Skip to content
Open
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
99 changes: 99 additions & 0 deletions .github/workflows/gpu-sim.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
name: CUDA on simulated H100

# Optional, non-required check. The hosted runners have no GPU, so this builds
# the C++11 CUDA kernels and runs them on a simulated NVIDIA H100
# (https://github.com/pantheongpu/pantheonsim). Each program checks its own
# result ("Solution validates") and exits nonzero on failure. A simulator
# cannot reproduce hardware-specific behavior or timing, so it complements
# real-GPU testing and does not replace it.

on:
workflow_dispatch:
pull_request:
branches: [ main ]
paths:
- "Cxx11/*cuda*"
- "Cxx11/*cublas*"
- "Cxx11/*thrust*"
- "Cxx11/prk_*.h"
- "Cxx11/Makefile.legacy"
- "common/make.defs.cuda"
- ".github/workflows/gpu-sim.yml"

concurrency:
group: ${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true

permissions:
contents: read

jobs:
build-and-test:
name: CUDA 13.0, simulated H100
runs-on: ubuntu-24.04
timeout-minutes: 60

steps:
- name: Checkout
uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

- name: Set up the simulated GPU and CUDA toolkit
uses: pantheongpu/setup-pantheonsim@ca87fe3143b29166822b923d6050945c5179408c # v0.1.5
with:
gpu: nvidia/h100
cuda-toolkit: "13.0"

# The thrust and cuBLAS programs include Boost.Range (RANGEFLAG in make.defs)
- name: Install Boost headers
run: |
sudo apt-get update
sudo apt-get install -y --no-install-recommends libboost-dev

# The simulator JIT-compiles PTX and does not run SASS, and it needs a
# shared CUDA runtime. common/make.defs.cuda targets sm_89 with the
# system CUDA 12.6, so point it at the toolkit installed above, ask for
# PTX only (compute_90) and link the shared runtime. The nvcc is called
# by its absolute path, not through the PATH drop-in of the action.
- name: Configure make.defs
run: |
cp common/make.defs.cuda common/make.defs
sed -i \
-e 's#^NVCC=.*#NVCC=/usr/local/cuda-13.0/bin/nvcc#' \
-e 's#--gpu-architecture=sm_89#--gpu-architecture=compute_90 -cudart shared#' \
-e 's#/usr/local/cuda-12.6/targets#/usr/local/cuda-13.0/targets#' \
common/make.defs
grep -E '^(NVCC|CUDAFLAGS)' common/make.defs

# Not built: the *-cudastf programs (need CUDA STF headers fetched by
# ci/install-cudastf.sh), and the MPI, NCCL, NVSHMEM and multi-GPU
# programs (no MPI or several GPUs here).
- name: Build CUDA kernels
working-directory: Cxx11
run: |
make -f Makefile.legacy -j3 \
nstream-cuda nstream-managed-cuda transpose-cuda stencil-cuda p2p-cuda \
nstream-device-thrust transpose-device-thrust \
nstream-cublas transpose-cublas dgemm-cublas sgemm-cublas

# Small sizes; the run arguments follow the usage message of each program.
- name: Run on the simulated H100
working-directory: Cxx11
run: |
set -e
run() { echo "::group::$*"; vgpu run "./$@"; echo "::endgroup::"; }
run nstream-cuda 10 2000000
run nstream-cuda 10 2000000 256 1
run nstream-managed-cuda 10 200000
run nstream-managed-cuda 10 200000 0 1 0 1
for v in 0 1 2 3; do run transpose-cuda 10 256 $v; done
run stencil-cuda 10 1000
run stencil-cuda 10 1000 32 grid 2
run p2p-cuda 10 1024
run p2p-cuda 10 1024 4
run nstream-device-thrust 10 2000000
run transpose-device-thrust 10 256
run nstream-cublas 10 2000000
run transpose-cublas 10 256
run dgemm-cublas 10 256
run dgemm-cublas 10 256 4
run sgemm-cublas 10 256
Loading