rtx: make minimum SM a declared knob, MDL follows it, defaulting to virtual targets - #357
Merged
jeffamstutz merged 1 commit intoAug 13, 2026
Conversation
The floor was picked twice: CMAKE_CUDA_ARCHITECTURES from a neural/CUDA-version branch (89, 75 or 52-virtual) feeding both host kernels and OptiX modules, and the MDL backend from a hardcoded "52". Nothing named the floor, so it could not be moved without editing both, and the supported GPU range stayed implicit in the branch. VISRTX_MIN_ARCH now names it and feeds all three consumers. It defaults to 50 -- OptiX's own floor for every version we build against, so 52 gave up first-gen Maxwell for nothing -- and to 75 under CUDA 13, which dropped pre-Turing codegen; an explicit floor below 75 there is a configure error rather than a silent arch shift. VISRTX_ENABLE_NEURAL raises the floor to 89 (cooperative vectors need Ada) instead of overriding one consumer of it. The MDL backend's sm_version is the highest entry its fixed set offers at or below the floor, so generated material PTX never outranks the GPU it links into. -DCMAKE_CUDA_ARCHITECTURES=<...> still overrides the host kernels (e.g. "75-real;80-real" for offline, JIT-free launches). This is a host-kernel-only knob: the OptiX module target stays pinned at compute_<floor> (a -real override would emit cubins optixModuleCreate cannot load), and the MDL backend's sm_version tracks VISRTX_MIN_ARCH, not this variable -- so all three consumers stay in agreement regardless of the host-side override.
jeffamstutz
requested changes
Aug 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The floor was picked twice: CMAKE_CUDA_ARCHITECTURES from a neural/CUDA-version branch (89, 75 or 52-virtual) feeding both host kernels and OptiX modules, and the MDL backend from a hardcoded "52". Nothing named the floor, so it could not be moved without editing both, and the supported GPU range stayed implicit in the branch.
VISRTX_MIN_ARCH now names it and feeds all three consumers. It defaults to 50 -- OptiX's own floor for every version we build against, so 52 gave up first-gen Maxwell for nothing -- and to 75 under CUDA 13, which dropped pre-Turing codegen; an explicit floor below 75 there is a configure error rather than a silent arch shift. VISRTX_ENABLE_NEURAL raises the floor to 89 (cooperative vectors need Ada) instead of overriding one consumer of it. The MDL backend's sm_version is the highest entry its fixed set offers at or below the floor, so generated material PTX never outranks the GPU it links into.
-DCMAKE_CUDA_ARCHITECTURES=<...> still overrides the host kernels (e.g. "75-real;80-real" for offline, JIT-free launches). This is a host-kernel-only knob: the OptiX module target stays pinned at compute_ (a -real override would emit cubins optixModuleCreate cannot load), and the MDL backend's sm_version tracks VISRTX_MIN_ARCH, not this variable -- so all three consumers stay in agreement regardless of the host-side override.