e432bfd63..f8410a1757 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index d4c22e1e0b..299801aa2d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index ff4a9d96c7..7ee36fefbf 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index f4e59a83b7..5d569b93c9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll index 63925d7bdd..a8cedd535d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index e360ec5554..2df0bf6db0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index 734c32c344..57a76256e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll index 4e748a7d03..c80f62fe81 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index 78aa88746a..bd156b2c66 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index 59c4581047..7b8fbc58d5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll index 164b4ecbaa..94386ac5b1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: fa1e911bf859988435fa06a072208a78569ad813901a16c46240fc298171828b. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 08a7b47bf5..f8d5c93c98 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index a328c13b54..f6aa4172cf 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index 47a90c1f91..b6bc087595 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index b440cc9a66..ca99453c1c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 42c8c3406e..bd956a1dda 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index af359d940b..b81c4d98bb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index 4ad7a9bbd9..cde3716885 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index cdf1812920..fd705ce50a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index 69068f4348..6d6ee437e5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index 39b8079ff1..8e23601455 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index a8488e950b..fbb9d1b5b1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index e9db974bc0..002ab772d6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index a1bfb7b666..f2127942d2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index e53fff68a4..80d2959625 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index 8edc1a7962..3b1f925ad9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index 3454107c84..42a2b67acb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index a93faef3d6..97cd772431 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index 81cd03fd9f..6c9c30b6eb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index 36750c05df..584b701882 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 460a9bca6d..ee9d4909b3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index 4e2cc0c1e6..19c3cbae2a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index a4dcb9af51..f2fa71d720 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index 9911f2ff92..c894f8e033 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index 5d64d411f9..01ed1f1b7c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index 636780c003..ab14acb4e1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index fcfd9f8f71..cbbf3727e4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index 851b12833b..dedddbe4e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index 57eba85ca6..4fcb854806 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index d442347227..86e44d8e22 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index 59875f1bd9..2d9d2bb845 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index 9aea0c597c..2ae37f44cb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index d1717d0542..a6a36ac8fe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 686b3a3452..a37d3e9465 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index 8ac11f8adb..eea7f519e8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index 3eba84995e..c91873dc06 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index 82af3065da..42b42a5df0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index 18f95c6b3d..8cc06b0e03 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index e3f194e2e7..f4f16b6d19 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index 5e9dc67265..0b4875b305 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index bab5881b9a..4ac76bbda7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index 7ed9a51e1d..82b9fb6e53 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index b2f6a782cb..b0d724dc15 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index e3ff3c3dbe..fe17ef8bf3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index 95f3385238..e3eab2383a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index a0f4588257..e449dabd0a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index d02c1e355d..4a370d31e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index 83f57a71a7..6a82737b5e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index adab16d057..1d1da94852 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index d5b4f9f0c4..698b703b7c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index feaf9c4b52..f6e79e25cc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index eb6f62f03c..f7ccb958ad 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 2910ba2c1c..1bc4a2c47e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index e07e14ea52..4f4bcc914f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index 87f260e836..76e9cfce05 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index 4fcf98c004..e84327ff93 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index e4ef06fc31..f0ebe02047 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index d803d51e02..3336bb6cbd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index a885575b6f..b6c04ef320 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index a8fdaadc12..7289e3d0e9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 4f20aabc36..49447c9630 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 72f78bf376..9f7da7d68f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index c57638e74b..47112c5c5e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index 0b3816e9da..c9796b53b2 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 0313e9fcbd..d44ec05c1f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index 572c3abda0..0684800498 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll index fcb98f7e95..052f0e5f88 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll index ddd4e30212..30cf4e78ea 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll index 92065d24e1..c2a13457f3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll index 152c11dce3..8be6e32b18 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll index 0e6b58d2b0..cea3ee60fc 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll index 91eed5c738..6ced4b6397 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll index 0b45e2547d..164af51c2a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll index aa2bfb45af..5ed88a2841 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll index 6b81d8bd58..da742891a5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll index c6fb788be1..798ebb5e6a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll index 9ae3549b39..fe7d15e012 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll index 4255a91037..26df93f72e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index dcbc9a3db5..64838e7da1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 41db650465..0af7393ba9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index 66ce94af51..a666f18201 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll index 4429f3f902..afff377824 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll index ed8825d63d..fad0504d41 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll index e87cebd799..0493a146dd 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll index 64c4b60d4e..b95ff64c56 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 7dbe4d58c6..198b91f590 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index 180687268b..597d5e8639 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index 9aaebadeae..358d0d1974 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 8a82e4f569..83a34555f0 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index 6109e6e7b1..d1abfdbeb3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index ca58ea76c8..fb4e380e08 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index e5e9348b81..19500ce50b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index e6a5e34364..a4a66fa8db 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll index 081544c233..0bb892f6d2 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index 5e934210d7..df07680c30 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index 5891836047..3f3eebf39f 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll index 0bb0c24789..dff829ede3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index f440693c28..ad47108f81 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index 1e7961cb68..d2b2ab9dfc 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll index 0b467a2c90..8913f5b5cf 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index a286c6d1bf..198f3643af 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index 7ab9c6003e..aea882afb1 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll index 299f50b531..1699d7f099 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index 84aa86de91..87af035c32 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index cc500b7317..5474c487df 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index 1713149464..346a8f82bc 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll index f476333136..d5949622e8 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index 3229ba9b78..4998500d2e 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll index db51c2bab7..72b07524e3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll index b76bdde10a..b07b8ae0cf 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll index fe7d3d6937..d1b8705209 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index 82ffd726e7..562d04368c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 74f5fc2ba1..1c9506aee2 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 63eef2db4c..caf8797096 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll index 419b47bfc3..dcde35438b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_f32.ll index a1d3da6717..7a42bbe5dc 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f64.ll b/cuda-oxide/intrinsics/probes/div_rm_f64.ll index 00c593b7cf..29061f0f75 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll index 0f51f288c5..c7b2c9189c 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_f32.ll index da743c6ea0..c80996935d 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f64.ll b/cuda-oxide/intrinsics/probes/div_rn_f64.ll index 47f2055c65..fb1fb654ad 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll index 016ec9e93a..29f7cb6506 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_f32.ll index 206e3ad108..7d0f0f9bfa 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f64.ll b/cuda-oxide/intrinsics/probes/div_rp_f64.ll index 6d5d138bff..ed58974abc 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll index 3433f00e97..22547b5bd4 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_f32.ll index 278d9fb167..630217b846 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f64.ll b/cuda-oxide/intrinsics/probes/div_rz_f64.ll index 3369341aec..642e6bda86 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll index 447898907d..2d2f5e51ae 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_s32.ll b/cuda-oxide/intrinsics/probes/dp2a_s32.ll index 33b9322ca8..d9261cfed7 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_u32.ll b/cuda-oxide/intrinsics/probes/dp2a_u32.ll index 5c4f90aac8..ce6c83d588 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_s32.ll b/cuda-oxide/intrinsics/probes/dp4a_s32.ll index 8030171ce9..2d41f630a3 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_u32.ll b/cuda-oxide/intrinsics/probes/dp4a_u32.ll index 7259fd2a04..2572e62911 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll index fcd64174aa..233350c6c2 100644 --- a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll +++ b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index c0947ece4e..965f3f0e2d 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index b5b176f529..a92c49207b 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll index b1113d8083..d2057ada53 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll index 27907c641f..a3081473d5 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll index 5a813a2d8c..2c420d59bf 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll index 70ecb71ba2..e405529d2f 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll index 3f3050d23f..2b118653fe 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll index e61b17ab3e..3149169d4a 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll index a19c13f319..3819cc7d10 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll index 56f292188f..d0191b77ca 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll index 64c080d159..5c3f9f3efe 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index 21aa09cb06..ac6874666f 100644 --- a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index 0ac2068b54..307b76d16d 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index 5c2acf1398..1261c0f261 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll index 49531a6d30..a2ade6ecde 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index e1422568b7..ab1ee0448f 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index 336e78599e..f68edcdb41 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index 2403be316d..49836c3e14 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index 49812021bc..5400396c20 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index 85452598d9..65531318e2 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index d18301c77e..d018df00b3 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index 8c93d75714..6c2ea273ea 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index 5e17dedb6a..458bd1e5ca 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll index 7ade7c9298..f2dccf646e 100644 --- a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_f16x2.ll index f98be54f13..a597f96147 100644 --- a/cuda-oxide/intrinsics/probes/fma_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_f32x2.ll index dd332b10eb..ae86d4e6d5 100644 --- a/cuda-oxide/intrinsics/probes/fma_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll index 7053c40911..9ca84e44a1 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll index 67522a4e4d..f2e3e561fd 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll index 8e3c9eb719..258330ba09 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll index f0aa509ed2..df131526f7 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll index ed3b27aec5..7488261367 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll index 1c97d0f674..5584d04ed2 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll index 92c934285e..4af2cc6183 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll index 343fa5d580..a5f113a6f7 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll index a6ad65b13b..507e315a2a 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll index 6d25197ac5..fb7eb628f8 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll index 075296a965..43aea59ead 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll index 315fa265e9..53840e35b7 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll index 9da6f952d3..6712e072e2 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll index aff9f07ff7..cb48e916e9 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll index 5c471fec57..157102fd1b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll index c8c4f68e47..7cee895c38 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll index 399094b7fe..b75d305559 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll index 4213b63aad..15f6beff4d 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll index 98cea80bd4..925796f88e 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll index ae1ee61678..834e3a7f0e 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll index 7ca0261283..4f2bacb797 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll index 365dea58a5..ab50a01c53 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll index 0e5ec433d2..559ac118f6 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll index 3188b45136..2ee08e700f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 3a0c7b627a..ff741ba691 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll index 9ca05876fc..46e5127db3 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll index 78b9c47da7..a4df1fd849 100644 --- a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll index 7d351266ff..562de52630 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll index c204923f75..cd48285911 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll index 5b0764b9c7..55f6b5b501 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll index 6c1202bfa0..2cbcfa0eee 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_x.ll b/cuda-oxide/intrinsics/probes/grid_dim_x.ll index 07aa2ea0b1..4821d0c54d 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_y.ll b/cuda-oxide/intrinsics/probes/grid_dim_y.ll index ad96d8b8df..47e3eaded5 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_z.ll b/cuda-oxide/intrinsics/probes/grid_dim_z.ll index c28d2b2cc8..572a3a5e04 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll index 4d46855c63..9983029740 100644 --- a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll index 4fdc43eb18..4b16a33b74 100644 --- a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lane_id.ll b/cuda-oxide/intrinsics/probes/lane_id.ll index 50f6c834cc..bb9bd2802c 100644 --- a/cuda-oxide/intrinsics/probes/lane_id.ll +++ b/cuda-oxide/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_eq.ll b/cuda-oxide/intrinsics/probes/lanemask_eq.ll index 61843c33c7..f9b99d6291 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_eq.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_ge.ll b/cuda-oxide/intrinsics/probes/lanemask_ge.ll index 1acedaa624..2a3fdf9bed 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_ge.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_gt.ll b/cuda-oxide/intrinsics/probes/lanemask_gt.ll index 1295da46e3..94c7501f6c 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_gt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_le.ll b/cuda-oxide/intrinsics/probes/lanemask_le.ll index e1c82790f0..3dd6c78734 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_le.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_lt.ll b/cuda-oxide/intrinsics/probes/lanemask_lt.ll index bb04d23139..01d1c2b18e 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_lt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index d6c0fc68ae..4594d1b137 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index ddc3b57aff..843d96fe53 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 3c9723fc8b..330c2ee70b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index 9c2ee7fca6..56fceee0b8 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index e2fa0648b2..984abf265c 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index c27457afd4..4a9c597af0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index 56bda53886..639ee3c0d5 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index 027688ac4b..34972cac93 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index 12ffeeed29..f386d7cc44 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 809f7982c5..fbf88e2d4d 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index 6de8a3e416..5a016ff865 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 894afe32fc..bccbf60148 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 453adbec66..99c994fa35 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 53cf01df1c..f16610cd19 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index f96143f6eb..686baf70a5 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index 8d7a76b4a3..f0ca18e74b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index dc7eb370f2..4a56256f46 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index d677b6c375..648ea122c4 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll index e47328e66d..a1653bc39e 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll index fb08af5e0b..cb0cbc19f0 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/map_shared_rank.ll b/cuda-oxide/intrinsics/probes/map_shared_rank.ll index 5358a9d506..9b04801881 100644 --- a/cuda-oxide/intrinsics/probes/map_shared_rank.ll +++ b/cuda-oxide/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll index fdd5192dcc..b18d94d216 100644 --- a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_sync.ll b/cuda-oxide/intrinsics/probes/match_all_sync.ll index 868dd4db47..74b22b2ed2 100644 --- a/cuda-oxide/intrinsics/probes/match_all_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll index bd6701cd5f..0cadc0b658 100644 --- a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_sync.ll b/cuda-oxide/intrinsics/probes/match_any_sync.ll index d130a1e5c2..9109f2ce04 100644 --- a/cuda-oxide/intrinsics/probes/match_any_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16.ll b/cuda-oxide/intrinsics/probes/max_bf16.ll index 4c4301746f..c19d49b1e8 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_bf16x2.ll index 4d54c4f928..2e4bc4e8af 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16.ll b/cuda-oxide/intrinsics/probes/max_f16.ll index 0bc3bc5009..3644e944cc 100644 --- a/cuda-oxide/intrinsics/probes/max_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16x2.ll b/cuda-oxide/intrinsics/probes/max_f16x2.ll index 1ee7701c9b..70deb44935 100644 --- a/cuda-oxide/intrinsics/probes/max_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll index db80c0b0df..9d6d1c2794 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll index 9108717041..9adcdb6146 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll index 1a2b90b3bf..04cbeab5aa 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll index 009bc5b38d..18d0996c47 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index b0d5d6da4b..8a5db4b169 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index b082a62ef8..ff20cce14a 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 2725f06bbb..effd7b96fc 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 97a59e57eb..08757fde1f 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 6f366cd931..75480a7586 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index 8775a69610..0a9d03bab0 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll index 915471f870..3c729d5986 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll index 3aa7dab90f..ba1284fa43 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_f16.ll index 8bf3acc56f..2c28b85397 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll index 408b2a979d..16ae094785 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index f32ff88d10..13b106899f 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index f1c1aba39f..4a27cdd089 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 74a5ac7431..0c293df2b4 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index 776b6820e3..787c605721 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 7d64f9b05d..a83d7b1623 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll index 26a3d223c3..e64f1fe375 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s32.ll b/cuda-oxide/intrinsics/probes/max_relu_s32.ll index d94f6d2429..479a66e294 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_s16x2.ll b/cuda-oxide/intrinsics/probes/max_s16x2.ll index 5a7722e611..d23da75a83 100644 --- a/cuda-oxide/intrinsics/probes/max_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_u16x2.ll b/cuda-oxide/intrinsics/probes/max_u16x2.ll index c5d4473cdc..6c635e6347 100644 --- a/cuda-oxide/intrinsics/probes/max_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll index f1ce7e9a88..45ed4e48ef 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll index e2a4004747..dbbda73403 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll index bf91eb2045..1c99646f41 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll index 8f9aae2372..2c8610d387 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll index 6389bee8a4..e2b2fee038 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll index a56003abd7..7977c6a44c 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll index 51de19fe1a..6aa1129259 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll index 736bb73f9a..2cc015f97b 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index cf1247a804..419f1f94f6 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll index 6210d63811..8f1a6f7d75 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_init.ll b/cuda-oxide/intrinsics/probes/mbarrier_init.ll index 047b17588c..ddea1c84ef 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_init.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll index baae5bf776..960026a0a0 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll index 0bbf799963..8090d9d56e 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll index 31af9e74bf..3e52b13de0 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll index b4896de30d..9c06fede43 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 6c9bdaeabe..c3b4a9fe72 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16.ll b/cuda-oxide/intrinsics/probes/min_bf16.ll index 14ed965fc9..b3fd7c4acf 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_bf16x2.ll index 2c6070ac9e..620b7b9563 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16.ll b/cuda-oxide/intrinsics/probes/min_f16.ll index 404f6960cb..2ad6d6bf92 100644 --- a/cuda-oxide/intrinsics/probes/min_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16x2.ll b/cuda-oxide/intrinsics/probes/min_f16x2.ll index f876d127d2..e6e359ff22 100644 --- a/cuda-oxide/intrinsics/probes/min_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll index bfc0cd2f62..1febefaa3a 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll index 85d02ee233..19b8e91144 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll index 50132f3274..0b18d5af19 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll index 3481f3f285..32d672fce3 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index 06b403ec3f..84a9df17eb 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index 7b1ebde1f8..11652e1a82 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index dbd6dd62f5..f3e3f32310 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index 958a82600c..25075ee272 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index 7c2e8409fe..37ba2174df 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index 2e1827fadd..39b58cfedc 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll index c110bc3aa7..c2fea41689 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll index c3a58dec08..40e598ecce 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_f16.ll index f8192f0932..3069deaebf 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll index f6fac96c8d..f47c8d0503 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 75ff0178de..e1e7a35ad4 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index b9b99edca6..ff72d766b1 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 57cf031d17..124cba8c70 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index 8d7e746214..a0b989fc63 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll index a798962b2c..d7af7d6af3 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll index c27c250591..de329394a2 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s32.ll b/cuda-oxide/intrinsics/probes/min_relu_s32.ll index cc7b6317dd..7b41a21a71 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_s16x2.ll b/cuda-oxide/intrinsics/probes/min_s16x2.ll index cd633cc6d1..fcf15664e8 100644 --- a/cuda-oxide/intrinsics/probes/min_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_u16x2.ll b/cuda-oxide/intrinsics/probes/min_u16x2.ll index f41b0c3327..8401b6d74a 100644 --- a/cuda-oxide/intrinsics/probes/min_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll index f7bde0891d..af95cb911b 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 4166605972..71f72dd737 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll index 64bf809f51..44959208e6 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll index b2c308038b..b45e2130f7 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll index f8510d0693..cc3f671f6b 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index 5fddf2f973..fa13725786 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index ef355d4797..7215251120 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll index 0926220c49..1f191d26f5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index bf74d189b9..0ec7209500 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 77fda80213..57e293c530 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index dfcd67fa9c..325cec17e9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 7cfbe1497d..ad7db34760 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 194677aa5a..b052cfc684 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index fb49b36c7a..0d17ea05bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index 6535323bb5..f0934d6591 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index f32e24c495..083649872c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index cebb70f36f..44447744b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index b5d157fdc7..a4db5c8b37 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 60d0e78603..99cb12537e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 6546046d16..d955d29da9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 383bd2770f..18b20d47f1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index 4d1a732b1f..556ce9334d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll index e5dc531dd1..9e6026e72c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index ba122b4e6c..f63c7a54da 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index 03bee05bdd..079f93f281 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index dc1b9bbf05..846de70044 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index b5201afde4..f15247556b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 476a4038dd..896a734197 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index 16f22fd9d9..278f9db388 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index 694133d6fe..fae8a3a2b2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 1a3a7b7054..ce07d8c25b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index df6658f4bb..547d0666e2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index abcedf9ca2..cce6fbeb8b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index 7e993480ed..027f635672 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index ba8a3d2399..d32c8a5ad1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index d2f6566bc5..118181a3a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index cb58731491..3bdcd6248a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index e46dc15509..3a760376de 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index d15a414e46..617fe96f0a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 9e9620c221..3240530dca 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index c5d2cf8119..4abdd46827 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index 65eaf27db0..69ee1937a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index 93e80f881a..691133df83 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 0c38aec935..afc4c3e7e3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index b1b3a4b4de..b4838e6e0c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index 9d04dc8c93..a3880915be 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 56313495d6..c52534c958 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index 204d1b2f75..21c9ec4ed7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index e48d800b41..062cf13352 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index cc3b2056cc..d94339408d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index e9a53d8649..8bafc78b67 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index bbba2ff0be..69751c224a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index ae1653eb58..0ff2b9137e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index dba4a512b8..a021186e6b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index a15fa7dbe7..3025a683a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index 0d75ea5930..25b03cdeda 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index 092f54ee29..c0cefaba81 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index a8af62e486..a5e6f953db 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index d00fa98966..37d1f536ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index e3a1580960..2b64a954e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index cbac0e3773..e063b4b4d8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index 1ca4d09147..246ce88133 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index e86249b588..f3f51e6e80 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index bc87ee24c7..a8c4296221 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index 091a7c67b9..b935ff1e63 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index 95b5342200..b56b2123d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index 4785e99b81..7a67c7b137 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index 42547cf2da..759a098236 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index 3d165cbae8..180157808d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 811dcce5d2..3721de21b5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index ff6d6cd68b..61482b9444 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index 27d37e69cf..4e18ef9af6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index 63b97ffcd0..99869663c7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index 508a2510cd..c45dbeaabd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index 8b913fd68c..e304d8bf2f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index 3e337b5e1c..498fec6b08 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index 56e5157434..773c74e1c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index 352fd5d01b..22e1d45cfc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index 0e07abce1e..68b769eea4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index c444bd307b..51fac63976 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 985769e694..6d33a1b21d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 21487f3252..f0b61fa737 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index 8ef69b5e08..d1adfc3fca 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index 99351c9bfa..b6e48f4640 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index c71d6021ec..da1b2375cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index 95abd6f977..4ebd37f67e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index 71a61bd426..1313d82205 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index 984b309228..9a41b862b2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index fc6f40f9c3..9c3351e8da 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index b89004ed20..8b571f3b4c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index d8cccb3c87..db991e0120 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index c5b129b074..912c4dace6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 9f26f91a45..c96c256e65 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 9c8ed180b7..734a4fc3e4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index e12f17e2cb..7b40096d12 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index 7b90b57cdd..7a67ef3d0f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index 42350b5441..1bd1e836fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index 4fd25d377e..31210f64cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index 2bd447cdc4..14b473d0fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index 0a7db968b8..da7b231193 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 36a4777c65..055104fb98 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index 662d835a9a..e858844729 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index bab546777b..327d254b18 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index fd7ff404de..1bff57a99c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 2d7af4dc62..9da54c118e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index eff1df7d5f..a87b8c09b4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index f8b0ca4421..9eb36193c6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 075addc2fd..49c707ef92 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index 74cb16cfba..b162131161 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index d4288b006c..f526670640 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index d7a93b5131..9bb17c4ff9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll index 2b05e6ac46..33d0d1ae49 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index a6d9df46fd..398c97425e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index 88978c7925..6623b80761 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 7347863db4..d6442d8e79 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll index 1b82a45cb3..36ba3bd354 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index 7f456cee1d..d84f39e5d4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index c48bc4588d..19c70460e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index 86d578415d..09b1e760fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll index 100d8d02fa..3b9a7607a3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index 8fb5e99e1e..ef5056fbb0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 363c34aa54..c1d3610575 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index 0a4700d71e..17ac021728 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll index 1cc31f091e..cdf865ea3f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index a8f3e8edcd..9f08c150b3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index b75a96d9ce..9ea53896ea 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index e8cddafd20..139f6f4e5e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index df0a653fb8..954476490c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll index 9910653161..125f74e0f7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index cc6ad92c93..9ebe5599f5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index bd201010c3..0edb124ada 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index fc5351eaa1..c9893ef9a1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll index 2cd84c241a..d09c984516 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index e9d3682d82..193e5eb2f7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index 45e334cca0..08299ee171 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index 94995327ab..51f1745692 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll index 4ddf6656ed..607ea2d3e8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index ca1249742e..4470b80636 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 31875b3d5e..7b083623dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index ff8cf0fd4b..8249863e9c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index c2c2b92355..aacee5dd95 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 1703de2ea9..3be88b3fb0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll index 56df23b767..729c2519b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index 3f2d45dfa1..e104fd966e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 88b613308d..17f0a55591 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 35397f1026..5266e35eeb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll index f64e01c530..93f5345f58 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index 393c6a46e3..c04b818f81 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index a62d22771b..8b536f8c09 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index 4e084fdc64..fa46c03dd2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll index 5779d1fa19..9ee6e89f9b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index 9ef79c8b34..667ebdfe6a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index c5d0fa2683..7e3b35fd00 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 8f121eec83..8f9ff8d844 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll index e5a8b93a30..07888edb90 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 75221e54a1..aa2bed3bb6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index 021fb95090..d7b5fbcbe1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index f0b419a655..82832000df 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll index 3e3e7712c7..1299aa0a7d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index 8077529cd6..7440bc82ee 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index aaf3101ba5..0a5fec3761 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index cbf2e9dcb5..ae1dafa6a8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index 6201f118e5..329468af5f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 62a20ced1e..55039ec6bb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index 24f0d33c71..2cb5091201 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index 5f39597b9d..cfe30cad88 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index df71dc6d0d..6d58ec174d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index 481a6cdfe5..f24fabd66d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 2dc97ff959..e952ce44cb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index 37e2562d47..a60f5cb653 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index 476882e1bb..4d64a83415 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index 8e7f547103..6d0e34be7f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index 95c75878ba..e4a784e0de 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index eeb6598400..3c23f9485d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index 4b32e8fc81..5c645518a1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 7207939f4f..b24157208c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index ad54063aaf..01a91d874d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index c126f91510..0fa2e8450f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index c9c1918d9f..2f8d3c4f5f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 619e06bc16..be2cc7c91d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index 37e8a6f059..3c04a35ccb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index c85c9292ad..5804d21c00 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index 943cfced44..12c8dd5c57 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index 64891baec8..e34ca13f97 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index ef1f8cb45c..17bd289fdd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index 9350335181..8368504320 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index 2af2dbbea4..d3755de6f2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index 0be4489728..50ffdace1b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index 50df175b70..28429bb564 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index 1f5e2846e2..8cfa70d25f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index c82d70105a..f1446edd8d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index e643082510..6f18f3517a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 4938e384b8..cd6fc278cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 1f7d24771d..051c6d4c74 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 11616cf36e..4e1efacab3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index 7f2fab77e0..800335d2a0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index 61c63143b7..c390c28b17 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index 8379d7c7c3..9f3db5d2b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index a0f4f44d79..6cb31a9de2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 536ceec7d3..7863299241 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 0fd960de52..6c1e92fd47 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index f4d0c1e716..6368aed6a7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index ac08ff4cbb..1d35a1e47e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index d1889712da..a731331f23 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index 82b0cbfcab..2f3082d7ba 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index 6d183e3edf..211dbd4ebd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index 64459b199d..7788ead153 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index cd041ca53c..4b4d733d85 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index eaa8c5c985..77f306baa1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index e38d992dc3..30182c35be 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index c5274b850f..ee5a016914 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index 4b76de0392..f4d49a28b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index c68144b194..fe0ba54f6b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index 6b07fa97b3..6efea8eedc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index 53b8ebb68c..46cb773cdd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index 4625dd5167..a7b2e75a4b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index fbd0f98935..bbe7abaf57 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index d0f9940106..787d860587 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 18c39df4d1..98a81aac1b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index badd9009b2..a2a3a67535 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 2954787465..6124bc0e88 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index e4f5f34ecb..01d7dc6239 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index d7cdfc3325..11c5b64541 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 49ab833b46..238b2faf02 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 5f5e552bf4..dac4744a82 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 61216b1c23..7361c5ab45 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index 37e2d843d2..95d4284f8b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index e355988fac..224116045e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index a48ae84609..af0bd28c7c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index 365d99861b..4e6bd73ae9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 0c10aa3c53..2e0db1ea69 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 7d2bf0ecda..7397268a7a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index 6d20bc3001..c9f7763c5d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index 04bcd8194a..e8983ac1db 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index dbc72511a9..440d5f6a0e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index 9bf94d05be..98939aca6b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 744ede663e..119503d4d3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index d169bcc771..6216ee7011 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index 896c6513a4..e7a9b99f26 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index 9921ae1078..509d491dd9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 3ad8555152..e1eaf9e218 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index 584298de52..84819143e5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index ed9e47ebd5..7eee7b9def 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index 4b94ed7fa2..f4f4923889 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index cb687c945e..250f13256b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index fa7b4aae5f..c7c9ca16cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index f925ccf290..e8d06477d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 338a035645..988f1d9803 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 523615ae7e..06a07052e8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 741ae6b002..1ca4142725 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 8ccf28a967..ccb779f515 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 5c195b2493..c09c411b55 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index 7eb7c201a5..4af5008099 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index 07e4ea7068..a6f9c983c5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index f0b4414207..366f21f1ae 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index 203ff011b6..07aad7bd29 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index 6eeeae590f..026fd87d75 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 40679ce7d1..3696acfe13 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index 7793fb1b6f..7dd708b563 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index 186b31988f..dde7c65b02 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index 4ecfffa2d3..9c495a3a1a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 00301b238d..ed4c140925 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index 4584a6b8d6..eb91c42287 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index 36f7dab63f..ebc6f84744 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index 5782ced891..a4c3b7880f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index 47e26c23ff..16beafec70 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index 3be3033ccb..6415441c0d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 7fb9a35414..252ed15628 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index b2ba7d3855..20483e4765 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index b06d99d190..e2cefbdfe7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index 02c0aa5154..d16920a061 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 9c406a2cac..355906beaf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 64cf990e67..c44eab75bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index 9c1a0875a9..14ee6b4e0c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 22b414bd57..0e961fc73b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 063d770895..b85e8a31ec 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index d41ee4bd57..ed3fcdde21 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index ff58561d3b..44b49ffda0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index ea82f4255a..895d4b076a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index 9a5019678d..31ea211631 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index ad871e7a4c..9e2876ddab 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index 1db5d5afc0..4e9861522e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index 011ce74c43..7c47e26576 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index 50ed30e6ee..c3af92fbeb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index 292f2dced7..7fa8cd9d1e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll index ac18f46e13..2d77d30f51 100644 --- a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll index b02cc4168b..94aa444917 100644 --- a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f16x2.ll b/cuda-oxide/intrinsics/probes/mul_f16x2.ll index f29398d954..682f3ac269 100644 --- a/cuda-oxide/intrinsics/probes/mul_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_f32x2.ll index 0e826e31cc..6686899b43 100644 --- a/cuda-oxide/intrinsics/probes/mul_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll index 1dd4cb4bdb..647cf62d7f 100644 --- a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll index 946d9ec9d8..8f7c5b6b24 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll index 38869a4d3c..705f6c6c75 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll index 930ba15bea..d67454d0a7 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll index 8845444e20..afa88e6fdb 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll index 9800e872ec..95dbef9aea 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll index 1d01f663e5..ac3828ba69 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll index 6b5594af01..d46918c1c8 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll index 68da7fde6f..7f4d1da32e 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll index 50d581c61a..7c2f165e12 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll index 75dd4ff99b..87c5a61177 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll index 2bfa2652fd..a12e557579 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll index 5c3c0abd74..3caaa9fa84 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nanosleep.ll b/cuda-oxide/intrinsics/probes/nanosleep.ll index 3d173156cc..d7e614d796 100644 --- a/cuda-oxide/intrinsics/probes/nanosleep.ll +++ b/cuda-oxide/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll index e9488c24bf..4ecbe2ba72 100644 --- a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_f16x2.ll b/cuda-oxide/intrinsics/probes/neg_f16x2.ll index 385e6df41c..7b748403e5 100644 --- a/cuda-oxide/intrinsics/probes/neg_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll index f9a0421868..99a2954481 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll index fdda0fb0d5..ecf9bcfc1f 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll index 37efd6dc21..00e93ab2a0 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll index 83519dc2c0..2ec773653b 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll index 4e227d2738..ed0dea8e2a 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll index b31c4bb7fe..c6589d0124 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/pmevent.ll b/cuda-oxide/intrinsics/probes/pmevent.ll index 0fdf57f45f..cee3b97421 100644 --- a/cuda-oxide/intrinsics/probes/pmevent.ll +++ b/cuda-oxide/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll index f6ea49f51a..69e33453c6 100644 --- a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt.ll b/cuda-oxide/intrinsics/probes/prmt.ll index 25e4ee79fd..54e589832f 100644 --- a/cuda-oxide/intrinsics/probes/prmt.ll +++ b/cuda-oxide/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_b4e.ll b/cuda-oxide/intrinsics/probes/prmt_b4e.ll index 484a1dfdc4..6b3114959e 100644 --- a/cuda-oxide/intrinsics/probes/prmt_b4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecl.ll b/cuda-oxide/intrinsics/probes/prmt_ecl.ll index 34e0535dd1..6455efc4b7 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecl.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecr.ll b/cuda-oxide/intrinsics/probes/prmt_ecr.ll index f83f4a1c08..4128351cd3 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecr.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_f4e.ll b/cuda-oxide/intrinsics/probes/prmt_f4e.ll index c4a7614224..aae6545972 100644 --- a/cuda-oxide/intrinsics/probes/prmt_f4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc16.ll b/cuda-oxide/intrinsics/probes/prmt_rc16.ll index 14192b0ebf..7c2a198257 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc16.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc8.ll b/cuda-oxide/intrinsics/probes/prmt_rc8.ll index 009328e41e..a862ad0456 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc8.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll index c8ba34faf5..04ce29599f 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll index e742043f1a..e7fb891f29 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll index 22099fbb9d..7df0f7d120 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll index 9350e2b746..7c8b7d824a 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll index 3662480176..63488a241e 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll index 9b321528a5..0a37391c9a 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll index 4474b209b6..ce085ea4b4 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll index 0706a5dd87..333cd323ad 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll index ed85e8c9de..48ad43f706 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll index 67e94691f7..0a5a25e5bf 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll index c55efd33b8..1314397bbb 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll index 8be6130b77..3834595528 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll index 78a380cd8d..2212561a32 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll index 01d8028fd2..08d83dfa22 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_add.ll b/cuda-oxide/intrinsics/probes/redux_sync_add.ll index 3c6fef28a4..cd4bf72700 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_add.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_and.ll b/cuda-oxide/intrinsics/probes/redux_sync_and.ll index d62cd46fb1..9664a17a4d 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_and.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll index 1977da8b38..f80cfa51c0 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 3da772f242..6773f431c9 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll index cd25209c11..ba1728e578 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll index f25556e96f..ec78783c2a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll index cac9c75c91..82b3688003 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll index 6b3c0ab201..a4a13febf5 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll index 8fea47dc6f..5f7dc786e9 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 6990a7d483..039a923201 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll index 6de0160a51..28a7cad706 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll index 5ef98bce2b..3c06926ae8 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll index e375542b35..c98ffa085a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll index 8b23ffc72f..9362796a08 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_or.ll b/cuda-oxide/intrinsics/probes/redux_sync_or.ll index 5def8ee0a3..09ecebd526 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_or.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll index ba934006d2..06c05198ba 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll index c3e17919e7..8c6179c4aa 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll index 379c790503..312f889f5e 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll index 345905cacb..cfdf0a2b24 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll index 3db9583452..aa084717bd 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll index 5b81fdbd03..915ea24445 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll index 07931bc541..88f014a29f 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll index 565a566ecf..92b62cccdf 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll index 9101ef9b0d..6da995cc9c 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll index d08405a0c0..75c48538fa 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_sync.ll index c095166597..5c6fb324a7 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll index 1598ac5170..20d248540f 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll index 968d5ae090..294c7862a6 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll index 9eab4166a3..1a65683779 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll index c3bdf2b0ea..0d5545634b 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll index 8306292e9f..2fb485631d 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll index f62c3be3fb..41cef2d48c 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll index 34180ef988..5d5a590cf3 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll index bac49742c0..77d2ba6110 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll index 4577db39c5..ff161742e8 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll index 1a59d8018a..1a9c8758f3 100644 --- a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll index 286922323f..50cce104ae 100644 --- a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll index 65b0fbe72c..899e546d13 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll index cd8eda50d4..262b98ba46 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll index 2f866fb9ae..d8928ed631 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll index dd08cfa8d1..da1f9136ec 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll index 9a52fb6c17..bfea216d84 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll index 6cb73537f2..73b1e6071c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll index 7f30e096c5..50c4b939f8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll index 111aff77cd..c82e435467 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll index 7bf5b85fe8..40f1721b68 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll index fcd654a822..efc26e9944 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll index e0064049b8..973d8f88c6 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll index 161e597e63..db08de80b6 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll index 1ac36b7dcd..1aba99e65c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll index ed9a2b35cf..dba46e92b2 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 23cecab80c..16922e61a4 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index ff62f080e7..9a406fbf8e 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 7e927c768d..208f8e3ba4 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 8dfad7bbda..1ea151eefd 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll index fb4c2e2fe2..7975adacf5 100644 --- a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f16x2.ll b/cuda-oxide/intrinsics/probes/sub_f16x2.ll index f7ce7ff1c4..9b500df35e 100644 --- a/cuda-oxide/intrinsics/probes/sub_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_f32x2.ll index 182ae4c0bd..a4e3cb2dd2 100644 --- a/cuda-oxide/intrinsics/probes/sub_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll index 0facbbe458..ba2a552a28 100644 --- a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_mask.ll b/cuda-oxide/intrinsics/probes/sync_mask.ll index f6a9812a13..671f5e68b1 100644 --- a/cuda-oxide/intrinsics/probes/sync_mask.ll +++ b/cuda-oxide/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_threads.ll b/cuda-oxide/intrinsics/probes/sync_threads.ll index 01a3bb1d2b..23ac2c69d8 100644 --- a/cuda-oxide/intrinsics/probes/sync_threads.ll +++ b/cuda-oxide/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll index 725f001d13..4e0dad07a5 100644 --- a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll index 4681ee92fe..d8d4947635 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll index 188cef95f3..d1b797d4ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll index b18d1467d6..297994ef50 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll index de11e3ecba..5f335632a2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll index d6b22fd8aa..680b6554e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index eb3f9cc385..8186df1a21 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll index 1db2dbdbad..93669f83ad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index f5aa13b3e6..c774b1911d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll index 1e5fe4bb0c..9d22837bce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index 19a6929919..e829843a37 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index c8d680b1cb..be15604c58 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index 915f8808dd..499bc4e5f1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 272b8348b3..21b4796691 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 934cd697ee..604c8c0170 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index 3f3585ca34..46521f2e36 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index bd9c2a8a6e..4a3775a23b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index 6915054db7..0734f6af1f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 2c93c4b8b6..53b5263b46 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index 243e2f4457..609b36a1a0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index d7318b8ecf..89fb8365b3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index 49f8360d73..db65f02265 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index 1b0b9c8428..46c530466d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index fb9294fc00..1c488431c7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index 1f7537312c..5ebb0818a4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll index 705b185f67..8b3b1622a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index 5ad6aabca9..314082c1fa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index 6123b0e8f9..52df3ab376 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 19e703d60c..9107b0378b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index 0a28e83671..98bd6d9825 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 8de4c9caf8..bf2ee26c90 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 8b83fcc6e3..5cf5b80760 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 80745c11f6..3fb56b30ae 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index e224d3f2fe..87a7dce952 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index b5fad6438a..3af2fa6e65 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index b4803dae61..b2219aa9af 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 9d52094864..98558fc3b8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index dab75d57b4..1f9def3fe8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index 4ada1ddd76..fb2acc9680 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index 4e70b6a4cf..d982785274 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 1033e04fb7..3fc4019883 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index 9165307648..f7547ba80d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 2645ba6aa8..27b05800c3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index 0325551ea7..df8ca16961 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index 493cc35b92..511b132a1d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll index fc9d5dd471..3951b4de41 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll index 834b5cd913..4df5378fe2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index badff34939..26cf54f7bc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index da4f8c90e7..3a19ad72e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index 510ca76c46..1ec0cdcaaa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index 39793770f4..c88ad51849 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index 6558b29937..da5e042913 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index 5943a61b0e..6303933af3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index 2c617ef281..42b4000456 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index 3f16f4c1b0..9cd59f5845 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index dabd22d996..b7d8da54ec 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index 9a21034963..55526b4094 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index 5404838fb9..67f98f9372 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index e4e364709f..3ab5f9d7fd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index 34e7e8520c..1b97819713 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index d10bdcb5c2..244065eed7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 49f9fad5c4..5194ca7360 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index 259a47baba..7760557e85 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index ecc652aaaf..54d00ae6ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index fedec7933d..85d311b1fd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index 501dac0681..51a1cc4bee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index c81c06e7bb..e2ce151b30 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index c77bd824dd..1d73cdf17d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index ac2162f914..e27fb796ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 8ba23ed5ab..4a9a9370d1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index 6e1f955780..ac05fcaf87 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index bb00237f03..f7af082e9f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 68482cc33a..6546558e0f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 59defae2ad..20f21f5a93 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index bef54212cc..72b713090f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index 8f5b6ac36f..6150ff47a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 070cbbc91c..a266721e61 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index ad594f196b..cfa637772b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index 60d37021f0..a353a9b6f5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index c57c9da0a7..d215a06150 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index 4d818e944a..98494570ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index ece5446590..1ce69e1644 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 376b7b8102..3eeac417c8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index 4fdc4403e8..6382b60a75 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 680541778b..72ebd6e3dd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index be22d6b1ff..0389053fc7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 6a1ea536a7..68220bafb0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index b584ba9d9b..8cddcb27f6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 34458c1887..fc10354b30 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index 3ca1383000..19337137cd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index 0dad601858..58793d9161 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index eea053e0e9..89a1b7d19e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index e56db75f78..cd85db48a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index f54a4b971e..bfe73d3fd2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index 662a79b2a6..758de0cd92 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index 3015ca3058..88d5b6978c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 5c9a851cfe..d0a4d2bfd2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index 548524be6e..241f318469 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index ee8e288721..d36eb8bc46 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index 6f7edcec98..8eb4a68077 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index 5098ca6a08..964c7a0ea9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 9236a7e66f..06b7225003 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index 9e8ccc97e2..722c1e76e0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index c3dc0adefe..99de889a06 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index 5630c662c3..987c80b112 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 18bf3328bd..469c64b5c9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index da5a8ac721..9cbb7e9dc6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index ea903e9e9d..8a64fef30d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 3ce9fe0bba..6799a8b98c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index 7f5c30cfa2..4fefdf518f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index 6a9ba29f79..55eab790c0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index c86e7781ff..de3a3688d5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index b90be4e669..48b5297598 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index 15b87faad5..a34f46a996 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index 33344e3c04..a297012fcb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index b17d114dc8..0b77d9014a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index af1d53993b..a4cbcd4a7c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index 8bee1d37af..5c6c12b1ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index 2c05ecb2ce..7f7c2d63c1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index 0e80981b8a..34cf97ee5f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index fa8fdfa09e..9c3dd07933 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index c3cbb87b0a..bb372c37cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index ea36ca43df..beae73deff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 6216ada9ac..409acdcb02 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 8da86c6721..74ad0bd0ea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll index 595b69fa29..3deb9f2564 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll index 573fed3b43..42800953d2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll index 27262fdf38..849a8ced59 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll index 6a61de230a..310f7492ef 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll index 98d37ad50d..1a454371be 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll index 6c40df6ff0..26fb8d385d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll index 5308f43130..29c6f223a5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 47f8ca3dd5..98de1423c6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll index 4cd8bba4af..27ee81805c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll index 8bb03d2af6..1e0cc19dd4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll index bb1f776258..6abcd38f66 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index fbd424203c..9195bdca3e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll index 9dbba42424..eff22a1823 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index 9c2515b6ae..ac63ae7be5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 8db4ffbae2..c7d14dd9fd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index c362ea1a36..96e1adbd6e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index b92b3aec0a..5031449cd7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index 042ce7ad76..fa0529a648 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index 9a620b8308..09d67dc49a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 9f3ce75d00..752410fb1b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll index e0e6c7c07b..0a13bb8f85 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll index 09ea38c767..5743c9c4c9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index ed121fd4c4..07d00be372 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index 9e0601053d..e43283a93c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index 1caaecb0f7..7b2b82371d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index 28ac79fae7..f6218e95ee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index e502338674..35ad46af86 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 3906ca1f48..b9863df9f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index 3289dc7937..22d912ebc0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll index cae6ddcfb1..6a5f3f71de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index fbad879555..b9bfa4e286 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index ddfcc64947..b476ca18ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll index 5086100059..709fd4c786 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll index 4003987731..cffdee7985 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index 1b81acc23a..96da6e4d75 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index 0d35a68a1a..0a275483e5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index aa5dcdad3d..9031753772 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index 24ba3b65f2..520c730fab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 7b375acf90..75ba7b4dcb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 1347bdbd3e..2149154a6e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index 8ce48b98ba..88f960628a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index 7c707115e2..df77e39719 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index 29fd5f8215..f24fa552a1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index 24a83eeca9..d99ccc3398 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index 7e08f3cc6c..8ee1f92bf3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index 7359c15811..4a04433402 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index 7508d9512c..a9dab820be 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 81c394680e..5ec0e410da 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index dd617432bb..f7ce12dcd5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index b1431a608d..908f347fc2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index dfd391a873..4efbb6d011 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 113cd6f2c5..8a693b260a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 65278745cb..9003d5c6f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index 32219dc011..cf50911247 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index c2616feaa9..fe984bd8d2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index bab355423a..3f54d5f4f4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index ef892661c9..4618b84509 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 1d16dd79e5..040c863cc2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index b3c367164a..a867eeb9a8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index 1649c07be1..1600215bad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index ea6a58618d..b11371319b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 5fc9cc0c34..5b4a21529b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index 76c9b5e126..7e07817a79 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index 281a56d4e4..63b694a8e5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index 2fa294582b..f192065fb5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index d0d8759e93..c466d4ba2d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index 6710bd6d41..6b35b0ac5a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index fa54eb1721..e2aa2f9668 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index ddda0f9b83..2595bb91e3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index 9989a00130..0bd7fc4c50 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 5cfb0ea15d..39f313995c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index da536ae700..296df3ecd1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index dc58a9dc25..2fae59724a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index 3af785f0bd..584ee36db4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 68bd22e8f9..f6471e38c4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index 7b36c3f64d..4f3563faa6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index eea2ea9957..3dc41182f9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index ac3c80e5a9..f245938c29 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 52564ec661..07db680da4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 2a3c96178a..66e6f15c97 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index cae93a7a62..ec1f065848 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index 506ea62571..0dd568d556 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index 2dddd7ce99..fc9886ad47 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 3994c8a4ba..84e29c19f1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index 15fb8957ff..78232c22b0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 5d05ee095c..d678d472c2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index f2fae2861f..9aee0baaf5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index 73289f8549..928d01c2d0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index acc70601f7..12a6c4e9d3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index 2574645ac3..2c528316f4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index cf8ac59413..399c847dcf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index f7fdb470e7..a98fc65a80 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index b2b72b7be7..dfe8fb10a0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index 51c111b66c..18dc360213 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index e905a65ccf..748b1f9b32 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index dfce971dbe..943c7cbecf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index a0daa1fa67..5f043c91c8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index 27f2dadbcd..7c7301e2f2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index 65292054b9..375ddcae95 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 8dc86b0fb2..7c0fad9857 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index 04be3d6a60..c7b71fa69b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 0ce49b7860..0fbd11cafd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index 3bd1c7461b..816202a829 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index 9e8e75fda8..245dc44bc5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index 269ef07736..9ec39e8d5d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index d24ba123b9..b295852821 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index 25074b9b4e..3f9ab7e0e4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index 423a7b40f7..0d02551a73 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll index a90834a825..4bcbb78e44 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll index 0b8a01a198..e570b7310a 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll index 576978840f..3e3826968d 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll index 786ce3e147..01f5d53a2c 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll index 983a5c7f45..4ebefec49a 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll index 6be7af57f0..c3dfc12236 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll index 69f632e4f0..046affc441 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll index c3528e6c92..ff3b827b2d 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll index dcf8a45027..31b193d505 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll index b82acad78c..64984bfa3e 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index e7ca4939ea..8b85971e9a 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll index f0faaacf67..79b2160a5f 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_x.ll b/cuda-oxide/intrinsics/probes/thread_idx_x.ll index 861f7b549e..4d8cb82890 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_y.ll b/cuda-oxide/intrinsics/probes/thread_idx_y.ll index b9c97e960a..17e9c68136 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_z.ll b/cuda-oxide/intrinsics/probes/thread_idx_z.ll index 074d5c6627..78dfc2d2ba 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence.ll b/cuda-oxide/intrinsics/probes/threadfence.ll index 81bb0b1da6..0b8fc8eb45 100644 --- a/cuda-oxide/intrinsics/probes/threadfence.ll +++ b/cuda-oxide/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_block.ll b/cuda-oxide/intrinsics/probes/threadfence_block.ll index 2422f916a0..50af1914b8 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_block.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_system.ll b/cuda-oxide/intrinsics/probes/threadfence_system.ll index d5181821c8..8afc25bda4 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_system.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll index 135eb11317..ccee72572f 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 9124402f87..9387c0646f 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/trap.ll b/cuda-oxide/intrinsics/probes/trap.ll index 91eec43597..23efc3bf94 100644 --- a/cuda-oxide/intrinsics/probes/trap.ll +++ b/cuda-oxide/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/uni_sync.ll b/cuda-oxide/intrinsics/probes/uni_sync.ll index b990da9cc0..5987a4d7c5 100644 --- a/cuda-oxide/intrinsics/probes/uni_sync.ll +++ b/cuda-oxide/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll index b36a3105c1..2d1ddd17ba 100644 --- a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll index c747459656..16fea2fef7 100644 --- a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll index 3584d85f0b..52ef66f4d0 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_fence.ll b/cuda-oxide/intrinsics/probes/wgmma_fence.ll index a2b40a4c89..6744b757fc 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_fence.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll index 8ba5bde0b4..e269d7ec9e 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 7e6b4b4b3b5958a0a36fb0c40486b8263e0799fbe1f23cb9be64b6687f27e3a0. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" From a5f14c7a24ba5efbd4d2cdadab580f74e93ce78c Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 20:57:31 +0530 Subject: [PATCH 09/17] docs(book): refresh the dialect-nvvm catalog stamp after the second merge The merged catalog hashes to 7e6b4b4b and resolves 1038 entries into 559 generated operations, 585 in all with the 26 hand-written ones. The tma and packed rows were already current. Signed-off-by: nihalpasham --- cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md index 3e9a1ba82e..7a22f93ffa 100644 --- a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `31b52388` (the stamp in every `ops/generated/` file -header), the dialect holds 580 operations across 42 modules, and they come +At catalog SHA-256 `7e6b4b4b` (the stamp in every `ops/generated/` file +header), the dialect holds 585 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `31b52388`, the counts on this page predate the +stamp no longer starts with `7e6b4b4b`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `cuda-oxide/crates/dialect-nvvm/src/ops/`. These are the @@ -327,7 +327,7 @@ describe. There are seven modules and 26 operations: **Generated**, under `ops/generated/`, from `intrinsics/catalog.json` by `cuda-intrinsics-gen`. Every file there opens with `// @generated ... DO NOT EDIT.`, and editing one by hand is undone by the next generator run. This is -the large majority -- 35 modules and 554 operations, resolved from 1033 catalog +the large majority -- 35 modules and 559 operations, resolved from 1038 catalog entries, since several intrinsics can share one structural op: | Area | Modules | Ops | From c28c841a1474833cee240535d00d74a793bc1a72 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 10/17] fix(cuda-device): make SharedArray::shared_ptr an unsafe constructor The constructor is what makes a SharedPtr's locality claim true, and the safe addr_u32 relies on it. It now carries the same Safety section as as_raw_mut_ptr: a static mut SharedArray of the current kernel, obtained with &raw mut. DynamicSharedArray::shared_ptr and shared_ptr_at stay safe, as get and offset are. The doc examples wrap the call in unsafe. Signed-off-by: nihalpasham --- cuda-oxide/crates/cuda-device/src/shared.rs | 23 +++++++++++++++------ 1 file changed, 17 insertions(+), 6 deletions(-) diff --git a/cuda-oxide/crates/cuda-device/src/shared.rs b/cuda-oxide/crates/cuda-device/src/shared.rs index 733ff98ba9..790e24078d 100644 --- a/cuda-oxide/crates/cuda-device/src/shared.rs +++ b/cuda-oxide/crates/cuda-device/src/shared.rs @@ -240,12 +240,23 @@ impl SharedArray { /// result keeps the array's CTA-shared address space in its type, so it /// stays a shared-memory pointer when it is passed to helper functions. /// + /// # Safety + /// + /// `shared` must point to a `static mut SharedArray` in the current CUDA + /// kernel, normally obtained with `&raw mut`. The returned pointer is valid + /// only within that kernel. This requirement is what makes the type's + /// claim true: a `SharedPtr` built from any other pointer hands + /// [`SharedPtr::addr_u32`] an address outside the shared window, where the + /// conversion yields an unspecified value. Callers must keep concurrent + /// accesses disjoint and use the CUDA synchronization required before + /// reading data written by another thread. + /// /// ```rust,ignore /// static mut TILE: SharedArray = SharedArray::UNINIT; - /// let tile: SharedPtr = SharedArray::shared_ptr(&raw mut TILE); + /// let tile: SharedPtr = unsafe { SharedArray::shared_ptr(&raw mut TILE) }; /// ``` #[inline(always)] - pub fn shared_ptr(shared: *mut Self) -> SharedPtr { + pub unsafe fn shared_ptr(shared: *mut Self) -> SharedPtr { SharedPtr { ptr: shared as *mut SharedSlot, } @@ -629,9 +640,9 @@ pub struct SharedSlot(T); /// through calls to helper functions, and emits the `.shared::cta` forms of /// instructions that take it without converting a generic address. /// -/// It is created from shared memory the CTA owns, with -/// [`SharedArray::shared_ptr`] or [`DynamicSharedArray::shared_ptr`], or from -/// a generic pointer with the `unsafe` [`SharedPtr::from_raw`]. Like a raw +/// It is created from shared memory the CTA owns, with the `unsafe` +/// [`SharedArray::shared_ptr`] or the safe [`DynamicSharedArray::shared_ptr`], +/// or from a generic pointer with the `unsafe` [`SharedPtr::from_raw`]. Like a raw /// pointer, it may dangle; dereferencing it goes through [`SharedPtr::as_ptr`] /// and `unsafe` code. /// @@ -649,7 +660,7 @@ pub struct SharedSlot(T); /// unsafe { tile.add(i).as_ptr().write(value) }; /// } /// -/// let tile = SharedArray::shared_ptr(&raw mut TILE); +/// let tile = unsafe { SharedArray::shared_ptr(&raw mut TILE) }; /// store(tile, thread::threadIdx_x() as usize, 1.0); /// ``` #[repr(transparent)] From a09a080b1ba84afc3fbe0f9a5d17dd5b52ec7a18 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 11/17] fix(llvm-export): refuse kernel parameters that hold shared or local pointers inside aggregates The kernel-parameter check downcast only the top-level type, so a struct { p: SharedPtr, n: u32 } parameter compiled, declared a 16-byte .param and applied cvta.shared.u64 to host bytes; a *mut Barrier field had the same hole. The check now looks through struct fields and array elements and names the field path it found, with a message distinct from the bare form so a fixture can pin the nested case. Unit tests cover a struct, an array, a pointer nested below an array element, and the accepted shapes (a struct of generic and global pointers on a kernel, the shared-holding struct on a device function). The new error_kernel_shared_ptr_param fixture pins the nested diagnostic in the smoketest error lane, with its STATUS.md row and verdict case. Signed-off-by: nihalpasham --- .../crates/llvm-export/src/export/function.rs | 76 +- .../tests/export_test/kernels_and_roots.rs | 176 +++- .../crates/rustc-codegen-cuda/STATUS.md | 1 + .../error_kernel_shared_ptr_param/Cargo.lock | 753 ++++++++++++++++++ .../error_kernel_shared_ptr_param/Cargo.toml | 13 + .../error_kernel_shared_ptr_param/README.md | 32 + .../error_kernel_shared_ptr_param/src/main.rs | 51 ++ cuda-oxide/scripts/smoketest.sh | 11 +- 8 files changed, 1099 insertions(+), 14 deletions(-) create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.lock create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.toml create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/README.md create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/src/main.rs diff --git a/cuda-oxide/crates/llvm-export/src/export/function.rs b/cuda-oxide/crates/llvm-export/src/export/function.rs index f8e6e1d4d0..db8900f9a2 100644 --- a/cuda-oxide/crates/llvm-export/src/export/function.rs +++ b/cuda-oxide/crates/llvm-export/src/export/function.rs @@ -65,6 +65,43 @@ fn sanitize_comment(text: &str) -> String { /// Public export callers need not have verified the source module. Pliron 0.19 /// also sums segment sizes as u32 without checking overflow, so validate the /// complete layout here before it can index or allocate from malformed counts. +/// Looks for a pointer into shared (address space 3) or local (address space +/// 5) memory inside a kernel parameter type, through struct fields and array +/// elements. On a hit it returns the address space and appends the path that +/// reaches the pointer to `path`: `.N` for struct field `N` and `[0]` for an +/// array element (every element has the same type, so the first stands for +/// all of them). A bare pointer leaves `path` empty. +fn device_only_pointer_in(ctx: &Context, ty: TypeHandle, path: &mut String) -> Option { + let ty_ref = ty.deref(ctx); + if let Some(pointer) = ty_ref.downcast_ref::() { + let space = pointer.address_space(); + return matches!(space, 3 | 5).then_some(space); + } + if let Some(array) = ty_ref.downcast_ref::() { + let mark = path.len(); + path.push_str("[0]"); + if let Some(space) = device_only_pointer_in(ctx, array.elem_type(), path) { + return Some(space); + } + path.truncate(mark); + return None; + } + if let Some(structure) = ty_ref.downcast_ref::() { + if structure.is_opaque() { + return None; + } + for (field_index, field) in structure.fields().enumerate() { + let mark = path.len(); + path.push_str(&format!(".{field_index}")); + if let Some(space) = device_only_pointer_in(ctx, field, path) { + return Some(space); + } + path.truncate(mark); + } + } + None +} + fn validate_conditional_branch_layout(ctx: &Context, branch: &ops::CondBrOp) -> Result<(), String> { let operation = branch.get_operation().deref(ctx); let sizes = operation @@ -999,23 +1036,40 @@ impl<'a> ModuleExportState<'a> { // parameter type today; if a frontend type ever surfaces it at a // kernel boundary, it needs a new arm here for the same reason as // plain shared. + // + // The check looks through struct fields and array elements as well. + // A struct or array parameter reaches the kernel as the bytes the + // host wrote, so a shared or local pointer inside one is as + // impossible to supply as a bare one. Before this recursion a + // `struct { p: SharedPtr, n: u32 }` parameter compiled and the + // kernel applied `cvta.shared.u64` to host bytes. if is_kernel { for (index, arg_ty) in func_ty.arg_types().iter().enumerate() { - let arg_ref = arg_ty.deref(self.ctx); - let Some(pointer) = arg_ref.downcast_ref::() else { + let mut path = String::new(); + let Some(address_space) = device_only_pointer_in(self.ctx, *arg_ty, &mut path) + else { continue; }; - let space = match pointer.address_space() { - 3 => "shared", - 5 => "local", - _ => continue, + let space = if address_space == 3 { + "shared" + } else { + "local" }; + let reason = format!( + "{space} memory is allocated by the device at launch, so the host has no \ + address to pass, and the driver refuses a module whose entry declares one" + ); + if path.is_empty() { + return Err(format!( + "kernel `@{fixed_func_name}` parameter {index} is a pointer into {space} \ + memory (address space {address_space}); {reason}" + )); + } + // The field path keeps this message apart from the bare form, + // so a fixture can pin the nested case on its own. return Err(format!( - "kernel `@{fixed_func_name}` parameter {index} is a pointer into {space} \ - memory (address space {}); {space} memory is allocated by the device at \ - launch, so the host has no address to pass, and the driver refuses a \ - module whose entry declares one", - pointer.address_space() + "kernel `@{fixed_func_name}` parameter {index} holds a pointer into {space} \ + memory (address space {address_space}) at field path `{path}`; {reason}" )); } } diff --git a/cuda-oxide/crates/llvm-export/tests/export_test/kernels_and_roots.rs b/cuda-oxide/crates/llvm-export/tests/export_test/kernels_and_roots.rs index c557166dfb..d80832d0a2 100644 --- a/cuda-oxide/crates/llvm-export/tests/export_test/kernels_and_roots.rs +++ b/cuda-oxide/crates/llvm-export/tests/export_test/kernels_and_roots.rs @@ -10,7 +10,7 @@ use llvm_export::{ export_module_with_externs_and_roots, }, ops::{FuncOp, GlobalOp, GlobalOpExt, ReturnOp}, - types::{FuncType, PointerType, VoidType}, + types::{ArrayType, FuncType, PointerType, StructLayout, StructType, VoidType}, }; use pliron::{ basic_block::BasicBlock, @@ -22,6 +22,7 @@ use pliron::{ context::{Context, Ptr}, identifier::Identifier, op::Op, + r#type::TypeHandle, }; use crate::common::module_top_block; @@ -246,10 +247,22 @@ fn pointer_param_function( address_spaces: &[u32], is_kernel: bool, ) { - let params: Vec<_> = address_spaces + let params: Vec = address_spaces .iter() .map(|space| PointerType::get(ctx, *space).into()) .collect(); + typed_param_function(ctx, module_block, name, params, is_kernel); +} + +/// Builds a `void` function with the given parameter types and an empty +/// body, optionally carrying the `gpu_kernel` marker. +fn typed_param_function( + ctx: &mut Context, + module_block: Ptr, + name: &str, + params: Vec, + is_kernel: bool, +) { let void_ty = VoidType::get(ctx); let func_ty = FuncType::get(ctx, void_ty.into(), params, false); let func = FuncOp::new(ctx, name.try_into().unwrap(), func_ty); @@ -307,6 +320,165 @@ fn kernel_rejects_a_local_memory_pointer_parameter() { ); } +/// The host fills a struct parameter byte by byte, so a shared pointer +/// inside one is as impossible to supply as a bare one. The exporter looks +/// through the aggregate and names the field path it found, with a message +/// distinct from the bare form. +#[test] +fn kernel_rejects_a_shared_memory_pointer_inside_a_struct_parameter() { + let mut ctx = Context::new(); + let module = ModuleOp::new(&mut ctx, "kernel_shared_struct_param".try_into().unwrap()); + let module_block = module_top_block(&mut ctx, &module); + let shared_ptr_ty = PointerType::get(&ctx, 3); + let i32_ty = IntegerType::get(&ctx, 32, Signedness::Signless); + let args_ty = StructType::get_unnamed( + &ctx, + ( + vec![shared_ptr_ty.into(), i32_ty.into()], + StructLayout::Unpacked, + ), + ); + typed_param_function( + &mut ctx, + module_block, + "shared_struct_param", + vec![args_ty.into()], + true, + ); + + let error = export_module_to_string(&ctx, &module) + .expect_err("export must refuse a kernel parameter struct that holds a shared pointer"); + assert!( + error.contains( + "kernel `@shared_struct_param` parameter 0 holds a pointer into shared memory \ + (address space 3) at field path `.0`" + ), + "{error}" + ); +} + +/// An array parameter is refused the same way, and the path names the +/// element position. +#[test] +fn kernel_rejects_a_shared_memory_pointer_inside_an_array_parameter() { + let mut ctx = Context::new(); + let module = ModuleOp::new(&mut ctx, "kernel_shared_array_param".try_into().unwrap()); + let module_block = module_top_block(&mut ctx, &module); + let generic_ptr_ty = PointerType::get(&ctx, 0); + let shared_ptr_ty = PointerType::get(&ctx, 3); + let pair_ty = ArrayType::get(&ctx, shared_ptr_ty.into(), 2); + typed_param_function( + &mut ctx, + module_block, + "shared_array_param", + vec![generic_ptr_ty.into(), pair_ty.into()], + true, + ); + + let error = export_module_to_string(&ctx, &module) + .expect_err("export must refuse a kernel parameter array of shared pointers"); + assert!( + error.contains( + "kernel `@shared_array_param` parameter 1 holds a pointer into shared memory \ + (address space 3) at field path `[0]`" + ), + "{error}" + ); +} + +/// A pointer buried under a struct, an array and another struct is found and +/// reported with its full path; local memory is refused on the same grounds. +#[test] +fn kernel_rejects_a_local_memory_pointer_nested_below_an_array_element() { + let mut ctx = Context::new(); + let module = ModuleOp::new(&mut ctx, "kernel_local_nested_param".try_into().unwrap()); + let module_block = module_top_block(&mut ctx, &module); + let local_ptr_ty = PointerType::get(&ctx, 5); + let i32_ty = IntegerType::get(&ctx, 32, Signedness::Signless); + let leaf_ty = StructType::get_unnamed( + &ctx, + ( + vec![local_ptr_ty.into(), i32_ty.into()], + StructLayout::Unpacked, + ), + ); + let pair_ty = ArrayType::get(&ctx, leaf_ty.into(), 2); + let args_ty = StructType::get_unnamed( + &ctx, + (vec![i32_ty.into(), pair_ty.into()], StructLayout::Unpacked), + ); + typed_param_function( + &mut ctx, + module_block, + "local_nested_param", + vec![args_ty.into()], + true, + ); + + let error = export_module_to_string(&ctx, &module) + .expect_err("export must refuse a nested local-memory pointer in a kernel parameter"); + assert!( + error.contains( + "kernel `@local_nested_param` parameter 0 holds a pointer into local memory \ + (address space 5) at field path `.1[0].0`" + ), + "{error}" + ); +} + +/// The recursion refuses only what the host cannot supply: a struct of +/// generic and global pointers stays a valid kernel parameter, and a device +/// (non-kernel) function may take the struct that holds a shared pointer. +#[test] +fn kernel_keeps_a_struct_of_host_addressable_pointers_and_device_functions_keep_nested_shared() { + let mut ctx = Context::new(); + let module = ModuleOp::new(&mut ctx, "kernel_allowed_struct_params".try_into().unwrap()); + let module_block = module_top_block(&mut ctx, &module); + let generic_ptr_ty = PointerType::get(&ctx, 0); + let global_ptr_ty = PointerType::get(&ctx, 1); + let shared_ptr_ty = PointerType::get(&ctx, 3); + let i32_ty = IntegerType::get(&ctx, 32, Signedness::Signless); + let host_ty = StructType::get_unnamed( + &ctx, + ( + vec![generic_ptr_ty.into(), global_ptr_ty.into()], + StructLayout::Unpacked, + ), + ); + let shared_ty = StructType::get_unnamed( + &ctx, + ( + vec![shared_ptr_ty.into(), i32_ty.into()], + StructLayout::Unpacked, + ), + ); + typed_param_function( + &mut ctx, + module_block, + "host_struct_param", + vec![host_ty.into()], + true, + ); + typed_param_function( + &mut ctx, + module_block, + "device_shared_struct", + vec![shared_ty.into()], + false, + ); + + let ir = export_module_to_string(&ctx, &module) + .expect("a host-addressable struct parameter and a device function's shared struct export"); + assert!( + ir.contains("define ptx_kernel void @host_struct_param({ ptr, ptr addrspace(1) } %v0)"), + "{ir}" + ); + assert!( + ir.contains("@device_shared_struct({ ptr addrspace(3), i32 } %v0)"), + "{ir}" + ); +} + /// The refusal is scoped to what the host cannot supply. Generic, global, /// and constant pointer parameters are host-addressable and stay accepted on /// kernels, and a device (non-kernel) function may carry any state space on diff --git a/cuda-oxide/crates/rustc-codegen-cuda/STATUS.md b/cuda-oxide/crates/rustc-codegen-cuda/STATUS.md index f42a90129f..94b339cbda 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/STATUS.md +++ b/cuda-oxide/crates/rustc-codegen-cuda/STATUS.md @@ -25,6 +25,7 @@ Run `cuda-oxide/scripts/check-error-example-status.sh` to verify both are in syn | `error_host_arch_intrinsic` | diagnostics-fixture | `core::arch::` intrinsic reached | | `error_host_target_feature` | diagnostics-fixture | `#[target_feature]` fn reached | | `error_kernel_shared_param` | diagnostics-fixture | AS3/AS5 pointer as kernel parameter | +| `error_kernel_shared_ptr_param` | diagnostics-fixture | AS3 pointer inside a kernel parameter struct | | `error_missing_device_attr` | diagnostics-fixture | `thread::index_*` stub (#76) | | `error_set_discriminant_uninhabited` | diagnostics-fixture | Invalid enum variant selection | | `error_union_shared_pointer_addr` | diagnostics-fixture | `&mut`/in-place write to a shared-pointer union field | diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.lock b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.lock new file mode 100644 index 0000000000..c46db5e9e1 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.lock @@ -0,0 +1,753 @@ +# This file is automatically @generated by Cargo. +# It is not intended for manual editing. +version = 4 + +[[package]] +name = "aho-corasick" +version = "1.1.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301" +dependencies = [ + "memchr", +] + +[[package]] +name = "anyhow" +version = "1.0.102" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c" + +[[package]] +name = "bindgen" +version = "0.69.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "271383c67ccabffb7381723dea0672a673f292304fcb45c01cc648c7a8d58088" +dependencies = [ + "bitflags", + "cexpr", + "clang-sys", + "itertools", + "lazy_static", + "lazycell", + "log", + "prettyplease", + "proc-macro2", + "quote", + "regex", + "rustc-hash", + "shlex", + "syn", + "which", +] + +[[package]] +name = "bitflags" +version = "2.13.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8" + +[[package]] +name = "block-buffer" +version = "0.10.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71" +dependencies = [ + "generic-array", +] + +[[package]] +name = "cexpr" +version = "0.6.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6fac387a98bb7c37292057cffc56d62ecb629900026402633ae9160df93a8766" +dependencies = [ + "nom", +] + +[[package]] +name = "cfg-if" +version = "1.0.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801" + +[[package]] +name = "clang-sys" +version = "1.8.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0b023947811758c97c59bf9d1c188fd619ad4718dcaa767947df1cadb14f39f4" +dependencies = [ + "glob", + "libc", + "libloading", +] + +[[package]] +name = "cpufeatures" +version = "0.2.17" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280" +dependencies = [ + "libc", +] + +[[package]] +name = "crc32fast" +version = "1.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9481c1c90cbf2ac953f07c8d4a58aa3945c425b7185c9154d67a65e4230da511" +dependencies = [ + "cfg-if", +] + +[[package]] +name = "crunchy" +version = "0.2.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5" + +[[package]] +name = "crypto-common" +version = "0.1.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a" +dependencies = [ + "generic-array", + "typenum", +] + +[[package]] +name = "cuda-artifact-finalizer" +version = "0.2.1" +dependencies = [ + "libnvvm-sys", + "nvjitlink-sys", + "serde", + "sha2", + "thiserror", +] + +[[package]] +name = "cuda-bindings" +version = "0.4.0" +dependencies = [ + "bindgen", + "libloading", + "prettyplease", + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "cuda-core" +version = "0.4.0" +dependencies = [ + "anyhow", + "cuda-bindings", + "cuda-core-derive", + "half", + "oxide-artifacts", +] + +[[package]] +name = "cuda-core-derive" +version = "0.4.0" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "cuda-device" +version = "0.2.1" +dependencies = [ + "cuda-macros", +] + +[[package]] +name = "cuda-host" +version = "0.2.1" +dependencies = [ + "cuda-artifact-finalizer", + "cuda-core", + "cuda-macros", + "half", + "libc", + "oxide-artifacts", + "ptx-parse", + "sha2", + "thiserror", +] + +[[package]] +name = "cuda-macros" +version = "0.2.1" +dependencies = [ + "oxide-tile-export", + "proc-macro2", + "quote", + "reserved-oxide-symbols", + "syn", +] + +[[package]] +name = "cuda-target-spec" +version = "0.1.0" + +[[package]] +name = "digest" +version = "0.10.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292" +dependencies = [ + "block-buffer", + "crypto-common", +] + +[[package]] +name = "either" +version = "1.16.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e" + +[[package]] +name = "equivalent" +version = "1.0.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f" + +[[package]] +name = "errno" +version = "0.3.14" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb" +dependencies = [ + "libc", + "windows-sys 0.61.2", +] + +[[package]] +name = "error_kernel_shared_ptr_param" +version = "0.1.0" +dependencies = [ + "cuda-core", + "cuda-device", + "cuda-host", +] + +[[package]] +name = "foldhash" +version = "0.1.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2" + +[[package]] +name = "generic-array" +version = "0.14.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a" +dependencies = [ + "typenum", + "version_check", +] + +[[package]] +name = "glob" +version = "0.3.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0cc23270f6e1808e30a928bdc84dea0b9b4136a8bc82338574f23baf47bbd280" + +[[package]] +name = "half" +version = "2.7.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b" +dependencies = [ + "cfg-if", + "crunchy", + "zerocopy", +] + +[[package]] +name = "hashbrown" +version = "0.15.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1" +dependencies = [ + "foldhash", +] + +[[package]] +name = "hashbrown" +version = "0.17.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a" + +[[package]] +name = "home" +version = "0.5.12" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "cc627f471c528ff0c4a49e1d5e60450c8f6461dd6d10ba9dcd3a61d3dff7728d" +dependencies = [ + "windows-sys 0.61.2", +] + +[[package]] +name = "indexmap" +version = "2.14.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9" +dependencies = [ + "equivalent", + "hashbrown 0.17.1", +] + +[[package]] +name = "itertools" +version = "0.12.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ba291022dbbd398a455acf126c1e341954079855bc60dfdda641363bd6922569" +dependencies = [ + "either", +] + +[[package]] +name = "itoa" +version = "1.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682" + +[[package]] +name = "lazy_static" +version = "1.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe" + +[[package]] +name = "lazycell" +version = "1.3.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "830d08ce1d1d941e6b30645f1a0eb5643013d835ce3779a5fc208261dbe10f55" + +[[package]] +name = "libc" +version = "0.2.186" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66" + +[[package]] +name = "libloading" +version = "0.8.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d7c4b02199fee7c5d21a5ae7d8cfa79a6ef5bb2fc834d6e9058e89c825efdc55" +dependencies = [ + "cfg-if", + "windows-link", +] + +[[package]] +name = "libnvvm-sys" +version = "0.2.1" +dependencies = [ + "cuda-target-spec", + "libloading", + "thiserror", +] + +[[package]] +name = "linux-raw-sys" +version = "0.4.15" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d26c52dbd32dccf2d10cac7725f8eae5296885fb5703b261f7d0a0739ec807ab" + +[[package]] +name = "log" +version = "0.4.29" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897" + +[[package]] +name = "memchr" +version = "2.8.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98" + +[[package]] +name = "minimal-lexical" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "68354c5c6bd36d73ff3feceb05efa59b6acb7626617f4962be322a825e61f79a" + +[[package]] +name = "nom" +version = "7.1.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d273983c5a657a70a3e8f2a01329822f3b8c8172b73826411a55751e404a0a4a" +dependencies = [ + "memchr", + "minimal-lexical", +] + +[[package]] +name = "nvjitlink-sys" +version = "0.2.1" +dependencies = [ + "libloading", + "thiserror", +] + +[[package]] +name = "object" +version = "0.36.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "62948e14d923ea95ea2c7c86c71013138b66525b86bdc08d2dcc262bdb497b87" +dependencies = [ + "crc32fast", + "hashbrown 0.15.5", + "indexmap", + "memchr", +] + +[[package]] +name = "once_cell" +version = "1.21.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50" + +[[package]] +name = "oxide-artifacts" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "74d9a941c0c43d2dceaf25f5c9ca4c9f9cceb28dfa5e4e3cfd14e476f003e7c9" +dependencies = [ + "object", +] + +[[package]] +name = "oxide-tile-export" +version = "0.2.1" +dependencies = [ + "serde", + "serde_json", + "thiserror", +] + +[[package]] +name = "prettyplease" +version = "0.2.37" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b" +dependencies = [ + "proc-macro2", + "syn", +] + +[[package]] +name = "proc-macro2" +version = "1.0.106" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934" +dependencies = [ + "unicode-ident", +] + +[[package]] +name = "ptx-parse" +version = "0.1.0" + +[[package]] +name = "quote" +version = "1.0.45" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924" +dependencies = [ + "proc-macro2", +] + +[[package]] +name = "regex" +version = "1.12.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276" +dependencies = [ + "aho-corasick", + "memchr", + "regex-automata", + "regex-syntax", +] + +[[package]] +name = "regex-automata" +version = "0.4.14" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f" +dependencies = [ + "aho-corasick", + "memchr", + "regex-syntax", +] + +[[package]] +name = "regex-syntax" +version = "0.8.11" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4" + +[[package]] +name = "reserved-oxide-symbols" +version = "0.2.1" + +[[package]] +name = "rustc-hash" +version = "1.1.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2" + +[[package]] +name = "rustix" +version = "0.38.44" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "fdb5bc1ae2baa591800df16c9ca78619bf65c0488b41b96ccec5d11220d8c154" +dependencies = [ + "bitflags", + "errno", + "libc", + "linux-raw-sys", + "windows-sys 0.59.0", +] + +[[package]] +name = "serde" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e" +dependencies = [ + "serde_core", + "serde_derive", +] + +[[package]] +name = "serde_core" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad" +dependencies = [ + "serde_derive", +] + +[[package]] +name = "serde_derive" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "serde_json" +version = "1.0.154" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e7e9cc8b1b85264074fbcc02a88680c4096b1e47df8f739dceb03bf482f04bd6" +dependencies = [ + "itoa", + "memchr", + "serde", + "serde_core", + "zmij", +] + +[[package]] +name = "sha2" +version = "0.10.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283" +dependencies = [ + "cfg-if", + "cpufeatures", + "digest", +] + +[[package]] +name = "shlex" +version = "1.3.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0fda2ff0d084019ba4d7c6f371c95d8fd75ce3524c3cb8fb653a3023f6323e64" + +[[package]] +name = "syn" +version = "2.0.117" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99" +dependencies = [ + "proc-macro2", + "quote", + "unicode-ident", +] + +[[package]] +name = "thiserror" +version = "2.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4" +dependencies = [ + "thiserror-impl", +] + +[[package]] +name = "thiserror-impl" +version = "2.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "typenum" +version = "1.20.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20" + +[[package]] +name = "unicode-ident" +version = "1.0.24" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75" + +[[package]] +name = "version_check" +version = "0.9.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a" + +[[package]] +name = "which" +version = "4.4.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "87ba24419a2078cd2b0f2ede2691b6c66d8e47836da3b6db8265ebad47afbfc7" +dependencies = [ + "either", + "home", + "once_cell", + "rustix", +] + +[[package]] +name = "windows-link" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5" + +[[package]] +name = "windows-sys" +version = "0.59.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1e38bc4d79ed67fd075bcc251a1c39b32a1776bbe92e5bef1f0bf1f8c531853b" +dependencies = [ + "windows-targets", +] + +[[package]] +name = "windows-sys" +version = "0.61.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc" +dependencies = [ + "windows-link", +] + +[[package]] +name = "windows-targets" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973" +dependencies = [ + "windows_aarch64_gnullvm", + "windows_aarch64_msvc", + "windows_i686_gnu", + "windows_i686_gnullvm", + "windows_i686_msvc", + "windows_x86_64_gnu", + "windows_x86_64_gnullvm", + "windows_x86_64_msvc", +] + +[[package]] +name = "windows_aarch64_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3" + +[[package]] +name = "windows_aarch64_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469" + +[[package]] +name = "windows_i686_gnu" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b" + +[[package]] +name = "windows_i686_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66" + +[[package]] +name = "windows_i686_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66" + +[[package]] +name = "windows_x86_64_gnu" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78" + +[[package]] +name = "windows_x86_64_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d" + +[[package]] +name = "windows_x86_64_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec" + +[[package]] +name = "zerocopy" +version = "0.8.52" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ce1022995ff5ff5d841ad7d994facc23098cd40152f2c1d11cd607c6f530653f" +dependencies = [ + "zerocopy-derive", +] + +[[package]] +name = "zerocopy-derive" +version = "0.8.52" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1ae7f38b72ec2a254e2b87ef277cf2cd4fb97cbebf944faa6f33354da0867930" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "zmij" +version = "1.0.23" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b" diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.toml b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.toml new file mode 100644 index 0000000000..15fee6ddd1 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/Cargo.toml @@ -0,0 +1,13 @@ +[package] +name = "error_kernel_shared_ptr_param" +version = "0.1.0" +edition = "2024" +authors = ["NVIDIA Corporation"] +license = "Apache-2.0" + +[workspace] + +[dependencies] +cuda-core = { path = "../../../../../cuda-core" } +cuda-device = { path = "../../../cuda-device" } +cuda-host = { path = "../../../cuda-host" } diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/README.md b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/README.md new file mode 100644 index 0000000000..0fcfdc5d95 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/README.md @@ -0,0 +1,32 @@ +# `error_kernel_shared_ptr_param` + +Negative test for a `#[kernel]` whose parameter is a struct that holds a +`SharedPtr`, a pointer that lowers to `addrspace(3)` by its type. + +A kernel receives its parameters in `.param` space, filled by the host at +launch. Shared memory is allocated per block by the device, so the host holds +no address in it to pass. `error_kernel_shared_param` pins the refusal of a +bare shared pointer parameter. This fixture pins the same refusal for a +pointer carried inside an aggregate: before the exporter looked through struct +fields and array elements, `Args { p: SharedPtr, n: u32 }` compiled, the +entry declared a 16-byte `.param`, and the kernel applied `cvta.shared.u64` to +the host-supplied bytes as if they were a shared address. + +The compiler must reject the signature at compile time instead: + +```bash +cargo oxide build error_kernel_shared_ptr_param +``` + +Expected diagnostic (pinned by `cuda-oxide/scripts/smoketest.sh`): + +```text +holds a pointer into shared memory +``` + +The diagnostic names the field path inside the parameter (`.0` here), which +keeps it distinct from the bare parameter's `is a pointer into shared memory`. +Structs of global (`addrspace(1)`) and constant (`addrspace(4)`) pointers stay +allowed as kernel parameters: the host allocates in those spaces, so it has an +address to supply. A `SharedPtr` belongs on the device side, created inside +the kernel from a `static mut SharedArray` or from `DynamicSharedArray`. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/src/main.rs b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/src/main.rs new file mode 100644 index 0000000000..74f8b80c79 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/error_kernel_shared_ptr_param/src/main.rs @@ -0,0 +1,51 @@ +/* + * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + * SPDX-License-Identifier: Apache-2.0 + */ + +//! Negative test: a `#[kernel]` parameter struct that holds a `SharedPtr`. +//! +//! A kernel receives its parameters in `.param` space, filled by the host at +//! launch. Shared memory is allocated per block by the device, so the host +//! holds no shared-memory address it could pass. A bare `SharedPtr` +//! parameter is refused for that reason (`error_kernel_shared_param` covers +//! the bare case with `*mut Barrier`), and a struct or array that carries one +//! must be refused the same way. Before the exporter looked inside +//! aggregates, `Args { p: SharedPtr, n: u32 }` compiled and the kernel +//! converted the host-supplied bytes with `cvta.shared.u64` as if they were a +//! shared address. +//! +//! Usage: +//! cargo oxide run error_kernel_shared_ptr_param +//! +//! Expected: the build FAILS with a diagnostic containing (pinned): +//! +//! ```text +//! holds a pointer into shared memory +//! ``` + +use cuda_device::{SharedPtr, kernel}; + +/// Kernel arguments that carry a typed shared pointer next to a plain field. +#[derive(Clone, Copy)] +pub struct Args { + pub p: SharedPtr, + pub n: u32, +} + +/// # Safety +/// +/// Never launched: the device compilation is expected to fail before a host +/// binary that could launch it exists. +#[kernel] +pub unsafe fn shared_ptr_param_kernel(args: Args) { + // BUG UNDER TEST: `args.p` points into shared memory, which the host + // cannot supply at launch. The refusal fires on the entry signature + // alone; the body deliberately does nothing else, so the diagnostic + // under test is the only error this fixture can produce. + let _ = args; +} + +fn main() { + println!("This negative example should fail during device compilation."); +} diff --git a/cuda-oxide/scripts/smoketest.sh b/cuda-oxide/scripts/smoketest.sh index 0753ef779a..fa442336ff 100755 --- a/cuda-oxide/scripts/smoketest.sh +++ b/cuda-oxide/scripts/smoketest.sh @@ -98,7 +98,7 @@ IKET_EXAMPLES=(iket_trace) BLACKWELL_COMPILE_EXAMPLES=(generated_intrinsics_blackwell) SM100_COMPILE_EXAMPLES=(redux_f32) NVVM_VERIFY_EXAMPLES=(constant_memory cp_async_small device_global enum_constant_provenance ex2_approx_f16 generated_intrinsics generated_intrinsics_blackwell generated_ldmatrix kernel_reference_validity legacy_atomic_fadd legacy_atomic_rmw_cas libdevice_math legacy_nvvm_pointer_shapes packed_atomic_add primitive_stress scoped_atomic_load_store shuffle_64 tcgen05 tuple_constant_provenance wgmma_mma_bf16) -ERROR_EXAMPLES=(error error_set_discriminant_uninhabited error_enum_bool_payload_addr error_enum_pointer_overlap error_enum_shared_pointer_layout error_heap_alloc error_host_arch_intrinsic error_host_target_feature error_kernel_shared_param error_missing_device_attr error_generated_intrinsic_abi error_generated_intrinsic_unknown_id error_generated_intrinsic_fn_pointer error_generated_intrinsic_callable error_union_shared_pointer_addr error_union_shared_pointer_layout) +ERROR_EXAMPLES=(error error_set_discriminant_uninhabited error_enum_bool_payload_addr error_enum_pointer_overlap error_enum_shared_pointer_layout error_heap_alloc error_host_arch_intrinsic error_host_target_feature error_kernel_shared_param error_kernel_shared_ptr_param error_missing_device_attr error_generated_intrinsic_abi error_generated_intrinsic_unknown_id error_generated_intrinsic_fn_pointer error_generated_intrinsic_callable error_union_shared_pointer_addr error_union_shared_pointer_layout) # Per-example rustc flags policy: an example that needs a special rustc flag # (e.g. disjoint_slice_len needs -Zinline-mir=no to keep its regression @@ -702,6 +702,15 @@ verdict_error() { return 1 fi ;; + error_kernel_shared_ptr_param) + # The nested form's marker differs from the bare parameter's + # `is a pointer into shared memory`, so this fixture proves the + # exporter looked inside the parameter struct. + if ! grep -Fq 'holds a pointer into shared memory' "${log}"; then + echo "FAIL (missing nested shared-memory kernel-parameter diagnostic)" + return 1 + fi + ;; error_generated_intrinsic_abi) if ! grep -Fq 'cuda-intrinsics ABI mismatch' "${log}" \ || ! grep -Fq '__cuda_oxide_intrinsic_abi_v2::i0001' "${log}"; then From 7d79f82afa3cc7fffe61ddbd8a9c9261ee97b617 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 12/17] test(examples): make the shared_memory_helper code check discriminating The module-wide cvta.to.shared search also passed on main's build of this example. The script now also requires the .ll call site to pass a ptr addrspace(3), a PTX .func for shared_ptr_probe that the kernel calls once with mbarrier.init.shared inside its body, and no cvta.to.shared or isspacep anywhere; when the .func is missing it says to build with CUDA_OXIDE_NO_OPT=1. The example joins NO_OPT_SHAPE_EXAMPLES so the smoketest reads the unoptimized artifacts, and the README says that the no-feature build runs the typed probe and that the PTX half reads the unoptimized build. Signed-off-by: nihalpasham --- .../examples/shared_memory_helper/README.md | 25 +++++- .../shared_memory_helper/verify-code-shape.sh | 77 ++++++++++++++++--- cuda-oxide/scripts/smoketest.sh | 2 +- 3 files changed, 90 insertions(+), 14 deletions(-) diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md index 922d21e73d..a022732b05 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/README.md @@ -13,8 +13,9 @@ unsafe fn helper(smem: *mut u8, offset: usize) -> u64 { } ``` -Every build also runs the same probe through `SharedPtr`, a CTA-shared -pointer typed as such: +The no-feature build also runs the same probe through `SharedPtr`, a +CTA-shared pointer typed as such (the feature builds fail MIR verification +before any probe runs): ```rust let smem = DynamicSharedArray::::shared_ptr(); @@ -43,10 +44,26 @@ this file: | Helper | Required code shape | | --- | --- | -| `shared_ptr_probe` | first parameter is `ptr addrspace(3)` | +| `shared_ptr_probe` | first parameter is `ptr addrspace(3)`, and its call site passes one | | `write_pair` | both `Pair` fields arrive as `ptr addrspace(3)` parameters | | `select_slot` | return type is `ptr addrspace(3)` | -| whole module | no `cvta.to.shared` in the PTX | +| `shared_ptr_probe` in PTX | a `.func` that the kernel calls once, with `mbarrier.init.shared` inside its body | +| whole module | no `cvta.to.shared` and no `isspacep` in the PTX | + +The PTX half reads the unoptimized build. `scripts/smoketest.sh` lists this +example in `NO_OPT_SHAPE_EXAMPLES`, so after the optimized build decides the +verdict it rebuilds once with `CUDA_OXIDE_NO_OPT=1` and runs the script on +those artifacts. To run the script by hand, build the same way first: + +```bash +CUDA_OXIDE_NO_OPT=1 cargo oxide build shared_memory_helper +bash crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh +``` + +Without the surviving `.func`, a module-wide search for `cvta.to.shared` +proves little: an optimized build inlines the helper, and the generic-pointer +control compiles to the same instructions. The call into the `.func` is where +a generic pointer would have needed a conversion. The launch supplies 2048 dynamic shared bytes. Four independent CTAs each use one thread. Nine launches cover three data offsets and three input values, diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh index e12ab06047..5bc0dc0612 100755 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_memory_helper/verify-code-shape.sh @@ -2,11 +2,23 @@ # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: Apache-2.0 # -# `SharedPtr` keeps a CTA-shared pointer typed `addrspace(3)` across a call: -# the non-inlined `shared_ptr_probe` helper takes one as its parameter, the -# two fields of a by-value `Pair` reach `write_pair` as two such parameters, -# `select_slot` returns one, and no PTX converts a generic address to a -# shared one (`cvta.to.shared`). +# `SharedPtr` keeps a CTA-shared pointer typed `addrspace(3)` across a call. +# +# LLVM IR: the non-inlined `shared_ptr_probe` helper takes one as its +# parameter and its call site passes one, the two fields of a by-value `Pair` +# reach `write_pair` as two such parameters, and `select_slot` returns one. +# +# PTX: `shared_ptr_probe` survives as a `.func` that the kernel calls once, +# and that `.func` initializes the barrier with `mbarrier.init.shared` on its +# parameter. No instruction converts a generic address to a shared one +# (`cvta.to.shared`) or asks which window an address is in (`isspacep`). The +# call into the `.func` is where a generic pointer would have needed one; a +# module-wide search alone proves little once the helper is inlined, because +# the generic-pointer control compiles to the same instructions. +# +# The PTX half reads the unoptimized build. scripts/smoketest.sh lists this +# example in NO_OPT_SHAPE_EXAMPLES and rebuilds it with CUDA_OXIDE_NO_OPT=1 +# before running this script; by hand, build the same way first. set -euo pipefail dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" @@ -17,12 +29,20 @@ for file in "${ll}" "${ptx}"; do [[ -f "${file}" ]] || { echo "FAIL: ${file} not found; build the example first" >&2; exit 1; } done +# --- LLVM IR ----------------------------------------------------------------- + if ! grep -Eq '^define .*shared_ptr_probe.*\(ptr addrspace\(3\) ' "${ll}"; then echo "FAIL: shared_ptr_probe does not take a ptr addrspace(3) parameter" >&2 grep -E '^define .*shared_ptr_probe' "${ll}" >&2 || true exit 1 fi +if ! grep -Eq 'call i64 @[^(]*shared_ptr_probe[^(]*\(ptr addrspace\(3\) ' "${ll}"; then + echo "FAIL: the call to shared_ptr_probe does not pass a ptr addrspace(3)" >&2 + grep -E 'call i64 @[^(]*shared_ptr_probe' "${ll}" >&2 || true + exit 1 +fi + if ! grep -Eq '^define .*write_pair.*\(ptr addrspace\(3\) %[A-Za-z0-9_.]+, ptr addrspace\(3\) %[A-Za-z0-9_.]+, ' "${ll}"; then echo "FAIL: write_pair does not take both Pair fields as ptr addrspace(3) parameters" >&2 grep -E '^define .*write_pair' "${ll}" >&2 || true @@ -35,10 +55,49 @@ if ! grep -Eq '^define ptr addrspace\(3\) @[^(]*select_slot' "${ll}"; then exit 1 fi -if grep -q 'cvta\.to\.shared' "${ptx}"; then - echo "FAIL: PTX converts a generic address to a shared one" >&2 - grep -n 'cvta\.to\.shared' "${ptx}" >&2 +# --- PTX --------------------------------------------------------------------- + +# The body of the PTX `.func` whose name contains the symbol: from the header +# line, past the parameter list, to the closing brace. A forward declaration +# ends in `;` before any `{` and is skipped. +func_body() { + awk -v needle="$1" ' + !emit && !candidate && index($0, ".func") && index($0, needle) { candidate = 1 } + candidate && $0 ~ /^[[:space:]]*;[[:space:]]*$/ { candidate = 0; next } + candidate && $0 ~ /^[[:space:]]*\{/ { emit = 1; candidate = 0 } + emit { print } + emit && $0 ~ /^\}/ { exit } + ' "${ptx}" +} + +probe_body="$(func_body shared_ptr_probe)" +if [[ -z "${probe_body}" ]]; then + echo "FAIL: no PTX .func for shared_ptr_probe in ${ptx}" >&2 + echo "hint: this half reads the unoptimized build; rebuild with" >&2 + echo " CUDA_OXIDE_NO_OPT=1 cargo oxide build shared_memory_helper" >&2 exit 1 fi -echo "PASS: SharedPtr stays addrspace(3) across shared_ptr_probe, write_pair and select_slot, no cvta.to.shared" +if ! grep -Eq 'mbarrier\.init\.shared(::cta)?\.b64' <<<"${probe_body}"; then + echo "FAIL: the shared_ptr_probe .func does not initialize the barrier with mbarrier.init.shared" >&2 + exit 1 +fi + +# One call into the helper: the symbol on a `call` line, outside `.func` +# headers and comments. +calls="$(grep -E 'shared_ptr_probe' "${ptx}" | grep -vE '\.func|^[[:space:]]*//' | grep -cE '(^|[[:space:]])call' || true)" +if [[ "${calls}" -ne 1 ]]; then + echo "FAIL: expected one PTX call into shared_ptr_probe, found ${calls}" >&2 + grep -nE 'call.*shared_ptr_probe' "${ptx}" >&2 || true + exit 1 +fi + +for forbidden in 'cvta\.to\.shared' 'isspacep'; do + if grep -Eq "${forbidden}" "${ptx}"; then + echo "FAIL: PTX contains ${forbidden}: a generic address was converted or classified" >&2 + grep -nE "${forbidden}" "${ptx}" >&2 + exit 1 + fi +done + +echo "PASS: SharedPtr stays addrspace(3) across shared_ptr_probe, write_pair and select_slot; the shared_ptr_probe .func is called once and initializes its barrier in shared memory; no cvta.to.shared or isspacep" diff --git a/cuda-oxide/scripts/smoketest.sh b/cuda-oxide/scripts/smoketest.sh index fa442336ff..05b2ce6d51 100755 --- a/cuda-oxide/scripts/smoketest.sh +++ b/cuda-oxide/scripts/smoketest.sh @@ -125,7 +125,7 @@ NO_LAUNCH_EXAMPLES=(wgmma_mma_bf16) # emitted before opt folds and reorders it. The optimized build still runs and # still decides the verdict; these get one extra CUDA_OXIDE_NO_OPT=1 build # afterwards purely to produce the artifacts the shape check reads. -NO_OPT_SHAPE_EXAMPLES=(const_bool_dead_branch) +NO_OPT_SHAPE_EXAMPLES=(const_bool_dead_branch shared_memory_helper) # This example intentionally configures a cubin artifact in Cargo metadata. # Its supported project route necessarily includes NVVM IR + libNVVM + From e7d7d47f1fac8506c0a701b8f2431a5f4f5fd9eb Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 13/17] test(examples): add shared_ptr_api covering SharedPtr across helper boundaries Eleven kernels move a SharedPtr through the boundaries the review's scratch batteries covered: a non-inlined helper returning one, &Region and &mut SharedPtr, a closure through a generic apply, [SharedPtr; 2] by value, Option>, a SharedArray of SharedPtrs in shared memory, transmute to usize and back, a from_raw round trip with a negative offset, addr_u32 against cvta_generic_to_shared_u32, a struct pointee with a cast to its first field, and f64 and u16 byte views. The host checks every value and prints one SUCCESS line. Listed in NVVM_VERIFY_EXAMPLES so the compile-only lane also runs it through libNVVM. Structs that mix a SharedPtr with a narrower field stay out: on the p3:32 route they are not yet supported in arrays or shared memory. Signed-off-by: nihalpasham --- .../examples/shared_ptr_api/Cargo.lock | 753 ++++++++++++++++++ .../examples/shared_ptr_api/Cargo.toml | 23 + .../examples/shared_ptr_api/README.md | 74 ++ .../examples/shared_ptr_api/src/main.rs | 535 +++++++++++++ cuda-oxide/scripts/smoketest.sh | 2 +- 5 files changed, 1386 insertions(+), 1 deletion(-) create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.lock create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.toml create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/README.md create mode 100644 cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/src/main.rs diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.lock b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.lock new file mode 100644 index 0000000000..0df03ea524 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.lock @@ -0,0 +1,753 @@ +# This file is automatically @generated by Cargo. +# It is not intended for manual editing. +version = 4 + +[[package]] +name = "aho-corasick" +version = "1.1.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301" +dependencies = [ + "memchr", +] + +[[package]] +name = "anyhow" +version = "1.0.102" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c" + +[[package]] +name = "bindgen" +version = "0.69.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "271383c67ccabffb7381723dea0672a673f292304fcb45c01cc648c7a8d58088" +dependencies = [ + "bitflags", + "cexpr", + "clang-sys", + "itertools", + "lazy_static", + "lazycell", + "log", + "prettyplease", + "proc-macro2", + "quote", + "regex", + "rustc-hash", + "shlex", + "syn", + "which", +] + +[[package]] +name = "bitflags" +version = "2.13.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8" + +[[package]] +name = "block-buffer" +version = "0.10.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71" +dependencies = [ + "generic-array", +] + +[[package]] +name = "cexpr" +version = "0.6.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6fac387a98bb7c37292057cffc56d62ecb629900026402633ae9160df93a8766" +dependencies = [ + "nom", +] + +[[package]] +name = "cfg-if" +version = "1.0.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801" + +[[package]] +name = "clang-sys" +version = "1.8.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0b023947811758c97c59bf9d1c188fd619ad4718dcaa767947df1cadb14f39f4" +dependencies = [ + "glob", + "libc", + "libloading", +] + +[[package]] +name = "cpufeatures" +version = "0.2.17" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280" +dependencies = [ + "libc", +] + +[[package]] +name = "crc32fast" +version = "1.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9481c1c90cbf2ac953f07c8d4a58aa3945c425b7185c9154d67a65e4230da511" +dependencies = [ + "cfg-if", +] + +[[package]] +name = "crunchy" +version = "0.2.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5" + +[[package]] +name = "crypto-common" +version = "0.1.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a" +dependencies = [ + "generic-array", + "typenum", +] + +[[package]] +name = "cuda-artifact-finalizer" +version = "0.2.1" +dependencies = [ + "libnvvm-sys", + "nvjitlink-sys", + "serde", + "sha2", + "thiserror", +] + +[[package]] +name = "cuda-bindings" +version = "0.4.0" +dependencies = [ + "bindgen", + "libloading", + "prettyplease", + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "cuda-core" +version = "0.4.0" +dependencies = [ + "anyhow", + "cuda-bindings", + "cuda-core-derive", + "half", + "oxide-artifacts", +] + +[[package]] +name = "cuda-core-derive" +version = "0.4.0" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "cuda-device" +version = "0.2.1" +dependencies = [ + "cuda-macros", +] + +[[package]] +name = "cuda-host" +version = "0.2.1" +dependencies = [ + "cuda-artifact-finalizer", + "cuda-core", + "cuda-macros", + "half", + "libc", + "oxide-artifacts", + "ptx-parse", + "sha2", + "thiserror", +] + +[[package]] +name = "cuda-macros" +version = "0.2.1" +dependencies = [ + "oxide-tile-export", + "proc-macro2", + "quote", + "reserved-oxide-symbols", + "syn", +] + +[[package]] +name = "cuda-target-spec" +version = "0.1.0" + +[[package]] +name = "digest" +version = "0.10.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292" +dependencies = [ + "block-buffer", + "crypto-common", +] + +[[package]] +name = "either" +version = "1.16.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e" + +[[package]] +name = "equivalent" +version = "1.0.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f" + +[[package]] +name = "errno" +version = "0.3.14" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb" +dependencies = [ + "libc", + "windows-sys 0.61.2", +] + +[[package]] +name = "foldhash" +version = "0.1.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2" + +[[package]] +name = "generic-array" +version = "0.14.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a" +dependencies = [ + "typenum", + "version_check", +] + +[[package]] +name = "glob" +version = "0.3.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0cc23270f6e1808e30a928bdc84dea0b9b4136a8bc82338574f23baf47bbd280" + +[[package]] +name = "half" +version = "2.7.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b" +dependencies = [ + "cfg-if", + "crunchy", + "zerocopy", +] + +[[package]] +name = "hashbrown" +version = "0.15.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1" +dependencies = [ + "foldhash", +] + +[[package]] +name = "hashbrown" +version = "0.17.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a" + +[[package]] +name = "home" +version = "0.5.12" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "cc627f471c528ff0c4a49e1d5e60450c8f6461dd6d10ba9dcd3a61d3dff7728d" +dependencies = [ + "windows-sys 0.61.2", +] + +[[package]] +name = "indexmap" +version = "2.14.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9" +dependencies = [ + "equivalent", + "hashbrown 0.17.1", +] + +[[package]] +name = "itertools" +version = "0.12.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ba291022dbbd398a455acf126c1e341954079855bc60dfdda641363bd6922569" +dependencies = [ + "either", +] + +[[package]] +name = "itoa" +version = "1.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682" + +[[package]] +name = "lazy_static" +version = "1.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe" + +[[package]] +name = "lazycell" +version = "1.3.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "830d08ce1d1d941e6b30645f1a0eb5643013d835ce3779a5fc208261dbe10f55" + +[[package]] +name = "libc" +version = "0.2.186" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66" + +[[package]] +name = "libloading" +version = "0.8.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d7c4b02199fee7c5d21a5ae7d8cfa79a6ef5bb2fc834d6e9058e89c825efdc55" +dependencies = [ + "cfg-if", + "windows-link", +] + +[[package]] +name = "libnvvm-sys" +version = "0.2.1" +dependencies = [ + "cuda-target-spec", + "libloading", + "thiserror", +] + +[[package]] +name = "linux-raw-sys" +version = "0.4.15" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d26c52dbd32dccf2d10cac7725f8eae5296885fb5703b261f7d0a0739ec807ab" + +[[package]] +name = "log" +version = "0.4.29" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897" + +[[package]] +name = "memchr" +version = "2.8.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98" + +[[package]] +name = "minimal-lexical" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "68354c5c6bd36d73ff3feceb05efa59b6acb7626617f4962be322a825e61f79a" + +[[package]] +name = "nom" +version = "7.1.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d273983c5a657a70a3e8f2a01329822f3b8c8172b73826411a55751e404a0a4a" +dependencies = [ + "memchr", + "minimal-lexical", +] + +[[package]] +name = "nvjitlink-sys" +version = "0.2.1" +dependencies = [ + "libloading", + "thiserror", +] + +[[package]] +name = "object" +version = "0.36.7" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "62948e14d923ea95ea2c7c86c71013138b66525b86bdc08d2dcc262bdb497b87" +dependencies = [ + "crc32fast", + "hashbrown 0.15.5", + "indexmap", + "memchr", +] + +[[package]] +name = "once_cell" +version = "1.21.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50" + +[[package]] +name = "oxide-artifacts" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "74d9a941c0c43d2dceaf25f5c9ca4c9f9cceb28dfa5e4e3cfd14e476f003e7c9" +dependencies = [ + "object", +] + +[[package]] +name = "oxide-tile-export" +version = "0.2.1" +dependencies = [ + "serde", + "serde_json", + "thiserror", +] + +[[package]] +name = "prettyplease" +version = "0.2.37" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b" +dependencies = [ + "proc-macro2", + "syn", +] + +[[package]] +name = "proc-macro2" +version = "1.0.106" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934" +dependencies = [ + "unicode-ident", +] + +[[package]] +name = "ptx-parse" +version = "0.1.0" + +[[package]] +name = "quote" +version = "1.0.45" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924" +dependencies = [ + "proc-macro2", +] + +[[package]] +name = "regex" +version = "1.12.3" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276" +dependencies = [ + "aho-corasick", + "memchr", + "regex-automata", + "regex-syntax", +] + +[[package]] +name = "regex-automata" +version = "0.4.14" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f" +dependencies = [ + "aho-corasick", + "memchr", + "regex-syntax", +] + +[[package]] +name = "regex-syntax" +version = "0.8.11" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4" + +[[package]] +name = "reserved-oxide-symbols" +version = "0.2.1" + +[[package]] +name = "rustc-hash" +version = "1.1.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2" + +[[package]] +name = "rustix" +version = "0.38.44" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "fdb5bc1ae2baa591800df16c9ca78619bf65c0488b41b96ccec5d11220d8c154" +dependencies = [ + "bitflags", + "errno", + "libc", + "linux-raw-sys", + "windows-sys 0.59.0", +] + +[[package]] +name = "serde" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e" +dependencies = [ + "serde_core", + "serde_derive", +] + +[[package]] +name = "serde_core" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad" +dependencies = [ + "serde_derive", +] + +[[package]] +name = "serde_derive" +version = "1.0.228" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "serde_json" +version = "1.0.154" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e7e9cc8b1b85264074fbcc02a88680c4096b1e47df8f739dceb03bf482f04bd6" +dependencies = [ + "itoa", + "memchr", + "serde", + "serde_core", + "zmij", +] + +[[package]] +name = "sha2" +version = "0.10.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283" +dependencies = [ + "cfg-if", + "cpufeatures", + "digest", +] + +[[package]] +name = "shared_ptr_api" +version = "0.1.0" +dependencies = [ + "cuda-core", + "cuda-device", + "cuda-host", +] + +[[package]] +name = "shlex" +version = "1.3.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0fda2ff0d084019ba4d7c6f371c95d8fd75ce3524c3cb8fb653a3023f6323e64" + +[[package]] +name = "syn" +version = "2.0.117" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99" +dependencies = [ + "proc-macro2", + "quote", + "unicode-ident", +] + +[[package]] +name = "thiserror" +version = "2.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4" +dependencies = [ + "thiserror-impl", +] + +[[package]] +name = "thiserror-impl" +version = "2.0.18" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "typenum" +version = "1.20.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20" + +[[package]] +name = "unicode-ident" +version = "1.0.24" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75" + +[[package]] +name = "version_check" +version = "0.9.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a" + +[[package]] +name = "which" +version = "4.4.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "87ba24419a2078cd2b0f2ede2691b6c66d8e47836da3b6db8265ebad47afbfc7" +dependencies = [ + "either", + "home", + "once_cell", + "rustix", +] + +[[package]] +name = "windows-link" +version = "0.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5" + +[[package]] +name = "windows-sys" +version = "0.59.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1e38bc4d79ed67fd075bcc251a1c39b32a1776bbe92e5bef1f0bf1f8c531853b" +dependencies = [ + "windows-targets", +] + +[[package]] +name = "windows-sys" +version = "0.61.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc" +dependencies = [ + "windows-link", +] + +[[package]] +name = "windows-targets" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973" +dependencies = [ + "windows_aarch64_gnullvm", + "windows_aarch64_msvc", + "windows_i686_gnu", + "windows_i686_gnullvm", + "windows_i686_msvc", + "windows_x86_64_gnu", + "windows_x86_64_gnullvm", + "windows_x86_64_msvc", +] + +[[package]] +name = "windows_aarch64_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3" + +[[package]] +name = "windows_aarch64_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469" + +[[package]] +name = "windows_i686_gnu" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b" + +[[package]] +name = "windows_i686_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66" + +[[package]] +name = "windows_i686_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66" + +[[package]] +name = "windows_x86_64_gnu" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78" + +[[package]] +name = "windows_x86_64_gnullvm" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d" + +[[package]] +name = "windows_x86_64_msvc" +version = "0.52.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec" + +[[package]] +name = "zerocopy" +version = "0.8.52" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ce1022995ff5ff5d841ad7d994facc23098cd40152f2c1d11cd607c6f530653f" +dependencies = [ + "zerocopy-derive", +] + +[[package]] +name = "zerocopy-derive" +version = "0.8.52" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1ae7f38b72ec2a254e2b87ef277cf2cd4fb97cbebf944faa6f33354da0867930" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + +[[package]] +name = "zmij" +version = "1.0.23" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b" diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.toml b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.toml new file mode 100644 index 0000000000..f7fd6f703c --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/Cargo.toml @@ -0,0 +1,23 @@ +[package] +name = "shared_ptr_api" +version = "0.1.0" +edition = "2024" +authors = ["NVIDIA Corporation"] +license = "Apache-2.0" +description = "SharedPtr through helper returns, references, closures, arrays, Option, shared storage and transmute" + +# Mark as standalone crate (not part of parent workspace) +[workspace] + +# Build with: +# cargo oxide run shared_ptr_api + +[dependencies] +# Device-side intrinsics (thread indexing, barriers, etc.) +cuda-device = { path = "../../../cuda-device" } + +# Host-side utilities (CudaKernel trait, etc.) +cuda-host = { path = "../../../cuda-host" } + +# Host-side CUDA runtime +cuda-core = { path = "../../../../../cuda-core" } diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/README.md b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/README.md new file mode 100644 index 0000000000..52917d7c09 --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/README.md @@ -0,0 +1,74 @@ +# shared_ptr_api + +`SharedPtr` is a raw pointer into the issuing block's shared memory whose +address space (`addrspace(3)`) is part of its type. A generic `*mut T` loses +that fact at the first helper boundary; a `SharedPtr` keeps it through +parameters, return values, struct fields, closures and arrays, so the compiler +can use the shared-memory forms of the instructions that take it. This example +moves one through each of those boundaries and checks every value on the host. + +| Kernel | Boundary | Check | +| :------------------ | :-------------------------------------------------------------- | :-------------------------------------- | +| `static_round_trip` | `SharedArray::shared_ptr`, `from_raw(as_ptr())`, negative `offset` | each thread reads back its own slot | +| `closure_capture` | a closure that captured a `SharedPtr`, called through a generic `#[inline(never)]` helper | `3 * tid` | +| `addr_u32_deltas` | `addr_u32` of `shared_ptr_at(64)`, `add(3)`, `byte_add(7)`, and of `get_raw()` through `cvta_generic_to_shared_u32` | `[64, 12, 1, 7]` | +| `behind_references` | `&Region` (a struct holding the pointer) and `&mut SharedPtr` through non-inlined helpers | three reads summed | +| `returned_pointer` | a non-inlined helper returning `SharedPtr` | odd lanes read the pool, even the array | +| `array_by_value` | `[SharedPtr; 2]` passed by value | `7 + tid` | +| `struct_pointee` | `SharedPtr` and `cast::()` to its first field | `a + b + a` | +| `byte_views` | `byte_add` into a `u8` pool, viewed as `f64` and `u16` | `0.25 * tid + tid + 7` | +| `option_pointer` | `Option>` returned and passed | `Some` stores, `None` does not | +| `stored_in_shared` | `SharedArray, 64>`: pointers kept in shared memory | the mirrored slot's value | +| `transmuted` | `transmute` to `usize` and back in non-inlined helpers | the store lands and the stride is 4 | + +The helpers carry `#[inline(never)]`, so each boundary is a real call. In the +LLVM IR the default build leaves next to this file, `smem_fill`, `smem_load`, +`pick`, `maybe` and the closure's `apply` take `ptr addrspace(3)`, +`store_via_array` takes `[2 x ptr addrspace(3)]`, `pick` and `from_bits` +return `ptr addrspace(3)`, and `SLOTS` is a `[64 x ptr addrspace(3)]` global +in address space 3. The PTX converts no generic address to a shared one +except in `from_bits`: the integer view of a shared pointer is its generic +address, as for unions, so the transmute back is one `cvta.to.shared`. + +Structs that mix a `SharedPtr` with a narrower field, such as +`{ p: SharedPtr, x: u32 }`, stay out of this example. On the libNVVM +route a shared pointer is 4 bytes wide, and such a struct is not yet +supported in arrays or in shared memory; the book's shared memory chapter +lists this limit. + +## Build and run + +```bash +cargo oxide run shared_ptr_api +``` + +The compile-only lane of `scripts/smoketest.sh` also compiles this example +through libNVVM (`NVVM_VERIFY_EXAMPLES`), where the shared pointer is 4 bytes +wide: + +```bash +cargo oxide emit-ltoir shared_ptr_api --arch=sm_120 +``` + +## Expected output + +```text +=== SharedPtr API (64 threads) === + ok static_round_trip + ok closure_capture + ok addr_u32_deltas + ok behind_references + ok returned_pointer + ok array_by_value + ok struct_pointee + ok byte_views + ok option_pointer + ok stored_in_shared + ok transmuted +SUCCESS: shared_ptr_api, 11 kernels, 0 mismatches +``` + +## Hardware requirements + +Any CUDA GPU; the kernels use plain shared memory and `sync_threads()`. The +launch supplies 1024 bytes of dynamic shared memory. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/src/main.rs b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/src/main.rs new file mode 100644 index 0000000000..12ba608b2b --- /dev/null +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/shared_ptr_api/src/main.rs @@ -0,0 +1,535 @@ +/* + * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + * SPDX-License-Identifier: Apache-2.0 + */ + +//! `SharedPtr` across every boundary a helper call can put it behind. +//! +//! `SharedPtr` is a raw pointer into the issuing block's shared memory +//! whose address space is part of its type. Each kernel here moves one +//! through a boundary where a generic `*mut T` would lose that fact, or where +//! the lowering has to rebuild the value from parts: +//! +//! - a `from_raw` round trip and a negative `offset` on a static array, +//! - a closure that captured one, called through a generic helper, +//! - `addr_u32` deltas against `cvta_generic_to_shared_u32(get_raw())`, +//! - `&Region` and `&mut SharedPtr` through non-inlined helpers, +//! - a non-inlined helper that returns one, +//! - `[SharedPtr; 2]` passed by value, +//! - a struct pointee and a `cast` to its first field, +//! - `f64` and `u16` views of a dynamic byte pool, +//! - `Option>` returned and passed, +//! - `SharedPtr` values stored in shared memory itself, +//! - `transmute` to `usize` and back. +//! +//! The host checks every value and prints one SUCCESS line. Structs that mix +//! a `SharedPtr` with a narrower field stay out of this example: on the +//! libNVVM route a shared pointer is 4 bytes wide, and such a struct is not +//! yet supported in arrays or in shared memory. +//! +//! Usage: +//! cargo oxide run shared_ptr_api + +use cuda_core::simt::LaunchConfig; +use cuda_core::{CudaContext, DeviceBuffer}; +use cuda_device::shared::cvta_generic_to_shared_u32; +use cuda_device::{DynamicSharedArray, SharedArray, SharedPtr, cuda_module, kernel, thread}; + +const THREADS: usize = 64; + +/// A pointee with two fields; a `cast` to `f32` reaches the first. +#[derive(Clone, Copy)] +#[repr(C)] +struct Pair { + a: f32, + b: f32, +} + +/// A typed shared pointer next to an element offset, read through `&Region`. +#[derive(Clone, Copy)] +struct Region { + base: SharedPtr, + offset: usize, +} + +static mut TILE_F32: SharedArray = SharedArray::UNINIT; +static mut TILE_U32: SharedArray = SharedArray::UNINIT; +static mut PAIRS: SharedArray = SharedArray::UNINIT; +static mut SLOTS: SharedArray, THREADS> = SharedArray::UNINIT; + +/// Stores through a typed shared pointer inside a non-inlined generic helper. +/// +/// # Safety +/// `p.add(i)` must be in bounds of live shared memory. +#[inline(never)] +unsafe fn smem_fill(p: SharedPtr, i: usize, v: T) { + unsafe { p.add(i).as_ptr().write(v) } +} + +/// Loads through a typed shared pointer inside a non-inlined generic helper. +/// +/// # Safety +/// `p.add(i)` must be in bounds of initialized shared memory. +#[inline(never)] +unsafe fn smem_load(p: SharedPtr, i: usize) -> T { + unsafe { p.add(i).as_ptr().read() } +} + +/// Returns one of two typed shared pointers. The return type lowers to +/// `ptr addrspace(3)`. +#[inline(never)] +fn pick(a: SharedPtr, b: SharedPtr, second: bool) -> SharedPtr { + if second { b } else { a } +} + +/// Calls a closure that captured a `SharedPtr` by value. +#[inline(never)] +fn apply(f: F, i: usize) { + f(i) +} + +/// Reads through `&Region`: a deref, a field of `Region`, then the pointer. +/// +/// # Safety +/// `r.base.add(r.offset + i)` must be in bounds of initialized shared memory. +#[inline(never)] +unsafe fn read_region(r: &Region, i: usize) -> T { + unsafe { r.base.add(r.offset + i).as_ptr().read() } +} + +/// Advances a `SharedPtr` in place through `&mut SharedPtr`. +/// +/// # Safety +/// The advanced pointer must stay inside the same allocation. +#[inline(never)] +unsafe fn bump(slot: &mut SharedPtr, by: usize) { + unsafe { *slot = slot.add(by) } +} + +/// Stores through one of two `SharedPtr`s passed in an array by value. +/// +/// # Safety +/// `ptrs[which].add(i)` must be in bounds of live shared memory. +#[inline(never)] +unsafe fn store_via_array(ptrs: [SharedPtr; 2], which: usize, i: usize, v: T) { + unsafe { ptrs[which].add(i).as_ptr().write(v) } +} + +/// Stores through an `Option>` and reports whether it did. +/// +/// # Safety +/// A `Some` pointer must be in bounds of live shared memory at `i`. +#[inline(never)] +unsafe fn store_opt(p: Option>, i: usize, v: u32) -> u32 { + match p { + Some(p) => { + unsafe { p.add(i).as_ptr().write(v) }; + 1 + } + None => 0, + } +} + +/// Wraps a `SharedPtr` in `Some` or drops it, in a non-inlined helper. +#[inline(never)] +fn maybe(p: SharedPtr, take: bool) -> Option> { + if take { Some(p) } else { None } +} + +/// A `SharedPtr` as its address bits. +#[inline(never)] +fn to_bits(p: SharedPtr) -> usize { + // SAFETY: `SharedPtr` is `repr(transparent)` over a thin pointer. + unsafe { core::mem::transmute::, usize>(p) } +} + +/// Address bits back to a `SharedPtr`. +/// +/// # Safety +/// `bits` must come from `to_bits` on a live `SharedPtr`. +#[inline(never)] +unsafe fn from_bits(bits: usize) -> SharedPtr { + unsafe { core::mem::transmute::>(bits) } +} + +#[cuda_module] +mod kernels { + use super::*; + + /// A static array through `shared_ptr`, a `from_raw` round trip, and a + /// negative `offset` back to this thread's slot. + #[kernel] + pub fn static_round_trip(out: *mut f32) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: TILE_F32 is a `static mut SharedArray` of this kernel; each + // thread writes and reads its own slot, with a barrier between. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_F32); + smem_fill(tile, tid, tid as f32 + 0.5); + thread::sync_threads(); + let again = SharedPtr::from_raw(tile.as_ptr()); + let last = again.add(THREADS - 1); + let mine = last.offset(-((THREADS - 1 - tid) as isize)); + *out.add(tid) = smem_load(mine, 0); + } + } + + /// A closure that captured a `SharedPtr`, called through a generic + /// non-inlined helper. + #[kernel] + pub fn closure_capture(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: TILE_U32 is a `static mut SharedArray` of this kernel. + let tile = unsafe { SharedArray::shared_ptr(&raw mut TILE_U32) }; + // SAFETY: each thread writes its own slot. + let store = move |i: usize| unsafe { tile.add(i).as_ptr().write(i as u32 * 3) }; + apply(store, tid); + thread::sync_threads(); + // SAFETY: the slot was written before the barrier. + unsafe { *out.add(tid) = tile.add(tid).as_ptr().read() }; + } + + /// `addr_u32` deltas within the dynamic pool, and equality with + /// `cvta_generic_to_shared_u32` on the pool's raw pointer. + #[kernel] + pub fn addr_u32_deltas(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + let base = DynamicSharedArray::::shared_ptr(); + let at64 = DynamicSharedArray::::shared_ptr_at(64); + // SAFETY: `get_raw` points into the dynamic pool; thread 0 writes the + // four output slots. + unsafe { + let raw = cvta_generic_to_shared_u32(DynamicSharedArray::::get_raw()); + if tid == 0 { + *out.add(0) = at64.addr_u32() - base.addr_u32(); + *out.add(1) = base.add(3).addr_u32() - base.addr_u32(); + *out.add(2) = (base.addr_u32() == raw) as u32; + *out.add(3) = base.cast::().byte_add(7).addr_u32() - base.addr_u32(); + } + } + } + + /// `SharedPtr` behind references: `&Region` and `&mut SharedPtr`. + #[kernel] + pub fn behind_references(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + let pool = DynamicSharedArray::::shared_ptr(); + // SAFETY: the launch supplies 1024 dynamic shared bytes, enough for + // the 64 `u32` slots; every slot is written before the barrier. + unsafe { + pool.add(tid).as_ptr().write(100 + tid as u32); + thread::sync_threads(); + let region = Region { + base: pool, + offset: 0, + }; + let a = read_region(®ion, tid); + let mut cursor = pool; + bump(&mut cursor, tid); + let b = cursor.as_ptr().read(); + let shifted = Region { + base: pool, + offset: THREADS / 2, + }; + let c = read_region(&shifted, tid % (THREADS / 2)); + *out.add(tid) = a + b + c; + } + } + + /// A non-inlined helper that returns a `SharedPtr` picks between a static + /// array and the dynamic pool. + #[kernel] + pub fn returned_pointer(out: *mut f32) { + let tid = thread::threadIdx_x() as usize; + let dynp = DynamicSharedArray::::shared_ptr(); + // SAFETY: TILE_F32 is a `static mut SharedArray` of this kernel and + // the pool holds 64 `f32`; each thread writes its own slot in both. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_F32); + smem_fill(tile, tid, 1000.0 + tid as f32); + smem_fill(dynp, tid, 2000.0 + tid as f32); + thread::sync_threads(); + let chosen = pick(tile, dynp, tid % 2 == 1); + *out.add(tid) = smem_load(chosen, tid); + } + } + + /// `[SharedPtr; 2]` passed by value to a non-inlined helper. + #[kernel] + pub fn array_by_value(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + let pool = DynamicSharedArray::::shared_ptr(); + // SAFETY: TILE_U32 is a `static mut SharedArray` of this kernel and + // the pool holds 64 `u32`; each thread writes its own slot in one. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_U32); + store_via_array([tile, pool], tid % 2, tid, 7 + tid as u32); + thread::sync_threads(); + let v = if tid.is_multiple_of(2) { + tile.add(tid).as_ptr().read() + } else { + pool.add(tid).as_ptr().read() + }; + *out.add(tid) = v; + } + } + + /// A struct pointee, and a `cast` to its first field. + #[kernel] + pub fn struct_pointee(out: *mut f32) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: PAIRS is a `static mut SharedArray` of this kernel; each + // thread writes and reads its own element. + unsafe { + let pairs = SharedArray::shared_ptr(&raw mut PAIRS); + smem_fill( + pairs, + tid, + Pair { + a: tid as f32, + b: 2.0 * tid as f32, + }, + ); + thread::sync_threads(); + let p = smem_load(pairs, tid); + let first: f32 = smem_load(pairs.add(tid).cast::(), 0); + *out.add(tid) = p.a + p.b + first; + } + } + + /// `byte_add` into a dynamic byte pool, viewed as `f64` and as `u16`. + #[kernel] + pub fn byte_views(out: *mut f64) { + let tid = thread::threadIdx_x() as usize; + let pool = DynamicSharedArray::::shared_ptr(); + // SAFETY: the pool holds 1024 bytes; the `f64` view uses the first + // 512 and the `u16` view the next 128, each thread its own slot. + unsafe { + let d = pool.byte_add(8 * tid).cast::(); + smem_fill(d, 0, tid as f64 * 0.25); + let h = pool.byte_add(8 * THREADS + 2 * tid).cast::(); + smem_fill(h, 0, tid as u16 + 7); + thread::sync_threads(); + let dv: f64 = smem_load(pool.cast::(), tid); + let hv: u16 = smem_load(pool.byte_add(8 * THREADS).cast::(), tid); + *out.add(tid) = dv + hv as f64; + } + } + + /// `Option>` returned from and passed into non-inlined + /// helpers. + #[kernel] + pub fn option_pointer(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: TILE_U32 is a `static mut SharedArray` of this kernel; each + // thread writes and reads its own slot. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_U32); + let took = store_opt(maybe(tile, true), tid, 5 + tid as u32); + let skipped = store_opt(maybe(tile, false), tid, 0); + thread::sync_threads(); + *out.add(tid) = tile.add(tid).as_ptr().read() + 10 * took + 100 * skipped; + } + } + + /// `SharedPtr` values stored in shared memory: a `SharedArray` of them. + #[kernel] + pub fn stored_in_shared(out: *mut u32) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: TILE_U32 and SLOTS are `static mut SharedArray`s of this + // kernel; each thread writes its own slot in both and reads one slot + // after the barrier. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_U32); + let slots = SharedArray::shared_ptr(&raw mut SLOTS); + tile.add(tid).as_ptr().write(tid as u32 * 11); + slots.add(tid).as_ptr().write(tile.add(THREADS - 1 - tid)); + thread::sync_threads(); + let p: SharedPtr = slots.add(tid).as_ptr().read(); + *out.add(tid) = p.as_ptr().read(); + } + } + + /// `transmute` to `usize` and back, in non-inlined helpers. The high word + /// records whether the bits of this thread's slot sit `4 * tid` bytes + /// past the bits of the array base. + #[kernel] + pub fn transmuted(out: *mut u64) { + let tid = thread::threadIdx_x() as usize; + // SAFETY: TILE_U32 is a `static mut SharedArray` of this kernel; each + // thread writes and reads its own slot. + unsafe { + let tile = SharedArray::shared_ptr(&raw mut TILE_U32); + let bits = to_bits(tile.add(tid)); + let back = from_bits(bits); + back.as_ptr().write(tid as u32 * 13); + thread::sync_threads(); + let v = tile.add(tid).as_ptr().read(); + let strided = bits == to_bits(tile) + 4 * tid; + *out.add(tid) = v as u64 | ((strided as u64) << 32); + } + } +} + +/// Counts the mismatches of one kernel and prints its line. +fn check(name: &str, bad: usize, mismatches: &mut usize, kernels: &mut usize) { + if bad == 0 { + println!(" ok {name}"); + } else { + println!(" BAD {name}: {bad} mismatches"); + } + *mismatches += bad; + *kernels += 1; +} + +fn main() -> Result<(), Box> { + let ctx = CudaContext::new(0)?; + let stream = ctx.default_stream(); + let module = kernels::load(&ctx)?; + let config = LaunchConfig { + grid_dim: (1, 1, 1), + block_dim: (THREADS as u32, 1, 1), + shared_mem_bytes: 1024, + }; + let mut mismatches = 0usize; + let mut kernels = 0usize; + println!("=== SharedPtr API ({THREADS} threads) ==="); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + // SAFETY: one block of THREADS threads; every output buffer holds one + // slot per thread (four slots for addr_u32_deltas, written by thread 0). + unsafe { module.static_round_trip(&stream, config, out.cu_deviceptr() as *mut f32)? }; + let got = out.to_host_vec(&stream)?; + check( + "static_round_trip", + (0..THREADS).filter(|&i| got[i] != i as f32 + 0.5).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.closure_capture(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + check( + "closure_capture", + (0..THREADS).filter(|&i| got[i] != i as u32 * 3).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, 4)?; + unsafe { module.addr_u32_deltas(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + let want = [64u32, 12, 1, 7]; + check( + "addr_u32_deltas", + (0..4).filter(|&i| got[i] != want[i]).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.behind_references(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + check( + "behind_references", + (0..THREADS) + .filter(|&i| { + let t = i as u32; + let half = (THREADS / 2) as u32; + got[i] != (100 + t) + (100 + t) + (100 + half + t % half) + }) + .count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.returned_pointer(&stream, config, out.cu_deviceptr() as *mut f32)? }; + let got = out.to_host_vec(&stream)?; + check( + "returned_pointer", + (0..THREADS) + .filter(|&i| { + let base = if i % 2 == 1 { 2000.0 } else { 1000.0 }; + got[i] != base + i as f32 + }) + .count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.array_by_value(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + check( + "array_by_value", + (0..THREADS).filter(|&i| got[i] != 7 + i as u32).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.struct_pointee(&stream, config, out.cu_deviceptr() as *mut f32)? }; + let got = out.to_host_vec(&stream)?; + check( + "struct_pointee", + (0..THREADS).filter(|&i| got[i] != 4.0 * i as f32).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.byte_views(&stream, config, out.cu_deviceptr() as *mut f64)? }; + let got = out.to_host_vec(&stream)?; + check( + "byte_views", + (0..THREADS) + .filter(|&i| got[i] != i as f64 * 0.25 + (i as f64 + 7.0)) + .count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.option_pointer(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + check( + "option_pointer", + (0..THREADS).filter(|&i| got[i] != 15 + i as u32).count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.stored_in_shared(&stream, config, out.cu_deviceptr() as *mut u32)? }; + let got = out.to_host_vec(&stream)?; + check( + "stored_in_shared", + (0..THREADS) + .filter(|&i| got[i] != (THREADS - 1 - i) as u32 * 11) + .count(), + &mut mismatches, + &mut kernels, + ); + + let out = DeviceBuffer::::zeroed(&stream, THREADS)?; + unsafe { module.transmuted(&stream, config, out.cu_deviceptr() as *mut u64)? }; + let got = out.to_host_vec(&stream)?; + check( + "transmuted", + (0..THREADS) + .filter(|&i| got[i] != (i as u64 * 13) | (1u64 << 32)) + .count(), + &mut mismatches, + &mut kernels, + ); + + if mismatches == 0 { + println!("SUCCESS: shared_ptr_api, {kernels} kernels, 0 mismatches"); + Ok(()) + } else { + println!("FAIL: shared_ptr_api, {mismatches} mismatches across {kernels} kernels"); + std::process::exit(1) + } +} diff --git a/cuda-oxide/scripts/smoketest.sh b/cuda-oxide/scripts/smoketest.sh index 05b2ce6d51..35d8eb57a2 100755 --- a/cuda-oxide/scripts/smoketest.sh +++ b/cuda-oxide/scripts/smoketest.sh @@ -97,7 +97,7 @@ AUTO_NVVM_EXAMPLES=(libdevice_math) IKET_EXAMPLES=(iket_trace) BLACKWELL_COMPILE_EXAMPLES=(generated_intrinsics_blackwell) SM100_COMPILE_EXAMPLES=(redux_f32) -NVVM_VERIFY_EXAMPLES=(constant_memory cp_async_small device_global enum_constant_provenance ex2_approx_f16 generated_intrinsics generated_intrinsics_blackwell generated_ldmatrix kernel_reference_validity legacy_atomic_fadd legacy_atomic_rmw_cas libdevice_math legacy_nvvm_pointer_shapes packed_atomic_add primitive_stress scoped_atomic_load_store shuffle_64 tcgen05 tuple_constant_provenance wgmma_mma_bf16) +NVVM_VERIFY_EXAMPLES=(constant_memory cp_async_small device_global enum_constant_provenance ex2_approx_f16 generated_intrinsics generated_intrinsics_blackwell generated_ldmatrix kernel_reference_validity legacy_atomic_fadd legacy_atomic_rmw_cas libdevice_math legacy_nvvm_pointer_shapes packed_atomic_add primitive_stress scoped_atomic_load_store shared_ptr_api shuffle_64 tcgen05 tuple_constant_provenance wgmma_mma_bf16) ERROR_EXAMPLES=(error error_set_discriminant_uninhabited error_enum_bool_payload_addr error_enum_pointer_overlap error_enum_shared_pointer_layout error_heap_alloc error_host_arch_intrinsic error_host_target_feature error_kernel_shared_param error_kernel_shared_ptr_param error_missing_device_attr error_generated_intrinsic_abi error_generated_intrinsic_unknown_id error_generated_intrinsic_fn_pointer error_generated_intrinsic_callable error_union_shared_pointer_addr error_union_shared_pointer_layout) # Per-example rustc flags policy: an example that needs a special rustc flag From 4aa492311cff4d0ce8c35ff6000648c13b051c01 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 14/17] docs(book): state the SharedPtr limits and make the snippet compilable The snippet imports thread and kernel, lives in a kernel, uses the dynamic pool pointer it declares, and wraps the unsafe constructor. A paragraph after the table lists the limits: no SharedPtr kernel parameter, bare or inside a struct or array; unions and MaybeUninit fail closed under the union storage rules; repr(packed) carriers follow the packed limits; and on the libNVVM route a struct that mixes a SharedPtr with a narrower field is not yet supported in arrays or shared memory, a gap that predates SharedPtr. The appendix carries the same note next to the union limits. Signed-off-by: nihalpasham --- .../shared-memory-and-synchronization.md | 37 +++++++++++++++---- .../appendix/supported-features.md | 7 ++++ 2 files changed, 36 insertions(+), 8 deletions(-) diff --git a/cuda-oxide/cuda-oxide-book/advanced/shared-memory-and-synchronization.md b/cuda-oxide/cuda-oxide-book/advanced/shared-memory-and-synchronization.md index 9d93002d37..5871909e1a 100644 --- a/cuda-oxide/cuda-oxide-book/advanced/shared-memory-and-synchronization.md +++ b/cuda-oxide/cuda-oxide-book/advanced/shared-memory-and-synchronization.md @@ -285,7 +285,7 @@ return values, struct fields and generic code, so instructions that take it use their shared-memory forms without converting a generic address. ```rust -use cuda_device::{DynamicSharedArray, SharedArray, SharedPtr}; +use cuda_device::{DynamicSharedArray, SharedArray, SharedPtr, kernel, thread}; #[inline(never)] unsafe fn fill(tile: SharedPtr, i: usize, value: f32) { @@ -294,23 +294,44 @@ unsafe fn fill(tile: SharedPtr, i: usize, value: f32) { } static mut TILE: SharedArray = SharedArray::UNINIT; -let tile: SharedPtr = SharedArray::shared_ptr(&raw mut TILE); -let pool: SharedPtr = DynamicSharedArray::::shared_ptr_at(1024); -unsafe { fill(tile, thread::threadIdx_x() as usize, 1.0) }; + +#[kernel] +pub fn fill_tile() { + // SAFETY: TILE is a `static mut SharedArray` of this kernel. + let tile: SharedPtr = unsafe { SharedArray::shared_ptr(&raw mut TILE) }; + // The same type for dynamic shared memory, 1024 bytes into the pool. + let pool: SharedPtr = DynamicSharedArray::::shared_ptr_at(1024); + // SAFETY: the launch supplies more than 1024 dynamic shared bytes, and + // each thread writes its own slot of TILE. + unsafe { + fill(tile, thread::threadIdx_x() as usize, 1.0); + pool.as_ptr().write(0); + } +} ``` | Operation | Result | |:---------------------------------------|:----------------------------------------------------| -| `SharedArray::shared_ptr(&raw mut S)` | `SharedPtr` to the first element of `S` | +| `unsafe SharedArray::shared_ptr(&raw mut S)` | `SharedPtr` to the first element of `S`, a `static mut` of this kernel | | `DynamicSharedArray::::shared_ptr()` / `shared_ptr_at(bytes)` | Like `get()` / `offset(bytes)` | | `unsafe SharedPtr::from_raw(p)` | Asserts that generic `p` points into this block's shared memory | | `add`, `offset`, `byte_add`, `cast` | Pointer arithmetic that keeps the type | | `as_ptr()` | The generic `*mut T`, for loads, stores and raw-pointer APIs | | `addr_u32()` | The 32-bit `.shared::cta` address, for `_shared_u32` forms | -`SharedPtr` cannot be a kernel parameter: the host has no shared-memory -address to pass. Pointers from `cluster::map_shared_rank` may name another -block's shared memory and are not `SharedPtr`s. +`SharedPtr` cannot be a kernel parameter, bare or inside a struct or array: +the host has no shared-memory address to pass, and the compiler refuses the +signature. Pointers from `cluster::map_shared_rank` may name another block's +shared memory and are not `SharedPtr`s. + +Inside aggregates the usual shared-pointer limits apply. A union or a +`MaybeUninit>` fails closed under the union storage rules, and a +`repr(packed)` struct that holds one follows the packed-carrier limits; both +are listed in [Supported Features](../appendix/supported-features.md). On the +libNVVM route a shared pointer is 4 bytes wide, so a struct that mixes a +`SharedPtr` with a narrower field, such as `{ p: SharedPtr, x: u32 }`, is +not yet supported in arrays or in shared memory. That gap predates +`SharedPtr` and applies to `*mut Barrier` fields in the same way. --- diff --git a/cuda-oxide/cuda-oxide-book/appendix/supported-features.md b/cuda-oxide/cuda-oxide-book/appendix/supported-features.md index 9cc4d6774d..f568257006 100644 --- a/cuda-oxide/cuda-oxide-book/appendix/supported-features.md +++ b/cuda-oxide/cuda-oxide-book/appendix/supported-features.md @@ -117,6 +117,13 @@ do not count toward that bound. Taking the in-place address of such a field (`u.ptr = p`, `&u.ptr`, `&mut u.ptr`) is refused with a diagnostic; build the union by value and read its fields by value. +`SharedPtr`, the typed shared pointer, follows the same limits: a union or a +`MaybeUninit>` fails closed under these union storage rules, a +`repr(packed)` struct that holds one follows the packed-carrier limits above, a +kernel parameter that holds one bare or inside a struct or array is refused, +and on the libNVVM route a struct that mixes one with a narrower field is not +yet supported in arrays or in shared memory. + Enum constants preserve payload relocations to device statics, including non-zero byte addends. This includes niche-encoded `Option<&T>` and direct-tagged enum layouts, both for direct thin-reference payloads and for From 5c3f1680b21fb69df7b5d281098a61ace9637043 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:10:19 +0530 Subject: [PATCH 15/17] docs(mir-importer): say why ThreadIndex needs no projection strip The strip exists because SharedPtr's scalar is itself a pointer, which the iterative place walker takes as an address to load from. An integer scalar newtype such as ThreadIndex already has identity extract_field arms in dialect-mir and mir-lower, so it needs none. The follow-up is to have the walker track address versus value and then drop the strip. Signed-off-by: nihalpasham --- .../mir-importer/src/translator/rvalue/place_read.rs | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/cuda-oxide/crates/mir-importer/src/translator/rvalue/place_read.rs b/cuda-oxide/crates/mir-importer/src/translator/rvalue/place_read.rs index 150c0493bb..dc3f96e9d1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/rvalue/place_read.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/rvalue/place_read.rs @@ -117,6 +117,14 @@ pub fn translate_place( /// Projecting that field is therefore the identity. Without this, the place /// walkers would read the pointer value of a `SharedPtr` nested in another /// place, such as `(region.base).ptr`, as the address of an aggregate. +/// +/// `ThreadIndex`, the other scalar-lowered newtype, needs no strip. Its scalar +/// is an integer, and `dialect-mir` and `mir-lower` already treat +/// `extract_field` on an integer scalar as the identity. The iterative place +/// walker in `place_iter.rs` takes any pointer-typed current value as an +/// address to load from, so only a newtype whose scalar is itself a pointer is +/// misread. Follow-up: have that walker track whether its current value is an +/// address or a value, then drop this strip. pub(crate) fn strip_shared_ptr_field_projections<'a>( body: &mir::Body, place: &'a mir::Place, From fc7e160df43ac94636819399122803648ae09c4a Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:36:24 +0530 Subject: [PATCH 16/17] docs(book): show the CTA-local TMA form next to the cluster form A one-page figure for the book and the pull request, in the style of the tile-libraries direct-linkage figure. Signed-off-by: nihalpasham --- .../advanced/images/tma-cta-local.svg | 139 ++++++++++++++++++ .../advanced/tensor-memory-accelerator.md | 11 ++ 2 files changed, 150 insertions(+) create mode 100644 cuda-oxide/cuda-oxide-book/advanced/images/tma-cta-local.svg diff --git a/cuda-oxide/cuda-oxide-book/advanced/images/tma-cta-local.svg b/cuda-oxide/cuda-oxide-book/advanced/images/tma-cta-local.svg new file mode 100644 index 0000000000..acbd534d5a --- /dev/null +++ b/cuda-oxide/cuda-oxide-book/advanced/images/tma-cta-local.svg @@ -0,0 +1,139 @@ + + + + + +CTA-local TMA loads: a shared pointer that is shared by type, and the .shared::cta copy form +The cluster form converts a generic *mut u8 to the cluster window; on sm_120a ptxas adds a runtime CTA check whose call drops setmaxnreg. The CTA form takes a SharedPtr<u8>. +TODAY (cp_async_bulk_tensor_2d_g2s, generic *mut u8) +THE TYPE (SharedPtr<T>, from #1341) +NEW (cp_async_bulk_tensor_2d_g2s_cta, SharedPtr<u8>) + +tma_copy_2d_test: the cluster form +static mut TILE: SharedArray<f32, TILE_SIZE, 128> + = SharedArray::UNINIT; +static mut BAR: Barrier = Barrier::UNINIT; + +// dst: a generic *mut u8, any CTA of the cluster +cp_async_bulk_tensor_2d_g2s( + &raw mut TILE as *mut u8, + tensor_map, tile_x, tile_y, + &raw mut BAR, +); +Unchanged by this PR: 39 in-tree callers keep this form and +its cluster contract; the PTX is byte for byte the same. + +tma_copy.ll: the generic pointer is converted +%v21 = addrspacecast ptr addrspace(3) @TILE to ptr +%v22 = addrspacecast ptr %v21 to ptr addrspace(7) +call void @llvm.nvvm.cp.async.bulk.tensor.g2s.tile.2d( + ptr addrspace(7) %v22, ptr addrspace(3) @BAR, ptr %v7, + i32 %v9, i32 %v10, i16 0, i64 0, i1 0, i1 0, i32 0) +The destination arrives as a generic pointer (the user's cast), +so the lowering can only name the cluster window, addrspace(7). +On sm_100a the cluster form is native. On sm_120a ptxas cannot +prove a non-multicast .shared::cluster destination local, so +it inserts a runtime CTA check with a syscall fallback, and +that call makes it ignore setmaxnreg (C7506): the #1238 kernel +ran with 24 registers instead of 168. + +SharedPtr<T>: addrspace(3) by type +let t = unsafe { SharedArray::shared_ptr(&raw mut TILE) }; +let d = DynamicSharedArray::<u8, 128>::shared_ptr_at(1024); +let p = unsafe { SharedPtr::from_raw(generic) }; // asserts +t.add(i) t.byte_add(b) t.cast::<u8>() // keep the type +t.as_ptr() // the generic *mut T for loads and stores +t.addr_u32() // the 32-bit .shared::cta address +repr(transparent) over *mut SharedSlot<T>; the importer types +that pointee as shared, next to SharedArray and Barrier, and +the newtype lowers as one scalar, like ThreadIndex. + +shared_memory_helper: across helper boundaries +struct Region<T> { base: SharedPtr<T>, offset: usize } +struct Pair { a: SharedPtr<f32>, b: SharedPtr<f32> } +#[inline(never)] +unsafe fn shared_ptr_probe<T>(region: Region<T>) -> u64 +#[inline(never)] +unsafe fn write_pair(pair: Pair, index: usize, value: f32) +#[inline(never)] +unsafe fn select_slot(pair: Pair, index: usize, second: bool) + -> SharedPtr<f32> + +the .ll: what each boundary keeps +define i64 @..shared_ptr_probe..( + ptr addrspace(3) %v0, i64 %v1) +define void @..write_pair( + ptr addrspace(3) %v0, ptr addrspace(3) %v1, i64, float) +define ptr addrspace(3) @..select_slot( + ptr addrspace(3) %v0, ptr addrspace(3) %v1, i64, i1) +Both routes; 0 cvta.to.shared in the PTX. A SharedPtr kernel +parameter, bare or in a struct or array, is a compile error. + +tma_copy_2d_cta_test: the CTA form (new) +// dst: SharedPtr<u8>, this CTA's shared memory by type +cp_async_bulk_tensor_2d_g2s_cta( + SharedArray::shared_ptr(&raw mut TILE).cast::<u8>(), + tensor_map, tile_x, tile_y, + &raw mut BAR, +); +cp_async_bulk_tensor_{1..5}d_g2s_cta, catalog records +i1034 to i1038, PTX 8.6, sm_90 and newer; tile mode only. + +LLVM NVPTX route: typed intrinsic (LLVM 22+) +call void @llvm.nvvm.cp.async.bulk.tensor.g2s.cta.tile.2d( + ptr addrspace(3) @TILE, ptr addrspace(3) @BAR, ptr %v7, + i32 %v9, i32 %v10, i64 0, i1 0) +No addrspacecast: the destination is already addrspace(3). + +libNVVM route: inline PTX, u32 addresses +asm sideeffect + "cp.async.bulk.tensor.2d.shared::cta.global.tile + .mbarrier::complete_tx::bytes [$0], [$2, {$3, $4}], [$1];", + "r,r,l,r,r,~{memory}" + +refused at compile time +a SharedPtr kernel parameter, bare or inside a struct or array +a barrier passed straight from map_shared_rank_mut (a located +error that names the cluster form to use instead) +the caller's contract: from_raw on a cluster pointer; the raw +cuda_intrinsics layer keeps *mut u8 and converts with +cvta.to.shared, and its safety text says CTA-local + + +THE PTX AND WHAT PTXAS MAKES OF IT ON sm_120a (tma_copy.ptx built for sm_90a; mnemonic column aligned; the copy line wrapped before .mbarrier; register counts from ptxas 13.4) + +cluster form: tma_copy_2d_test (unchanged) +mov.b64 %rd15, TILE; +cvta.shared.u64 %rd16, %rd15; +cvta.to.shared::cluster.u64 %rd17, %rd16; +cp.async.bulk.tensor.2d.shared::cluster.global.tile + .mbarrier::complete_tx::bytes [%rd17], [%rd2, {%r3, %r4}], [BAR]; +// ptxas sm_120a: 24 registers, 8-byte stack; with setmaxnreg: C7506 ignored +// SASS: CALL.ABS.NOINC, then UTMALDG.2D; no USETMAXREG + +CTA form: tma_copy_2d_cta_test +// no address conversion: the shared symbols are the operands +cp.async.bulk.tensor.2d.shared::cta.global.tile + .mbarrier::complete_tx::bytes [TILE], [%rd2, {%r3, %r4}], [BAR]; +// ptxas sm_120a: 10 registers, no stack; with setmaxnreg: 168 registers +// no C7506; SASS: USETMAXREG.DEALLOC.CTAPOOL 0x28, UTMALDG.2D, +// USETMAXREG.TRY_ALLOC.CTAPOOL 0xe8; no CALL +What this proves: the cluster form converts the generic destination twice (to generic, then to the +cluster window) and ptxas on sm_120a adds the runtime CTA check whose call drops setmaxnreg. The CTA +form names the shared symbols directly and keeps register reallocation. On the 5090, tma_copy copies +4096 of 4096 values through the new kernel next to the unchanged ones, through both routes, with +memcheck and synccheck clean; ranks 1 to 5 are probed and assembled, and the cluster PTX is unchanged. + diff --git a/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md b/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md index 783edaf4ed..2ddc63def0 100644 --- a/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md +++ b/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md @@ -140,6 +140,17 @@ On SM12x this form matters for performance: ptxas implements a non-multicast `.shared::cluster` TMA load it cannot prove CTA-local with a runtime check and a system call, which also makes it ignore `setmaxnreg`. +```{figure} images/tma-cta-local.svg +:align: center +:width: 100% + +The two G2S destination forms. Left, the cluster form converts a generic `*mut u8` to the cluster +window, and on sm_120a ptxas adds a runtime CTA check whose call makes it ignore `setmaxnreg`. +Middle, `SharedPtr` keeps the shared address space in its type across helper parameters, struct +fields and return values. Right, `cp_async_bulk_tensor_2d_g2s_cta` takes a `SharedPtr` and emits +the `.shared::cta` form with no address conversion, so the kernel keeps `setmaxnreg` and its 168 registers. +``` + ### Only one thread issues the copy TMA copies are **not** collective operations across the block. Exactly From fbe27a694e2cdfbb8971d87858761145d4098792 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 21:36:46 +0530 Subject: [PATCH 17/17] docs(book): wrap the CTA-form TMA snippet in unsafe SharedArray::shared_ptr is an unsafe constructor now, and the copy itself is unsafe, so the fragment shows the block and its safety argument like the shared-memory chapter's snippet does. Signed-off-by: nihalpasham --- .../advanced/tensor-memory-accelerator.md | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-) diff --git a/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md b/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md index 2ddc63def0..f2397747bf 100644 --- a/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md +++ b/cuda-oxide/cuda-oxide-book/advanced/tensor-memory-accelerator.md @@ -129,11 +129,15 @@ a CTA-shared pointer by type, so the locality holds wherever the pointer came from, including through helper functions: ```rust -cp_async_bulk_tensor_2d_g2s_cta( - SharedArray::shared_ptr(&raw mut TILE).cast::(), - tensor_map, tile_x, tile_y, - &raw mut BAR, -); +// SAFETY: TILE and BAR are `static mut` shared-memory objects of this kernel, +// and the copy completes on BAR before any thread reads TILE. +unsafe { + cp_async_bulk_tensor_2d_g2s_cta( + SharedArray::shared_ptr(&raw mut TILE).cast::(), + tensor_map, tile_x, tile_y, + &raw mut BAR, + ); +} ``` On SM12x this form matters for performance: ptxas implements a non-multicast