gpu-ci.yml

# Runs the test suite on a self-hosted GPU machine with CUDA and OpenCL enabled
name: GPU tests

on:
  pull_request:
    types: [opened, synchronize, reopened, ready_for_review]
    branches: [master]
  merge_group:

env:
  CARGO_TERM_COLOR: always
  # Disable incremental compilation.
  #
  # Incremental compilation is useful as part of an edit-build-test-edit cycle,
  # as it lets the compiler avoid recompiling code that hasn't changed. However,
  # on CI, we're not making small edits; we're almost always building the entire
  # project from scratch. Thus, incremental compilation on CI actually
  # introduces *additional* overhead to support making future builds
  # faster...but no future builds will ever occur in any given CI environment.
  #
  # See https://matklad.github.io/2021/09/04/fast-rust-builds.html#ci-workflow
  # for details.
  CARGO_INCREMENTAL: 0
  # Allow more retries for network requests in cargo (downloading crates) and
  # rustup (installing toolchains). This should help to reduce flaky CI failures
  # from transient network timeouts or other issues.
  CARGO_NET_RETRY: 10
  RUSTUP_MAX_RETRIES: 10
  # Don't emit giant backtraces in the CI logs.
  RUST_BACKTRACE: short
  RUSTFLAGS: -D warnings

concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
  cancel-in-progress: true

jobs:
  cuda:
    name: Rust tests on CUDA
    runs-on: [self-hosted, gpu-ci]
    env:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITITES: compute,utility
    steps:
      - uses: actions/checkout@v4
        with:
          submodules: recursive
      - uses: actions-rs/toolchain@v1
      - uses: taiki-e/install-action@nextest
      - uses: Swatinem/rust-cache@v2
      # Check we have access to the machine's Nvidia drivers
      - run: nvidia-smi
      # The `compute`/`sm` number corresponds to the Nvidia GPU architecture
      # In this case, the self-hosted machine uses the Ampere architecture, but we want this to be configurable
      # See https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
      - name: Set env for CUDA compute
        run: echo "CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | sed 's/\.//g')" >> $GITHUB_ENV
      - name: set env for EC_GPU
        run: echo 'EC_GPU_CUDA_NVCC_ARGS=--fatbin --gpu-architecture=sm_${{ env.CUDA_ARCH }} --generate-code=arch=compute_${{ env.CUDA_ARCH }},code=sm_${{ env.CUDA_ARCH }}' >> $GITHUB_ENV
      - run: echo "${{ env.EC_GPU_CUDA_NVCC_ARGS}}"
      # Check that CUDA is installed with a driver-compatible version
      # This must also be compatible with the GPU architecture, see above link
      - run: nvcc --version
      - name: CUDA tests
        env: 
          EC_GPU_FRAMEWORK: cuda
        run: |
          cargo nextest run --profile ci --cargo-profile dev-ci --features cuda

  opencl:
    name: Rust tests on OpenCL
    runs-on: [self-hosted, gpu-ci]
    env:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITITES: compute,utility
    steps:
      - uses: actions/checkout@v4
        with:
          submodules: recursive
      - uses: actions-rs/toolchain@v1
      - uses: taiki-e/install-action@nextest
      - uses: Swatinem/rust-cache@v2
      # Check we have access to the machine's Nvidia drivers
      - run: nvidia-smi
      # The `compute`/`sm` number corresponds to the Nvidia GPU architecture
      # In this case, the self-hosted machine uses the Ampere architecture, but we want this to be configurable
      # See https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
      - name: Set env for CUDA compute
        run: echo "CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | sed 's/\.//g')" >> $GITHUB_ENV
      - name: set env for EC_GPU
        run: echo 'EC_GPU_CUDA_NVCC_ARGS=--fatbin --gpu-architecture=sm_${{ env.CUDA_ARCH }} --generate-code=arch=compute_${{ env.CUDA_ARCH }},code=sm_${{ env.CUDA_ARCH }}' >> $GITHUB_ENV
      - run: echo "${{ env.EC_GPU_CUDA_NVCC_ARGS}}"
      # Check that CUDA is installed with a driver-compatible version
      # This must also be compatible with the GPU architecture, see above link
      - run: nvcc --version
      # Check that we can access the OpenCL headers
      - run: clinfo
      - name: OpenCL tests
        env:
          EC_GPU_FRAMEWORK: opencl
        run: |
          cargo nextest run --profile ci --cargo-profile dev-ci --features cuda,opencl