Skip to content

Add an optional workflow that runs the solution test on a simulated H100 - #45

Open
saqibkh wants to merge 1 commit into
UoB-HPC:mainfrom
saqibkh:add-gpu-sim-ci
Open

saqibkh wants to merge 1 commit into
UoB-HPC:mainfrom
saqibkh:add-gpu-sim-ci

Conversation

@saqibkh

@saqibkh saqibkh commented Nov 10, 2026

Copy link
Copy Markdown

The existing CI builds the cuda and cuda-thrust models but does not run them (skip_run: true, no GPU on the runners), so the ctest solution case is never run on those models. This adds a separate, optional workflow .github/workflows/gpu-sim.yml that builds both and runs solution (bm1, 1024 poses, energies checked against the deck) on a simulated NVIDIA H100 using PantheonSim (no GPU needed). I am the author of PantheonSim, so please weigh this accordingly.

It touches no source file, does not change linux.yml and is not meant to be a required check. A simulator cannot reproduce hardware-specific behavior or timing, so it complements real-GPU testing and does not replace it.

What it does:

  • installs the CUDA 13.0 toolkit and the simulated H100
  • builds the cuda model with -DCUDA_ARCH=sm_90 and the thrust model with -DTHRUST_IMPL=CUDA -DBACKEND=CUDA -DCUDA_ARCH=90-virtual, both with -DCUDA_EXTRA_FLAGS="-cudart shared". The simulator runs PTX and needs a shared CUDA runtime. The thrust model passes CUDA_ARCH straight to CMAKE_CUDA_ARCHITECTURES, so it needs the -virtual form; for the cuda model sm_90 also embeds PTX
  • runs vgpu run ctest -L solution for each
  • runs on manual dispatch and on pull requests to main that touch src/cuda, src/thrust, src/main.cpp, src/bude.h, data, cmake or CMakeLists.txt, with a concurrency group that cancels superseded runs

Measured:

  • Run on my fork: https://github.com/saqibkh/miniBUDE/actions/runs/42748051241 . Both jobs passed, 1 of 1 ctest test (solution) each.
  • The jobs took about 20 minutes each, almost all of it simulator and toolkit setup; the solution test took 33 seconds in the cuda job.
  • Locally the simulated H100 reports max_diff_% 0.002 for the cuda model (0.001 for thrust), the same as an RTX 3060 for the cuda model.

Not covered:

  • Only ppwi 1 and wgsize 1 (the solution case defaults) are run, not the other PPWI values.
  • The AdaptiveCpp, HIP and other models are untouched.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant