diff --git a/.github/workflows/perf.yml b/.github/workflows/perf.yml index 3833a851..7ea98820 100644 --- a/.github/workflows/perf.yml +++ b/.github/workflows/perf.yml @@ -23,15 +23,15 @@ name: Performance # # Locally: python3 tools/perf/bench_gate.py all --base origin/main +# +# The Callgrind job is a required check on main, so it has to report on every pull request: +# pull requests run this workflow whatever they change, and the `changes` job skips both +# comparisons when nothing they measure changed (a skipped job counts as passed). Pushes to +# main keep their path filter. + on: pull_request: types: [opened, synchronize, reopened, labeled, unlabeled] - paths: - - 'engine/**' - - 'crates/dstudio-core/**' - - 'Cargo.lock' - - 'tools/perf/**' - - '.github/workflows/perf.yml' push: branches: [main] paths: @@ -60,8 +60,38 @@ env: PERF_ACCEPT: ${{ contains(github.event.pull_request.labels.*.name, 'perf-regression-accepted') && '1' || '' }} jobs: + changes: + name: Benchmarked code changed + runs-on: ubuntu-24.04 + timeout-minutes: 5 + outputs: + perf: ${{ steps.changes.outputs.perf }} + steps: + - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1 + if: github.event_name == 'pull_request' + with: + persist-credentials: false + fetch-depth: 0 + - name: Paths the benchmarks depend on + id: changes + env: + EVENT: ${{ github.event_name }} + run: | + # Pushes and manual runs always compare. A pull request compares when it touches + # what the benchmarks build, or when the diff can't be read. + perf=true + if [ "$EVENT" = pull_request ] && files=$(git diff --name-only "$BASE_REF" HEAD); then + if ! printf '%s\n' "$files" | grep -qE '^(engine/|crates/dstudio-core/|Cargo\.lock$|tools/perf/|\.github/workflows/perf\.yml$)'; then + perf=false + fi + fi + echo "perf=$perf" | tee -a "$GITHUB_OUTPUT" + instructions: name: Instructions (Linux, Callgrind) + needs: changes + # Skipped (which counts as passed) only when the changes job said so; if it failed, compare anyway. + if: ${{ !cancelled() && needs.changes.outputs.perf != 'false' }} runs-on: ubuntu-24.04 timeout-minutes: 90 steps: @@ -102,8 +132,9 @@ jobs: time: name: Time (${{ matrix.name }}) + needs: changes # Adding or removing the accept label only matters to the gate above. - if: github.event.action != 'labeled' && github.event.action != 'unlabeled' + if: ${{ !cancelled() && needs.changes.outputs.perf != 'false' && github.event.action != 'labeled' && github.event.action != 'unlabeled' }} strategy: fail-fast: false matrix: diff --git a/docs/building.md b/docs/building.md index afbe7f97..0829be7d 100644 --- a/docs/building.md +++ b/docs/building.md @@ -174,10 +174,12 @@ A **lint** job checks the workflows (actionlint, which also runs shellcheck on their scripts), the scripts in `tools/`, the change detection below and the agent skills, in under a minute. -The required checks are **build-test (macOS)**, **build-test (Windows)** and -**build-test (Linux)**. Each passes only when lint and every tests and app job -passed, so a failure on one system turns all three red; the failing job is -named in the check list. +The required checks are **build-test (macOS)**, **build-test (Windows)**, +**build-test (Linux)**, **lint** and the performance gate **Instructions (Linux, +Callgrind)** (below). Each build-test check passes only when lint and every +tests and app job passed, so a failure on one system turns all three red; the +failing job is named in the check list. The Security workflow's checks and the +timing jobs are advisory. Pull requests run only what their changes can reach ([`tools/ci/changes.sh`](../tools/ci/changes.sh), tested by @@ -238,10 +240,12 @@ they are updated. ### Performance checks -[`.github/workflows/perf.yml`](../.github/workflows/perf.yml) runs on pull +[`.github/workflows/perf.yml`](../.github/workflows/perf.yml) compares on pull requests that touch `engine/`, `crates/dstudio-core/`, `Cargo.lock`, `tools/perf/` or the workflow itself, and on -pushes to `main`. It builds every engine benchmark (`engine/bench/`) and the +pushes to `main`. It starts on every pull request, because the Instructions +check is required: for any other change its `changes` job skips both +comparisons, and a skipped job counts as passed. It builds every engine benchmark (`engine/bench/`) and the core's Rust benchmarks (`crates/dstudio-core/benches/`) twice on the same runner, at the base commit and with the change, and compares them row by row with [`tools/perf/bench_gate.py`](../tools/perf/bench_gate.py):