NumSharp/.github/workflows/benchmark.yml at master · SciSharp/NumSharp · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
# =============================================================================
# Benchmark — decoupled, post-release, semi-manual. Runs the ENTIRE NumSharp-vs-
# NumPy comparison via benchmark/run_benchmark.py: the official op/dtype/N matrix
# (BenchmarkDotNet + NumPy across 1K/100K/10M), the NDIter iterator benchmark
# (benchmark/nditer, + README cards), and the matrix subsystems that fill the
# axes the op matrix omits — layout (benchmark/layout), cast (benchmark/cast),
# fusion (benchmark/fusion) — each appended as its own report section.
#
# Trigger is the GitHub Release being published (the "after a successful release"
# signal — 'Build and Release' creates it on a v* tag) or a manual dispatch. It
# is a SEPARATE workflow from the release pipeline on purpose: a slow/failed
# benchmark must never gate a release.
#
# It commits the refreshed report + cards + the benchmark/history/<date>_<sha>/
# snapshot (and the `latest` symlink — see benchmark/scripts/snapshot_history.py)
# straight to master with '[skip ci]' (so the push cannot re-trigger anything).
# GITHUB_TOKEN + contents:write — no PAT.
#
# NOTE on numbers: GitHub runners are shared, variable hardware. ABSOLUTE ms are
# not comparable run-to-run; every ratio is NumPy / NumSharp on the SAME runner,
# which stays meaningful. numpy is pinned. The NDIter harness IGNORES NumSharp's
# known intermittent AccessViolation: a crashing section is reported NA.
# =============================================================================
name: Benchmark

on:
  release:
    types: [published]
  workflow_dispatch: {}

permissions:
  contents: write
  actions: write   # dispatch the Deploy Docs workflow after committing refreshed report pages

concurrency:
  group: benchmark
  cancel-in-progress: false

jobs:
  benchmark:
    runs-on: ubuntu-latest
    timeout-minutes: 180
    steps:
      - name: Checkout master
        uses: actions/checkout@v4
        with:
          ref: master
          fetch-depth: 0

      - name: Setup .NET (8 + 10)
        uses: actions/setup-dotnet@v4
        with:
          dotnet-version: |
            8.0.x
            10.0.x
          dotnet-quality: 'preview'

      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.12'

      - name: Install Python deps (pin numpy for cross-release comparability)
        run: pip install "numpy==2.4.2" matplotlib tabulate

      - name: Run full benchmark (official op suites + NDIter)
        # run_benchmark.py builds the CSharp bench project + Core, runs the official suites, then
        # the NDIter sheet (section-isolated, AV-ignored→NA) + cards, and appends
        # the NDIter section to benchmark-report.md.
        run: python benchmark/run_benchmark.py

      - name: Render full reports into the docs site
        # The op matrix + the iterator sheet become searchable DocFX pages (a
        # different result model from the cards, so they are whole-page includes,
        # not merged). Regenerated here from the just-produced reports so the site
        # always matches the run. The op matrix drops its own H1 (tail -n +2) so the
        # page carries a single title; the iterator sheet is a fenced block with no H1.
        run: |
          { printf '# Operation Matrix — NumSharp vs NumPy (full report)\n\n'
            printf '> _Auto-generated after each release by the [Benchmark workflow](https://github.com/SciSharp/NumSharp/blob/master/.github/workflows/benchmark.yml) — do not edit by hand. Discussion + cards: [Benchmarks vs NumPy](benchmarks.md)._\n\n'
            tail -n +2 benchmark/benchmark-report.md
          } > docs/website-src/docs/benchmark-matrix.md
          { printf '# Iterator Benchmark — NDIter vs NumPy (full sheet)\n\n'
            printf '> _Auto-generated after each release by the [Benchmark workflow](https://github.com/SciSharp/NumSharp/blob/master/.github/workflows/benchmark.yml) — do not edit by hand. This is the canonical sheet the cards on [Benchmarks vs NumPy](benchmarks.md) render from. speedup = NumPy ÷ NumSharp (>1.0× = NumSharp faster)._\n\n'
            cat benchmark/nditer/nditer_results.md
          } > docs/website-src/docs/benchmark-iterator.md

      - name: Commit report + cards, redeploy docs
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          git config user.name  "github-actions[bot]"
          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
          git add benchmark/benchmark-report.md \
                  benchmark/benchmark-report.csv \
                  benchmark/nditer/nditer_results.md \
                  benchmark/nditer/nditer_results.tsv \
                  benchmark/nditer/cards/ops.png \
                  benchmark/nditer/cards/cat.png \
                  benchmark/layout/layout_results.md \
                  benchmark/layout/layout_results.tsv \
                  benchmark/operand/operand_results.md \
                  benchmark/operand/operand_results.tsv \
                  benchmark/cast/cast_results.md \
                  benchmark/cast/cast_results.tsv \
                  benchmark/fusion/fusion_results.md \
                  benchmark/history/ \
                  docs/website-src/docs/benchmark-matrix.md \
                  docs/website-src/docs/benchmark-iterator.md
          # benchmark/history/ holds the committable snapshot run_benchmark.py wrote
          # (via benchmark/scripts/snapshot_history.py): the report + every subsystem
          # sheet + cards + the json/csv/numpy-results that are gitignored at the
          # benchmark root, under history/<date>_<sha>/, plus the `latest` symlink —
          # the durable provenance and the stable path the docs reference.
          if git diff --cached --quiet; then
            echo "No benchmark changes to commit."
            exit 0
          fi
          git commit -m "bench: refresh report + NDIter cards + history snapshot [skip ci]"
          git push origin HEAD:master
          # The commit carries [skip ci] so it cannot re-trigger this workflow or the
          # release pipeline (push-to-master). The refreshed report PAGES + cards live
          # under docs/website-src/**, so explicitly redeploy the docs to publish them
          # (workflow_dispatch ignores docs.yml's path filter).
          gh workflow run docs.yml --ref master