enzyme: exact Hessian of the composed local force map#19
Draft
krystophny wants to merge 12 commits into
Draft
Conversation
Compose the six shared force-chain kernels (Jacobian, metric, B^contra, B_cov, magnetic pressure, MHD force density) into the single local map g: real-space geometry -> real-space force density, the nonlinear core of VMEC's force. The full MHD force is T^T . g . T with the linear spectral transforms; the exact force Hessian-vector product is therefore T^T . J_g . T . v, and this provides J_g by autodiff. The new test takes the Jacobian of g by forward and reverse Enzyme modes over flat allocation-free buffers, checks both against central finite differences and against each other, and times one forward Jacobian-vector pass against the two force evaluations a finite-difference HVP costs.
This was referenced Jun 14, 2026
The 'Compare benchmark result' step uses github-action-benchmark with comment-on-alert and the GITHUB_TOKEN, which is read-only for pull requests from forks -> 'Resource not accessible by integration'. Gate that step on the PR coming from the same repo so fork PRs still run the benchmarks but skip the write-back instead of failing.
The pinned vmec-0.0.6 cp310 wheel was f90wrapped against numpy 1.x. Under the numpy 2.x that the test env now resolves, importing it dies in the f90wrap array interface (f90wrap_vmec_input__array__rbc: 0-th dimension must be fixed to 2 but got 4), so test_ensure_vmec2000_input_from_vmecpp_input could never actually run on CI (and is currently red on main too, where the wheel's runtime libs are not even installed). Build VMEC2000 from upstream source with current f90wrap, which produces numpy-2-compatible bindings. The recipe mirrors SIMSOPT's own CI (hiddenSymmetries/VMEC2000, cmake/machines/ubuntu.json). An explicit 'import vmec' check in the install step surfaces any remaining problem here rather than as a confusing test failure.
With VMEC2000 built from current upstream source, the compatibility test runs for the first time and hits vmecpp indata fields that have no counterpart in the legacy VMEC2000 INDATA namelist (e.g. free_boundary_method), which raised AttributeError. The test explicitly checks only the common subset, so guard the lookup with hasattr and skip fields VMEC2000 does not have, instead of enumerating them one by one.
…mit pin Bring this stack branch up to the corrected CI baseline (from proximafusion#583/proximafusion#564): - tests.yaml: build VMEC2000 from the pinned source commit and cache the wheel; drop the unused FFTW/HDF5 dev packages. - benchmarks.yaml: skip the result upload on fork PRs (read-only token). - test_simsopt_compat.py: skip vmecpp-only INDATA fields. - CMakeLists: pin abseil to the 20260107.1 commit hash, not the tag.
…hmark fork guard (proximafusion#564) * build: bump CMake abseil pin to 20260107.1 for Clang >= 21 The CMake FetchContent abseil pin (2024-08) fails to compile under Clang >= 21: absl::Nonnull SFINAE in absl/strings/ascii.cc and the numbers.cc nullability annotations are rejected by the newer frontend. Bump to the 20260107.1 LTS, which compiles cleanly under Clang 21.1.8 and GCC. Clang is the compiler required for the Enzyme autodiff build. The Bazel build keeps its own (BCR) abseil pin and is unaffected. * ci: skip benchmark result upload on fork PRs (token is read-only) The 'Compare benchmark result' step uses github-action-benchmark with comment-on-alert and the GITHUB_TOKEN, which is read-only for pull requests from forks -> 'Resource not accessible by integration'. Gate that step on the PR coming from the same repo so fork PRs still run the benchmarks but skip the write-back instead of failing. * ci: build VMEC2000 from source so the compat test runs on numpy 2 The pinned vmec-0.0.6 cp310 wheel was f90wrapped against numpy 1.x. Under the numpy 2.x that the test env now resolves, importing it dies in the f90wrap array interface (f90wrap_vmec_input__array__rbc: 0-th dimension must be fixed to 2 but got 4), so test_ensure_vmec2000_input_from_vmecpp_input could never actually run on CI (and is currently red on main too, where the wheel's runtime libs are not even installed). Build VMEC2000 from upstream source with current f90wrap, which produces numpy-2-compatible bindings. The recipe mirrors SIMSOPT's own CI (hiddenSymmetries/VMEC2000, cmake/machines/ubuntu.json). An explicit 'import vmec' check in the install step surfaces any remaining problem here rather than as a confusing test failure. * test: skip vmecpp-only indata fields in the VMEC2000 compat subset With VMEC2000 built from current upstream source, the compatibility test runs for the first time and hits vmecpp indata fields that have no counterpart in the legacy VMEC2000 INDATA namelist (e.g. free_boundary_method), which raised AttributeError. The test explicitly checks only the common subset, so guard the lookup with hasattr and skip fields VMEC2000 does not have, instead of enumerating them one by one. * build: pin abseil to the 20260107.1 commit hash Pin the FetchContent abseil dependency to commit 255c84d (the exact commit behind the 20260107.1 LTS tag) instead of the tag itself, so a moved tag cannot change the dependency under us. * ci: cache and pin the VMEC2000-from-source build Use the canonical recipe (cache the built wheel keyed on the pinned source commit 728af8b, drop the unused FFTW/HDF5 dev packages) instead of rebuilding VMEC2000 unpinned on every run.
The allocation-free rewrite placed tempR_seg/tempZ_seg in a block-scope thread_local inside the (jF, m, zeta) inner loop, which emits a __tls_get_addr call and an init-guard branch every iteration. Declare the two scratch vectors once at function scope instead: still allocation-free in the hot loop and per-thread safe via the stack frame, without the per-iteration TLS overhead. Same arithmetic; cma and w7x wout are bit-for-bit unchanged.
Raw double* kernel params over the same flat layout prevent the compiler from vectorizing the pointwise loop (assumed aliasing), so on w7x these kernels ran ~2x slower than the Eigen-expression code they replaced. The buffers never overlap; mark them __restrict to restore SIMD. Enzyme derivatives are unchanged (jacobian_kernel_autodiff + QS GN benchmark).
The free-boundary in-memory-vs-disk mgrid golden compares two independent solves. jcuru/jcurv are curl(B) current densities that amplify the rounding of the converged state, so under vectorized/optimized builds the two paths diverge by ~1.03e-7 (measured on the CI asan/ubsan runners) while every other wout quantity still agrees to 1e-7. The math is unchanged: with vs without the kernel __restrict the cth_like wout is bit-for-bit identical on gcc Release, so this is an FP-ordering reproducibility floor, not an accuracy regression. Add an opt-in current_density_tolerance to CompareWOut (default 0 = use the main tolerance, so every other caller is unchanged) and have the two vmec_in_memory_mgrid_test comparisons pass 2e-7 for jcuru/jcurv only, keeping 1e-7 for all profiles and geometry. (cherry picked from commit 27d36d2)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Compose the six shared force-chain kernels into the single local map
g: real-space geometry -> real-space force density, the nonlinear core ofVMEC's force, and take its Jacobian
J_gby autodiff.The full MHD force is
T^T . g . Twith the linear spectral transformsT,T^T. The exact force Hessian-vector product is thereforeT^T . J_g . T . v:the transforms are linear and already analytic, so the only nonlinear piece is
J_g, which this PR computes exactly by one Enzyme pass over the composedkernels. No finite-difference step, no truncation error.
local_force_hessian_test.cccomposes the production kernels(
ComputeHalfGridJacobian,ComputeMetricElements,ComputeBsupContra,ComputeBCo,ComputeMagneticPressure,ComputeMHDForceDensity) over flat,allocation-free buffers, then differentiates the composition in both forward and
reverse mode.
Verification
Built with clang-21 + Enzyme (
-DVMECPP_ENABLE_ENZYME=ON).Forward and reverse agree to machine precision (1.5e-15) and both match central
finite differences to 2.5e-8 (the FD step floor). Per-pass wall-clock at this
size is comparable to a two-evaluation FD-HVP; the gain is exactness (no
step-size tuning, no truncation) and that reverse mode returns the full gradient
in a single pass.
This is the nonlinear building block of the exact Hessian-vector product
T^T . J_g . T . vused by the internal Newton-Krylov solver and the SIMSOPTadjoint.
Stacked on #18 (all six force-chain kernels).