|
| 1 | +name: CUDA tests (simulated H100) |
| 2 | + |
| 3 | +# Optional, non-required check. The hosted runners have no GPU, so this builds |
| 4 | +# the test suite and runs the GPU-tagged tests on a simulated NVIDIA H100 |
| 5 | +# (https://github.com/pantheongpu/pantheonsim). A simulator cannot reproduce |
| 6 | +# hardware-specific behavior or timing, so it complements real-GPU testing and |
| 7 | +# does not replace it. |
| 8 | + |
| 9 | +on: |
| 10 | + workflow_dispatch: |
| 11 | + pull_request: |
| 12 | + branches: ["main"] |
| 13 | + paths: |
| 14 | + - "include/**" |
| 15 | + - "tests/**" |
| 16 | + - "examples/**" |
| 17 | + - "CMakeLists.txt" |
| 18 | + - ".github/workflows/gpu-sim.yml" |
| 19 | + |
| 20 | +concurrency: |
| 21 | + group: ${{ github.workflow }}-${{ github.ref }} |
| 22 | + cancel-in-progress: true |
| 23 | + |
| 24 | +permissions: |
| 25 | + contents: read |
| 26 | + |
| 27 | +jobs: |
| 28 | + cuda-sim: |
| 29 | + name: CUDA 13.0, simulated H100 |
| 30 | + runs-on: ubuntu-24.04 |
| 31 | + timeout-minutes: 120 |
| 32 | + |
| 33 | + steps: |
| 34 | + - uses: actions/checkout@v4 |
| 35 | + with: |
| 36 | + submodules: true |
| 37 | + persist-credentials: false |
| 38 | + |
| 39 | + - name: Set up the simulated GPU and CUDA toolkit |
| 40 | + uses: pantheongpu/setup-pantheonsim@ca87fe3143b29166822b923d6050945c5179408c # v0.1.5 |
| 41 | + with: |
| 42 | + gpu: nvidia/h100 |
| 43 | + cuda-toolkit: "13.0" |
| 44 | + |
| 45 | + # The simulator runs PTX only (90-virtual) and needs a shared CUDA runtime. |
| 46 | + - name: Configure CMake |
| 47 | + run: | |
| 48 | + cmake -B build -S . \ |
| 49 | + -DCMAKE_BUILD_TYPE=Release \ |
| 50 | + -DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.0/bin/nvcc \ |
| 51 | + -DCMAKE_CUDA_ARCHITECTURES=90-virtual \ |
| 52 | + -DCMAKE_CUDA_RUNTIME_LIBRARY=Shared \ |
| 53 | + -DKERNEL_FLOAT_BUILD_TEST=1 \ |
| 54 | + -DKERNEL_FLOAT_BUILD_EXAMPLE=1 |
| 55 | +
|
| 56 | + - name: Build |
| 57 | + run: cmake --build build -j3 |
| 58 | + |
| 59 | + # Four test cases are skipped (see the filter below): |
| 60 | + # - binary operators and constant eq tests for __nv_bfloat16: the |
| 61 | + # simulator compares .bf16 values by raw bits in setp |
| 62 | + # (pantheongpu/pantheonsim#502). |
| 63 | + # - ternary operators for __half: the reference expression x * y + z |
| 64 | + # compiles to mul.f16 and add.f16 without a rounding modifier. Hardware |
| 65 | + # fuses them into an fma, the simulator rounds twice, so the last bit |
| 66 | + # can differ. |
| 67 | + # - ternary operators for __nv_bfloat16: also fails on an RTX 3060 (sm_86). |
| 68 | + - name: Run tests |
| 69 | + working-directory: build |
| 70 | + run: | |
| 71 | + vgpu run ./tests/kernel_float_tests --durations=yes \ |
| 72 | + '~"binary operators - GPU - __nv_bfloat16" ~"constant eq tests - GPU - __nv_bfloat16" ~"ternary operators - GPU - __half" ~"ternary operators - GPU - __nv_bfloat16"' |
| 73 | +
|
| 74 | + # vector_add and vector_add_tiling are not run: they index empty |
| 75 | + # std::vectors and crash on real hardware as well. |
| 76 | + - name: Run example |
| 77 | + working-directory: build/examples |
| 78 | + run: vgpu run ./pi/kernel_float_pi |
0 commit comments