open-standards/.github/workflows/jouleclaw-fabrics.yml

name: jouleclaw fabrics

# Every compute fabric gets built, on hardware that has the device.
#
# `metal` did not compile at all -- 3063 errors -- because no job ever built
# it: 40 architecture files carried `#[cfg(feature = "metal")]` blocks that had
# never once been compiled. An unbuilt feature does not stay still, it drifts,
# and nothing here would have noticed. So each fabric's error count is pinned
# and may fall but never rise.
#
# A baseline above zero is a debt, not a pass. Zero means the fabric compiles,
# which is still not the same as working: none of these runs a kernel yet.
#
# `rocm` has no job because `rocm = []` declares a feature with no dependencies
# and no code. It compiles perfectly while providing nothing, and a green check
# would report health it does not have.
#
# Nightly and path-filtered: the fleet's queue is the scarce resource, and these
# numbers move only when someone edits this crate.

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

on:
  schedule:
    # 07:00 UTC daily. Forge cron is UTC.
    - cron: '0 7 * * *'
  pull_request:
    paths:
      - 'jouleclaw/jouleclaw-rs/crates/jouleclaw-omni/**'
      - 'jouleclaw/jouleclaw-rs/Cargo.toml'
      - 'jouleclaw/jouleclaw-rs/Cargo.lock'
      - 'scripts/check-fabric-errors.py'
      - 'scripts/check-metal-shaders.py'
      - '.github/workflows/jouleclaw-fabrics.yml'
  workflow_dispatch:

jobs:
  metal:
    name: metal (macOS, Apple GPU)
    # Metal is macOS-only, so this is the one runner that can build it.
    runs-on: macos
    timeout-minutes: 60
    steps:
      - uses: actions/checkout@v5
      # launchd's PATH omits cargo on the macOS host runner.
      - name: Put cargo on PATH
        shell: bash
        run: echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
      - name: metal's error count may fall, never rise
        shell: bash
        run: python3 scripts/check-fabric-errors.py metal
      # compile_pipeline builds every shader at runtime, so an MSL error is
      # invisible to cargo. xcrun is the same frontend the runtime uses.
      - name: every Metal shader source must compile
        shell: bash
        run: python3 scripts/check-metal-shaders.py
      # Compiling a kernel says nothing about what it computes. This runner
      # has a GPU, so the kernels are dispatched and checked against CPU
      # references here.
      - name: Metal kernels must compute the right thing
        shell: bash
        run: |
          set -euo pipefail
          M=jouleclaw/jouleclaw-rs/Cargo.toml
          # tests/metal_kernels.rs is behind `#![cfg(feature = "metal")]`. If
          # that cfg ever stops holding the file compiles to nothing, the
          # target runs 0 tests, and cargo exits 0 -- a green check for a
          # suite that did not run. Count the tests before trusting the run.
          n=$(cargo test --manifest-path "$M" -p jouleclaw-omni --features metal \
                --test metal_kernels -- --list 2>/dev/null | grep -c ': test$')
          echo "metal_kernels declares $n tests"
          if [ "$n" -lt 26 ]; then
            echo "::error::metal_kernels declares $n tests, expected at least 26."
            echo "::error::Either the metal cfg stopped holding or tests were removed."
            exit 1
          fi
          # Run every target, not just that one. `--test metal_kernels` left
          # the crate's 46 metal-gated *lib* tests unrun: a feature's unit
          # tests are as invisible to CI as its code was.
          cargo test --manifest-path "$M" -p jouleclaw-omni --features metal

  linux-fabrics:
    name: ${{ matrix.feature }} (linux, RTX 4050 + Vulkan loader)
    # The host runner, not the docker one: cudarc wants the CUDA toolkit at
    # /usr/local/cuda and ash wants the system Vulkan loader, and this host has
    # both beside an actual NVIDIA GPU. These features have only ever been
    # built on macOS, where neither device exists.
    runs-on: linux-hw
    timeout-minutes: 90
    strategy:
      fail-fast: false
      matrix:
        feature: [cuda, vulkan]
    steps:
      # This host has no Node, so actions/checkout cannot run here.
      - name: Fetch the source
        shell: bash
        run: |
          set -euo pipefail
          want="${{ github.event.pull_request.head.sha || github.sha }}"
          rm -rf src && mkdir src && cd src
          git init -q
          # Fetch the commit itself, not the ref: a later commit on the ref
          # that this workflow's paths do not match would be fetched instead.
          git fetch -q --depth 1 http://100.64.224.0:3400/Transaction-Science/open-standards.git "$want"
          git checkout -q FETCH_HEAD
          got="$(git rev-parse HEAD)"
          [ "$got" = "$want" ] || { echo "::error::fetched $got, not $want"; exit 1; }
      - name: ${{ matrix.feature }}'s error count may fall, never rise
        shell: bash
        working-directory: src
        run: |
          # nvcc is not on this host's PATH, and cudarc's
          # cuda-version-from-build-system feature asks nvcc which CUDA this is.
          export PATH="$HOME/.cargo/bin:/usr/local/cuda/bin:$PATH"
          python3 scripts/check-fabric-errors.py ${{ matrix.feature }}