Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,9 @@ if(BLASTGPU_BUILD_TESTS)
add_executable(test_hydro tests/test_hydro.cpp)
target_include_directories(test_hydro PRIVATE src/eos src/hydro)
add_test(NAME hydro COMMAND test_hydro)

add_executable(m1b_p0_reference tools/m1b_p0_reference.cpp)
target_include_directories(m1b_p0_reference PRIVATE src/eos src/hydro)
endif()

# ---- M0 1D driver (no AMReX) ---------------------------------------------
Expand All @@ -45,6 +48,11 @@ if(BLASTGPU_BUILD_SOLVER)
else()
set(AMReX_GPU_BACKEND ${BLASTGPU_GPU_BACKEND} CACHE STRING "" FORCE)
if(BLASTGPU_GPU_BACKEND STREQUAL "CUDA")
if(NOT CMAKE_CUDA_ARCHITECTURES)
message(FATAL_ERROR
"CUDA builds require -DCMAKE_CUDA_ARCHITECTURES=<integer>, "
"for example 75 for an RTX 2070 SUPER")
endif()
enable_language(CUDA)
endif()
endif()
Expand All @@ -58,6 +66,20 @@ if(BLASTGPU_BUILD_SOLVER)
target_link_libraries(blastgpu PRIVATE AMReX::amrex)
if(BLASTGPU_GPU_BACKEND STREQUAL "CUDA")
set_source_files_properties(${BLASTGPU_SRC} PROPERTIES LANGUAGE CUDA)
setup_target_for_cuda_compilation(blastgpu)
set_property(TARGET blastgpu PROPERTY CUDA_ARCHITECTURES
"${CMAKE_CUDA_ARCHITECTURES}")
endif()
if(BLASTGPU_BUILD_TESTS AND BLASTGPU_GPU_BACKEND STREQUAL "NONE")
find_package(Python3 REQUIRED COMPONENTS Interpreter)
add_test(NAME m1b-u0-smoke COMMAND ${Python3_EXECUTABLE}
${CMAKE_SOURCE_DIR}/tests/m1b_p0_check.py --uniform
$<TARGET_FILE:blastgpu> ${CMAKE_SOURCE_DIR}/inputs/m1b_p0_u0.inp
"amr.n_cell=8 8 8" max_step=2 stop_time=1)
add_test(NAME m1b-sod-x-smoke COMMAND ${Python3_EXECUTABLE}
${CMAKE_SOURCE_DIR}/tests/m1b_p0_check.py
$<TARGET_FILE:blastgpu> ${CMAKE_SOURCE_DIR}/inputs/m1b_p0_sx.inp
"amr.n_cell=32 8 8" max_step=2 stop_time=1)
endif()
endif()
endif()
5 changes: 5 additions & 0 deletions docs/perf/m1b-p0-authorization.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
# M1b-P0 authorization

Authority: `BLASTGPU-M1B-P0-OWNER-AUTHORIZATION-SINGLE-LEVEL-THREE-DIMENSIONAL-MUSCL-HLLC-SSP-RK2-HYDRO-PERFORMANCE-SLICE-FROM-983CE171C0865789EACA6D878B0DD68D9A6C6022-2026-08-22`.

This branch implements only the authorized single-level Cartesian hydro slice. It does not execute multilevel hydro, reflux, regridding, remap, EB, MPI, reactive physics, tuning, or blastCFD comparison. blastCFD was not accessed.
4 changes: 4 additions & 0 deletions docs/perf/m1b-p0-build-manifest.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
build source_commit compiler flags architecture executable_sha256 build_id
cpu abcec70687e86382510fd182b493d829ee03d424 GCC-13.3 -O3 -DNDEBUG -fopenmp CPU 89f04c83ae9809f14027904b9faf57cf3f22ddd57602ad462bfb12290095fcbd d0d0f5eaabbba741b40c8a2e988b2fedbb22e091
cuda-opt abcec70687e86382510fd182b493d829ee03d424 NVCC-12.6.85 -O3 -DNDEBUG --use_fast_math -maxrregcount=255 --generate-line-info sm_75 2662fde779aa47a6cb1faf61a4696f0ae3326d8a386f464f5afa3242eb54d2d6 d6c0ec1533a6cfe89470b147bac7ab6613764f56
cuda-prof abcec70687e86382510fd182b493d829ee03d424 NVCC-12.6.85 -O3 -DNDEBUG --use_fast_math -maxrregcount=255 --generate-line-info sm_75 160142147f2e2280865cdc695d15f3e9ef82551017e7eee5690f8393a7847d8a 94e603c05d0a721f9228ab3786697a22e434128a
5 changes: 5 additions & 0 deletions docs/perf/m1b-p0-c1-authorization.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
# M1b-P0-C1 authorization

Authorization: `BLASTGPU-M1B-P0-C1-OWNER-AUTHORIZATION-CONSOLIDATE-PER-STAGE-DIRECTIONAL-FLUX-AND-CELL-UPDATE-MFITER-REGIONS-WITHOUT-CHANGING-NUMERICS-FACE-STORAGE-KERNELS-LAUNCH-DIMENSIONS-OR-REFLUX-CONTRACT-2026-08-22`.

The experiment was limited to replacing four hydro-owned `MFIter` lifetimes per RK stage with one lifetime that enqueued x flux, y flux, z flux, and the combined update for each box. No numerical formula, device kernel body, launch dimension, storage layout, stream configuration, input, tolerance, or reflux contract changed. No blastCFD work was performed.
28 changes: 28 additions & 0 deletions docs/perf/m1b-p0-c1-paired-timing.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
case pair control_hydro_s candidate_hydro_s speedup_percent
P64 1 3.868392944 3.861083600 0.189308
P64 2 3.915733783 3.839798750 1.977578
P64 3 3.893767722 3.881358292 0.319719
P64 4 3.887339670 3.880021630 0.188608
P64 5 3.886211541 3.879493847 0.173159
P64 6 3.893417659 3.886099986 0.188304
P64 7 3.892887575 3.904904430 -0.307737
P128 1 30.963360350 30.808829210 0.501581
P128 2 30.883421200 30.783218620 0.325510
P128 3 30.972175890 30.861715390 0.357921
P128 4 30.944392030 30.760813300 0.596794
P128 5 30.945798280 30.913120160 0.105710
P128 6 31.232841610 31.121937100 0.356355
P128 7 31.167226470 31.091800770 0.242590
P128 8 31.148733100 31.043899820 0.337694
P128 9 31.153316650 31.149445820 0.012427
P128 10 31.291841880 31.038931050 0.814818
P192 1 104.986156700 104.942506300 0.041595
P192 2 104.966131400 105.125216900 -0.151330
P192 3 104.901928100 104.961307400 -0.056573
P192 4 104.821090700 104.895476700 -0.070914
P192 5 104.848120900 104.880481900 -0.030855
P192 6 105.107488100 104.890759800 0.206623
P192 7 105.060223200 104.894155600 0.158319
P192 8 104.882113400 105.009893800 -0.121684
P192 9 104.894422800 104.944744100 -0.047950
P192 10 104.998942800 105.084878000 -0.081777
8 changes: 8 additions & 0 deletions docs/perf/m1b-p0-c1-raw-artifact-manifest.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
artifact external_relative_path sha256
control-nsys profiles/P128-control.nsys-rep e932891b96e6949a20d586cfda1ddb91b50051ecf9d8b464764063cce862932d
candidate-nsys profiles/P128-candidate.nsys-rep 6cc9449b82f3c7df2c01e53937793b1b576eb9c11669626d33f3d3e70edc4b54
control-flux-sass builds/control-flux-body.sass ab1c528bd6054c5c4d203a14f0bf6c4d5d1d210a41a85a305deee2ea2d2b6faa
candidate-flux-sass builds/candidate-flux-body.sass ab1c528bd6054c5c4d203a14f0bf6c4d5d1d210a41a85a305deee2ea2d2b6faa
control-update-sass builds/control-update-body.sass c6e82c9c654f5145c49131b8811662dc4d720d3397fe040a9d0a69938d25b324
candidate-update-sass builds/candidate-update-body.sass c6e82c9c654f5145c49131b8811662dc4d720d3397fe040a9d0a69938d25b324
paired-summary timing/paired-summary.txt c207f293706b4944cf4625e7907fd963d1aa1affe366049518b2b145c8dc16a2
35 changes: 35 additions & 0 deletions docs/perf/m1b-p0-c1-result.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,35 @@
# M1b-P0-C1 result

Disposition: `BLASTGPU_M1B_P0_C1_REVERT_NO_ELIGIBLE_SYNCHRONIZATION_OR_PERFORMANCE_BENEFIT`.

The candidate was numerically and architecturally sound but did not meet the frozen performance gates. It was preserved as commit `a0a77baf64b992f69f8316d412a37e6a2e4bb8c9` and explicitly reverted by `13ddca2392c995e9b113758adfbf3bdbc5a317c0`.

## Correctness and execution invariants

- CPU CTest: 4/4 passed.
- Canonical M1a: CPU and CUDA passed; all seven plotfile files were byte-identical.
- U0, SX, SY, SZ, R0-64, and R0-128 passed unchanged tolerances with zero violations.
- Candidate CUDA SX/SY/SZ/R0 TSVs were identical to the authoritative P0 TSVs.
- Compute Sanitizer memcheck, synccheck, initcheck, and racecheck passed.
- Flux and update normalized SASS were byte-identical between control and candidate.
- Kernel count, launch dimensions, storage, memory/copy topology, and reflux-compatible face components were unchanged.

## Timing adjudication

Balanced order was alternated by pair; bootstrap seed was `20260822` with 100,000 resamples.

| Case | Control median (s) | Candidate median (s) | Paired median | 95% bootstrap interval | Candidate wins | Decision |
|---|---:|---:|---:|---:|---:|---|
| P64 | 3.892888 | 3.880022 | +0.189% | +0.173% to +0.320% | 6/7 | no material regression |
| P128 | 31.060454 | 30.976026 | +0.347% | +0.216% to +0.502% | 10/10 | FAIL: below 1% |
| P192 | 104.934030 | 104.943625 | -0.052% | -0.096% to +0.064% | 3/10 | FAIL: below 1%, interval includes zero |

Candidate timing CV was 0.530% (P64), 0.480% (P128), and 0.081% (P192). Fresh control drift from the original P0 medians was +4.04%, +2.39%, and +1.65%; paired results are authoritative.

## Nsight Systems delta

Matched four-step P128 traces contained 466 kernels each, including 192 directional flux kernels and 64 combined updates. Control made 307 `cudaStreamSynchronize` calls; candidate made 187: the exact predicted reduction of 120, with no compensating synchronization family. Core launch count stayed 32 per stage for eight boxes. Candidate streams showed `flux, flux, flux, update` pipelines. Kernel time remained materially unchanged and flux work remained about 91% of kernel time, explaining why removal of host convergence delivered only a sub-percent P128 gain and no P192 benefit.

Arena maximum-used memory was unchanged: 111 MiB (P64), 837 MiB (P128), and 2807 MiB (P192). No steady `cudaMalloc`/`cudaFree`, new copy, or face-to-cell atomic appeared. Average candidate power differed by less than 1%, so no >2% energy regression was established.

The original four-region topology is retained. If recurrent synchronization is pursued again, the next bounded question is whole-level or batched multi-box launch feasibility; C1 shows that iterator consolidation alone is insufficient.
23 changes: 23 additions & 0 deletions docs/perf/m1b-p0-c1-source-audit.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
# M1b-P0-C1 source and AMReX stream audit

## Control topology

At `43dd468`, `Hydro3D::computeFluxes` used three sequential direction loops, each with its own `MFIter(border_)`; `Hydro3D::update` used a fourth `MFIter(out)`. `BlastAmr::AdvanceLevel` invoked fill, flux, update twice per step.

AMReX commit `2cf4fbcde02a700e01c62bebb4a730aeeb66c661` establishes the relevant behavior:

- `Src/Base/AMReX_MFIter.cpp:196-221` constructs the iterator with up to the configured stream count.
- `AMReX_MFIter.cpp:371-382` performs the multi-box pre-sync and assigns `currentIndex % streams`.
- `AMReX_MFIter.cpp:526-542` advances boxes and selects the next stream deterministically.
- `AMReX_MFIter.cpp:225-255` synchronizes each active stream at iterator destruction and resets stream zero.
- GPU `ParallelFor` launches use the current `Gpu::gpuStream()`; no hydro helper changes it.

For eight boxes and four streams, each control region produces one pre-sync plus four post-sync calls. One combined region therefore predicted 15 fewer `cudaStreamSynchronize` calls per stage, or 120 over the profiled eight RK stages.

## Dependency proof

The cell state, border, and three direction-nodal face `MultiFab`s share the same DistributionMapping and one FAB entry per cell BoxArray entry. `array(mfi)` resolves the same FAB index across them. Each box writes its own face-FAB allocation; geometrically duplicated nodal boundary faces are separate storage populated from consistent border ghosts. The update reads the three face FABs at that same index and writes only its cell-valid box. There is no cross-box face dependency or residual atomic; diagnostic counters are the only shared atomic target. CPU execution remains serial within each MFIter body.

The candidate commit `a0a77baf64b992f69f8316d412a37e6a2e4bb8c9` introduced `Hydro3D::advanceStage`, one `MFIter(out)` per stage, and enqueued x/y/z/update on the stream assigned to each box. It changed 13 lines and removed 20 across `src/hydro/Hydro3D.H`, `src/hydro/Hydro3D.cpp`, and `src/amr/BlastAmr.cpp`. It introduced no helper, allocation, manual stream, event, or explicit synchronization.

Static comparison proved identical flux and update SASS. The candidate trace proved repeated per-stream `flux, flux, flux, update` ordering with unchanged grid sizes (1040 flux blocks, 1024 update blocks), 256-thread blocks, and zero dynamic shared memory.
5 changes: 5 additions & 0 deletions docs/perf/m1b-p0-case-manifest.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
case grid boxes valid_cells steps stages max_grid_size blocking_factor output
P64 64x64x64 1 262144 100 200 64 8 disabled
P128 128x128x128 8 2097152 100 200 64 8 disabled
P192 192x192x192 27 7077888 100 200 64 8 disabled
P256A 256x256x256 64 16777216 0 0 64 8 allocation-only
9 changes: 9 additions & 0 deletions docs/perf/m1b-p0-correctness-results.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
# M1b-P0 correctness results

- Existing EOS and M0 hydro tests pass without changed M0 numerical interfaces.
- Canonical M1a passes on CPU and CUDA; all seven plotfile files remain byte-identical.
- U0 preserves all seven components with exactly zero spread for 100 steps on CPU and CUDA. Counters are all zero.
- SX/SY/SZ complete 537 steps to t=0.2. CPU/CUDA profile differences are at most `8.9e-15`; rotated conserved/thermodynamic differences are at most `4.5e-15`; tangential momentum is exactly zero.
- Against M0 N=256, normalized L1 errors are `8.60e-5` density, `2.22e-4` velocity, `9.04e-5` pressure, and `9.52e-5` energy.
- R0 completes 44 steps at 64 cubed and 90 at 128 cubed. Axis differences are at floating-point roundoff; CPU/CUDA relative differences are below `6.7e-15`; all counters are zero. CPU 256 cubed was not practical and was not run.
- Compute Sanitizer memcheck, synccheck, initcheck, and racecheck all report zero errors/hazards.
11 changes: 11 additions & 0 deletions docs/perf/m1b-p0-implementation.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
# M1b-P0 implementation

The additive `Cons3D` layout is `(alpha1, arho1, arho2, rhou, rhov, rhow, rhoE)`; `Prim3D` is `(alpha1, rho1, rho2, u, v, w, p)`. M0 types and overloads are unchanged. Direction-local HLLC reconstructs every primitive on demand, treats one velocity as normal and two as tangential, adds pressure only to normal momentum, and stores seven global-order fluxes plus contact speed.

`Hydro3D` persistently owns one three-ghost border MultiFab, three nodal eight-component face MultiFabs, and five device counters. `phi_old` is `U0`; `phi_new` is stage one and then the in-place final RK2 state. The paired alpha update is `-sum[(Falpha_hi-a*Sstar_hi)-(Falpha_lo-a*Sstar_lo)]/dx`. CFL is the maximum of the sum of three directional spectral radii.

Face components 0–6 are unweighted physical fluxes for future `FluxRegister` `srccomp=0`, `destcomp=0`, `numcomp=7`; component 7 is excluded. Area, timestep, and sign scaling must occur exactly once in future `CrseInit/FineAdd` calls.

CUDA builds require an explicit `CMAKE_CUDA_ARCHITECTURES`; AMReX's setup helper and the application target property propagate the same architecture. Identity mode remains the default and allocates no Hydro3D storage.

Measured limitation: the first valid implementation uses separate MFIter regions for the three flux directions and update. P128 therefore issues 32 hydro-core launches per stage and performs iterator-boundary synchronization. This was recorded without a second tuning pass.
2 changes: 2 additions & 0 deletions docs/perf/m1b-p0-kernel-results.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
interval steps stages boxes kernels flux update cfl_box boundary_fill copy streams stream_sync steady_cuda_malloc_free
P128-hydro-span 4 8 8 353 192 64 24 56 14 4 226 0
5 changes: 5 additions & 0 deletions docs/perf/m1b-p0-memory-results.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
case raw_baseline_mib raw_peak_mib raw_net_mib arena_reserved_mib arena_max_used_mib arena_used_bytes_per_cell modeled_mib modeled_bytes_per_cell
P64 648 6577 5929 5823 111 444 103.705 414.818
P128 648 6577 5929 5823 837 418.5 829.636 414.818
P192 641 6571 5930 5823 2807 416.1 2800.022 414.818
P256A 641 7487 6846 6652 6644 415.3 6637.090 414.818
10 changes: 10 additions & 0 deletions docs/perf/m1b-p0-numerical-results.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
case backend grid steps final_time primary_metric value floor_nan_inf result
U0 CPU 64x64x64 100 0.001377486203 max_component_spread 0 0/0/0 PASS
U0 CUDA 64x64x64 100 0.001377486203 max_component_spread 0 0/0/0 PASS
SX CPU+CUDA 256x16x16 537 0.2 max_cpu_cuda_profile_abs 4.06e-15 0/0/0 PASS
SY CPU+CUDA 16x256x16 537 0.2 max_cpu_cuda_profile_abs 4.06e-15 0/0/0 PASS
SZ CPU+CUDA 16x16x256 537 0.2 max_cpu_cuda_profile_abs 8.89e-15 0/0/0 PASS
SX/SY/SZ CPU+CUDA rotated 537 0.2 max_rotated_abs 4.45e-15 0/0/0 PASS
Sod-M0 CUDA 256-normal 537 0.2 max_normalized_L1 2.22e-4 0/0/0 PASS
R0 CPU+CUDA 64x64x64 44 0.0002 max_cpu_cuda_relative 1.59e-15 0/0/0 PASS
R0 CPU+CUDA 128x128x128 90 0.0002 max_cpu_cuda_relative 6.61e-15 0/0/0 PASS
13 changes: 13 additions & 0 deletions docs/perf/m1b-p0-performance-results.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
# M1b-P0 CUDA baseline

This is the untuned first valid hydro implementation on the RTX 2070 SUPER. It is not a blastCFD comparison.

| Case | Boxes | Median hydro / 100 steps | CV | Cell-stage/s | Face eval/s | Arena max used |
|---|---:|---:|---:|---:|---:|---:|
| P64 | 1 | 3.7418 s | 0.198% | 14.01 M | 42.69 M | 111 MiB |
| P128 | 8 | 30.3362 s | 0.346% | 13.83 M | 41.80 M | 837 MiB |
| P192 | 27 | 103.2347 s | 0.513% | 13.71 M | 41.35 M | 2807 MiB |

Every measured run completed 100 steps/200 stages with identical case-specific numerical checksums and zero floor/finite-value violations. Raw baseline-subtracted device footprint remained about 5929 MiB because the default AMReX arena reserved approximately 5823 MiB. Actual maximum-used storage scales near the frozen 415 B/valid-cell model.

P256A allocated and initialized successfully: arena maximum used 6644 MiB and only 277 MiB remained free. It was therefore not timed and does not satisfy 20% headroom.
7 changes: 7 additions & 0 deletions docs/perf/m1b-p0-profile-summary.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
# M1b-P0 Nsight Systems summary

The P128 profiling build captured four complete physical steps (eight stages), including initialization and final reports. The steady hydro span contains 353 kernels across four streams: 192 flux, 64 update, 24 included CFL box reductions, 56 physical-boundary fill, seven copy, seven FillBoundary copy, and three reduction-finalization kernels.

The hydro core is exactly three directional flux operations plus one update per box and stage: 32 physical core launches/stage for eight boxes. Flux kernels consume 1189.45 ms (90.8% of all traced GPU-kernel time), updates 51.54 ms, and included CFL box reductions 36.79 ms. No face-to-cell atomic exists; atomics are limited to diagnostic counters and AMReX reductions.

The hydro span contains 226 `cudaStreamSynchronize` calls with 1228.75 ms inclusive time. There is no steady `cudaMalloc` or `cudaFree`; the only `cudaMalloc`, managed allocation, and host allocation occur during initialization, and frees occur at teardown. This synchronization topology is the material P0 limitation to correct before adding multilevel work.
10 changes: 10 additions & 0 deletions docs/perf/m1b-p0-raw-artifact-manifest.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
artifact location sha256
external-manifest /home/tim/Downloads/blastgpu-perf-runs/m1b-p0-20260822T142031Z/manifests/manifest.sha256 0f6ebd27d1d5af10801cd0ca075eeb1f46b2d78d49b9190d3b67708e5ec07019
nsight-P128 profiles/P128-steady.nsys-rep d96e1d3bf4be09adaa10356fb52069166b1ac66e524b500a150170eae5d7d210
m0-sod-N256 references/sod-m0-n256.tsv 1bcc73742e2f94c326e1930bed0b29490143da0b711d1a1ac31e943b82bbcc68
m0-sod-N512 references/sod-m0-n512.tsv a75a9edab9f67343f20482078447ca7a3df733fb1575fe6fea6b79b76119e85a
m0-sod-N1024 references/sod-m0-n1024.tsv 349c2cd47498ed18e70434d28b837ef33e2101e9b1fb39bfdc6024d7bcb6e91b
cuda-U0-log correctness/authoritative-cuda/u0.log 919659daac4ac31a1c41565c30619d765b8080666b60b85954ea77940e058ffa
cuda-SX-log correctness/authoritative-cuda/sx.log 8ade98882efcb55155db8fade66e775eb0c8b5ba9dc56e7f26cb8fd7ac9982a8
cuda-SY-log correctness/authoritative-cuda/sy.log f47c6482ad030b4cb0890654463b93493f6f960bcdb58b09032fbd5e2102dd80
cuda-SZ-log correctness/authoritative-cuda/sz.log fcf3b84b93e8011edaadb80b42b8eac49f2bc8ee3972d1cd164d9286d809dc53
5 changes: 5 additions & 0 deletions docs/perf/m1b-p0-reference-generation.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
# M1b-P0 reference generation

`m1b_p0_reference` runs the frozen M0 planar Sod case at N=256, 512, and 1024 using the exact primitive states, floors, van Leer MUSCL, HLLC, SSP-RK2, CFL 0.45, outflow boundaries, and final time 0.2. All three references reported zero violations.

The N=256-to-512 normalized L1 differences were 0.00104 density, 0.00213 velocity, 0.000764 pressure, and 0.000830 energy. The N=512-to-1024 values decreased to 0.000566, 0.00106, 0.000377, and 0.000416. Tolerances in `m1b-p0-tolerances.json` were committed before performance measurements.
4 changes: 4 additions & 0 deletions docs/perf/m1b-p0-timing-results.tsv
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
case hydro_runs_s median_s mean_s min_s max_s cv_percent step_ms cell_stage_per_s face_eval_per_s
P64 3.733731498,3.748369142,3.741828144,3.749401300,3.731769491,3.733964577,3.749514113 3.741828144 3.741225466 3.731769491 3.749514113 0.1983 37.4183 14011547 42691432
P128 30.147818350,30.393191330,30.336207410,30.255957270,30.277352760,30.445533850,30.464592920 30.336207410 30.331521984 30.147818350 30.464592920 0.3456 303.3621 13826066 41802246
P192 103.314296300,103.567200900,103.784572600,103.234737200,102.828022400,102.880156900,102.059773500 103.234737200 103.095537114 102.059773500 103.784572600 0.5126 1032.3474 13712222 41350918
Loading
Loading