ideal_mhd_model: share the Jacobian kernel with exact autodiff (fwd vs rev)

.github/actions/run-benchmarks/action.yml

@@ -0,0 +1,188 @@
+name: Run VMEC++ benchmarks
+description: >
+  Install VMEC++ and run the benchmark suites in a fixed environment, then store
+  or compare the results. Two suites run here: the Python pytest end-to-end
+  benchmarks and the C++ Google Benchmark function-level microbenchmarks. This is
+  the single source of truth for the benchmark environment so the pull-request
+  and main/docs runs cannot drift apart; the only intended difference between
+  callers is whether results are written back to the benchmark-runs branch
+  (auto-push) and whether a regression is commented on the PR (comment-on-alert).
+
+inputs:
+  auto-push:
+    description: "Push results to the benchmark-runs branch (main/docs run)."
+    default: "false"
+  comment-on-alert:
+    description: "Comment on the PR when a regression exceeds the alert threshold."
+    default: "false"
+  github-token:
+    description: "Token used to push results and/or comment on alerts."
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: "3.10"
+
+    - name: Install system dependencies
+      shell: bash
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y build-essential cmake libnetcdf-dev liblapack-dev libomp-dev libhdf5-dev
+
+    - name: Mount bazel cache
+      uses: actions/cache@v4
+      with:
+        path: "~/.cache/bazel"
+        key: bazel-bench
+
+    - name: Install vmecpp with benchmark dependencies
+      shell: bash
+      run: |
+        pip install uv
+        uv pip install --system -v .[benchmark]
+
+    # --- Python end-to-end benchmarks (pytest-benchmark) ---
+
+    - name: Run Python benchmarks
+      shell: bash
+      run: |
+        export OMP_WAIT_POLICY=passive
+        pytest benchmarks/test_benchmarks.py --benchmark-json=benchmark_results.json
+
+    # github-action-benchmark's "pytest" tool always plots pytest-benchmark's
+    # `ops` field as iter/sec, but the C++ suite emits seconds.
+    # We convert both to the same format, so plotting is the same.
+    - name: Convert Python benchmark result to seconds
+      shell: bash
+      run: |
+        python3 benchmarks/to_seconds_json.py pytest benchmark_results.json benchmark_results_seconds.json
+
+    - name: Store / compare Python benchmark result
+      # Skip the result upload/compare for fork PRs: their GITHUB_TOKEN is
+      # read-only, so comment-on-alert/auto-push hit 'Resource not accessible
+      # by integration'. The benchmarks still run above; only the write-back is
+      # skipped for forks. Main/docs runs (auto-push) always store.
+      if: ${{ inputs.auto-push == 'true' || github.event.pull_request.head.repo.full_name == github.repository }}
+      uses: benchmark-action/github-action-benchmark@v1.21.0
+      with:
+        tool: "customSmallerIsBetter"
+        output-file-path: benchmark_results_seconds.json
+        gh-pages-branch: benchmark-runs
+        benchmark-data-dir-path: benchmarks
+        alert-threshold: "150%"
+        fail-on-alert: false
+        auto-push: ${{ inputs.auto-push }}
+        comment-on-alert: ${{ inputs.comment-on-alert }}
+        github-token: ${{ inputs.github-token }}
+
+    # --- C++ function-level microbenchmarks (Google Benchmark) ---
+    #
+    # Builds the four benchmark cc_binary targets, runs each with
+    # --benchmark_format=json, and merges their output into a single
+    # cpp_benchmark_results.json (schema: {"context": ..., "benchmarks": [...]})
+    # for the conversion step below. Entries with "error_occurred" (e.g. an
+    # FFT-path benchmark skipped because no FFTX codelet exists for that
+    # resolution) are dropped: they carry no valid timing to track.
+
+    - name: Run C++ microbenchmarks
+      shell: bash
+      run: |
+        set -euo pipefail
+
+        TARGETS=(
+          "//vmecpp/vmec/ideal_mhd_model:fft_toroidal_bench"
+          "//vmecpp/vmec/ideal_mhd_model:dealias_constraint_force_bench"
+          "//vmecpp/free_boundary/laplace_solver:laplace_solver_bench"
+          "//vmecpp/vmec/output_quantities:output_quantities_bench"
+        )
+
+        OUT_FILE="${GITHUB_WORKSPACE}/cpp_benchmark_results.json"
+        TMP_DIR="$(mktemp -d)"
+        trap 'rm -rf "${TMP_DIR}"' EXIT
+
+        cd src/vmecpp/cpp
+
+        # --config=perf: optimized build with frame pointers and symbols
+        # retained, matching the existing .bazelrc convention for
+        # performance work.
+        bazel build --config=perf -- "${TARGETS[@]}"
+
+        # Run each target. `bazel run` so runfiles (e.g. test_data for the
+        # output-quantities benchmark) are available in the working
+        # directory. --benchmark_min_time keeps CI runtime bounded while
+        # still averaging enough iterations for a stable measurement.
+        for target in "${TARGETS[@]}"; do
+          name="${target##*:}"
+          bazel run --config=perf -- "${target}" \
+            --benchmark_min_time=0.2s \
+            --benchmark_format=json \
+            --benchmark_out="${TMP_DIR}/${name}.json"
+        done
+
+        python3 - "${TMP_DIR}" "${OUT_FILE}" <<'PY'
+        import glob
+        import json
+        import os
+        import sys
+
+        tmp_dir, out_file = sys.argv[1], sys.argv[2]
+        merged = {"context": None, "benchmarks": []}
+        skipped = 0
+        for path in sorted(glob.glob(os.path.join(tmp_dir, "*.json"))):
+            with open(path) as f:
+                data = json.load(f)
+            if merged["context"] is None:
+                merged["context"] = data.get("context")
+            for bench in data.get("benchmarks", []):
+                if bench.get("error_occurred"):
+                    skipped += 1
+                    continue
+                merged["benchmarks"].append(bench)
+
+        with open(out_file, "w") as f:
+            json.dump(merged, f, indent=2)
+
+        print(f"Merged {len(merged['benchmarks'])} benchmark entries "
+              f"({skipped} skipped/errored).")
+        PY
+
+    # github-action-benchmark's "pytest" tool always plots pytest-benchmark's
+    # `ops` field as iter/sec, but the C++ suite emits seconds.
+    # We convert both to the same format, so plotting is the same.
+    - name: Convert C++ benchmark result to seconds
+      shell: bash
+      run: |
+        python3 benchmarks/to_seconds_json.py googlecpp cpp_benchmark_results.json cpp_benchmark_results_seconds.json
+
+    - name: Store / compare C++ benchmark result
+      # Same fork-PR skip rationale as the Python store step above.
+      #
+      # Runs sequentially after the Python store step (composite-action steps
+      # are serial), so the two auto-push writes to benchmark-runs never race.
+      # A distinct suite name keeps the C++ results under their own key in
+      # data.js (window.BENCHMARK_DATA.entries) so they never overwrite the
+      # Python "Benchmark" suite, even though both share benchmark-data-dir-path.
+      #
+      # skip-fetch-gh-pages: the Python store step above already fetched the
+      # benchmark-runs branch into a local ref and (in compare mode) advanced
+      # it, so a second `git fetch benchmark-runs:benchmark-runs` here is
+      # rejected as non-fast-forward. Reuse the already-fetched local ref
+      # instead of re-fetching.
+      if: ${{ inputs.auto-push == 'true' || github.event.pull_request.head.repo.full_name == github.repository }}
+      uses: benchmark-action/github-action-benchmark@v1.21.0
+      with:
+        name: "C++ Microbenchmarks"
+        tool: "customSmallerIsBetter"
+        output-file-path: cpp_benchmark_results_seconds.json
+        gh-pages-branch: benchmark-runs
+        benchmark-data-dir-path: benchmarks
+        skip-fetch-gh-pages: true
+        alert-threshold: "150%"
+        fail-on-alert: false
+        auto-push: ${{ inputs.auto-push }}
+        comment-on-alert: ${{ inputs.comment-on-alert }}
+        github-token: ${{ inputs.github-token }}

.github/workflows/benchmarks.yaml

@@ -15,40 +15,11 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
 
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Install system dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y build-essential cmake libnetcdf-dev liblapack-dev libomp-dev libhdf5-dev
-
-      - name: Install vmecpp with benchmark dependencies
-        run: |
-          pip install uv
-          uv pip install --system -v .[benchmark]
-
+      # Shared environment + run; PR runs compare and comment but never write
+      # back (auto-push stays false).
       - name: Run benchmarks
-        run: |
-          export OMP_WAIT_POLICY=active
-          pytest benchmarks/test_benchmarks.py --benchmark-json=benchmark_results.json
-
-      - name: Compare benchmark result
-        # Skip the result upload/compare for fork PRs: their GITHUB_TOKEN is
-        # read-only, so comment-on-alert/auto-push hit 'Resource not accessible
-        # by integration'. The benchmarks still run above; only the write-back
-        # is skipped for forks.
-        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository
-        uses: benchmark-action/github-action-benchmark@v1.21.0
+        uses: ./.github/actions/run-benchmarks
         with:
-          tool: "pytest"
-          output-file-path: benchmark_results.json
-          gh-pages-branch: benchmark-runs
-          benchmark-data-dir-path: benchmarks
-          auto-push: false
-          comment-on-alert: true
-          alert-threshold: "150%"
-          fail-on-alert: false
+          auto-push: "false"
+          comment-on-alert: "true"
           github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/clang_tidy.yaml

@@ -31,9 +31,11 @@ jobs:
           apt_packages: libhdf5-dev, liblapack-dev, libnetcdf-dev, gfortran, python3-dev, libomp-dev
           build_dir: build
           cmake_command: cmake -B build && cmake -DCMAKE_EXPORT_COMPILE_COMMANDS:STRING=ON build
-          # bazel-built test files have no cmake compile_commands entry, so
-          # clang-tidy cannot resolve <gtest/gtest.h>; exclude them from review
-          exclude: "./src/vmecpp/cpp/bazel-*,./src/vmecpp/cpp/external,./src/vmecpp/cpp/third_party,*_test.cc"
+          # bazel-built test and benchmark files have no cmake compile_commands
+          # entry, so clang-tidy cannot resolve <gtest/gtest.h> /
+          # <benchmark/benchmark.h> (both are Bazel-only deps); exclude them
+          # from review
+          exclude: "./src/vmecpp/cpp/bazel-*,./src/vmecpp/cpp/external,./src/vmecpp/cpp/third_party,*_test.cc,*_bench.cc"
           # turn off "LGTM" comments: we only want pings about warnings/errors
           lgtm_comment_body: ""
         # If there are any comments, fail the check

.github/workflows/docs.yaml

@@ -24,36 +24,13 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
 
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Install system dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y build-essential cmake libnetcdf-dev liblapack-dev libomp-dev libhdf5-dev
-
-      - name: Install vmecpp with benchmark dependencies
-        run: |
-          pip install uv
-          uv pip install --system -v .[benchmark]
-
+      # Shared environment + run; main/docs runs write results back to the
+      # benchmark-runs branch (auto-push) so they become the PR baseline.
       - name: Run benchmarks
-        run: |
-          pytest benchmarks/test_benchmarks.py --benchmark-json=benchmark_results.json
-
-      - name: Store benchmark result
-        uses: benchmark-action/github-action-benchmark@v1.21.0
+        uses: ./.github/actions/run-benchmarks
         with:
-          tool: "pytest"
-          output-file-path: benchmark_results.json
-          gh-pages-branch: benchmark-runs
-          benchmark-data-dir-path: benchmarks
-          alert-threshold: "150%"
-          fail-on-alert: false
+          auto-push: "true"
           github-token: ${{ secrets.GITHUB_TOKEN }}
-          auto-push: true
 
   build-docs:
     if: always()

.github/workflows/full_validation.yaml

@@ -0,0 +1,70 @@
+name: Full V&V against reference VMEC
+
+# Runs the full Verification & Validation suite from proximafusion/vmecpp-validation
+# (all ~219 input configurations against reference Fortran VMEC2000), using the
+# vmecpp version built from this commit rather than the version pinned on PyPI/GitHub.
+# This is a lot more exhaustive (and slower) than the "short" validation that
+# vmecpp-validation itself runs on its own PRs, so it is not run on every push here.
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref || github.run_id }}
+  cancel-in-progress: ${{ github.ref_name != 'main' }}
+
+jobs:
+  full-validation:
+    name: Run full VMEC++ validation
+    runs-on: ubuntu-22.04
+    # The full parameter scan runs ~219 configurations, each computing a reference
+    # wout with Fortran VMEC2000 in Docker plus one with VMEC++: budget generously.
+    timeout-minutes: 360
+    steps:
+      - name: Check out VMEC++
+        uses: actions/checkout@v4
+        with:
+          path: vmecpp
+          lfs: true
+
+      - name: Check out vmecpp-validation
+        uses: actions/checkout@v4
+        with:
+          repository: proximafusion/vmecpp-validation
+          path: vmecpp-validation
+          lfs: true
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Install required system packages for Ubuntu
+        run: |
+          sudo apt-get update && sudo apt-get install -y \
+            build-essential cmake libnetcdf-dev liblapack-dev liblapacke-dev libopenmpi-dev \
+            libomp-dev libeigen3-dev nlohmann-json3-dev libhdf5-dev
+
+      - name: Install vmecpp-validation's Python requirements
+        run: |
+          cd vmecpp-validation
+          # Install everything except vmecpp itself: we build and install vmecpp
+          # from this commit below instead of the version pinned in requirements.txt.
+          grep -v '^vmecpp@' requirements.txt > requirements.no-vmecpp.txt
+          python -m pip install -r requirements.no-vmecpp.txt
+
+      - name: Install VMEC++ from this commit
+        run: python -m pip install ./vmecpp
+
+      - name: Run full validation
+        working-directory: vmecpp-validation
+        run: python validate_vmec.py
+
+      - name: Upload V&V results
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: vnvresults
+          path: vmecpp-validation/vnvresults_*/
+          retention-days: 30

.github/workflows/test_bazel.yaml

@@ -35,15 +35,6 @@ jobs:
         run: |
           cd src/vmecpp/cpp
           bazel build --config=${{ matrix.config }} -- //...
-      - uses: actions/checkout@v4
-        with:
-          repository: proximafusion/vmecpp_large_cpp_tests
-          path: src/vmecpp/cpp/vmecpp_large_cpp_tests
-          lfs: true
-      - name: Build VMEC++ C++ tests via bazel
-        run: |
-          cd src/vmecpp/cpp
-          bazel build --config=${{ matrix.config }} -- //vmecpp_large_cpp_tests/...
       - name: Run full C++ test suite
         run: |
           cd src/vmecpp/cpp

.github/workflows/test_bazel_tsan.yaml

@@ -39,11 +39,6 @@ jobs:
           IMAGE=ghcr.io/${{ github.repository_owner }}/vmecpp-tsan:latest
           docker pull "${IMAGE,,}"
           docker tag "${IMAGE,,}" proximafusion/tsan
-      - uses: actions/checkout@v4
-        with:
-          repository: proximafusion/vmecpp_large_cpp_tests
-          path: src/vmecpp/cpp/vmecpp_large_cpp_tests
-          lfs: true
       - name: Mount bazel cache
         uses: actions/cache@v4
         with: