examples: adjoint boundary sensitivities + SIMSOPT analytic gradient#11
Draft
krystophny wants to merge 13 commits into
Draft
examples: adjoint boundary sensitivities + SIMSOPT analytic gradient#11krystophny wants to merge 13 commits into
krystophny wants to merge 13 commits into
Conversation
krystophny
force-pushed
the
simsopt-adjoint
branch
from
a86496f to
50ec5ab
Compare
This was referenced Jun 14, 2026
…proximafusion#580) * 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. * enzyme: opt-in Clang/Enzyme build option and AD smoke test Add VMECPP_ENABLE_ENZYME (OFF by default), which requires a Clang compiler and a ClangEnzyme plugin path and builds a self-contained autodiff smoke test. The test differentiates a scalar objective written over Eigen::Map'd caller buffers and checks reverse- and forward-mode Enzyme gradients against the closed form and central finite differences. enzyme.h documents the intrinsic ABI and the allocation constraint that shapes the differentiable kernels: Enzyme cannot track Eigen's aligned allocator, so differentiable paths use Eigen::Map over caller-owned buffers and avoid heap expression temporaries. With the option off the build is unchanged. * pybind: expose the unpreconditioned internal-basis gradient Add a precondition flag to VmecModel.evaluate (default true, unchanged behaviour). With precondition=false the forward model returns at the INVARIANT_RESIDUALS checkpoint, so get_forces() yields the raw, unpreconditioned force: the gradient of VMEC's augmented functional (MHD energy plus the spectral-condensation and lambda constraints) with respect to the decomposed internal-basis state. This is the consistent state/gradient pair an external optimizer needs to minimise in VMEC's own basis. The native solver's preconditioned search direction (precondition=true) is a different vector; the raw gradient is the equilibrium residual and vanishes at convergence. Tests: raw force is finite and differs in direction from the preconditioned force, and drops by >1e6 from the initial guess to the converged equilibrium. * examples: drive VMEC++ from external optimizers in the internal basis Treat the equilibrium as the root problem F(x) = 0, where F is the raw internal-basis force (gradient of VMEC's augmented functional) exposed by evaluate(precondition=False). Wire it to two solvers that reuse VMEC++'s forward model: native-style preconditioned descent and Jacobian-free Newton-Krylov (matrix-free Hessian information). Both reach the native solver's equilibrium. This is the external-differentiability path: VMEC++ as a differentiable equilibrium component an outside optimizer can drive. Quasi-Newton root-finders without a preconditioner diverge on this stiff system, which motivates exposing VMEC's preconditioner as an operator next. Tests assert both solvers reach force balance and recover the native energy and state. * pybind: expose VMEC preconditioner as an operator; preconditioned JFNK Add VmecModel.apply_preconditioner(v): applies VMEC's preconditioner M^-1 (m=1, radial, lambda steps) to a vector in the decomposed basis. M^-1 is VMEC's hand-built approximate inverse Hessian; this exposes it as a reusable linear operator for preconditioned Krylov / quasi-Newton and for the Hessian solve in adjoint sensitivities. It requires a prior evaluate(precondition=true), which assembles the radial preconditioner. Validated exactly: apply_preconditioner(raw force) equals the native preconditioned search direction; the operator is linear and, once assembled, state-invariant. Use it as the inner Krylov preconditioner in Newton-Krylov: on solovev (ns=11) this cuts force evaluations from 2242 to 505 (4.4x) versus unpreconditioned JFNK, converging to the same equilibrium. * pybind: Hessian-vector product inside VMEC++; internal Newton-Krylov Add VmecModel.hessian_vector_product(v): the curvature of VMEC's augmented functional, computed inside VMEC++ as a central directional derivative of the analytic force (its gradient). The force is exact; only the directional step is finite-differenced. Add a force_eval_count for fair cross-optimizer cost comparison (counts evaluations hidden in the Hessian-vector products). Drive a true Newton-Krylov from this HVP plus the preconditioner: it reaches the equilibrium in ~7 outer iterations (second order) versus ~1300 descent steps. This is the inside-the-solver Hessian path; together with the external optimizers it gives differentiability inside and out. Benchmark (solovev, ns=11, force evals counted in VMEC++): preconditioned descent 2606 evals 1302 iters Newton-Krylov (JFNK) 2243 evals Newton-Krylov (preconditioned) 507 evals Newton (VMEC++ HVP + M^-1) 9194 evals 7 iters The HVP-Newton's higher force-eval count (two evals per finite-difference HVP) is what the exact Enzyme Hessian will remove. * ideal_mhd_model: make computeMHDForces allocation-free The force kernel allocated 17 dynamic Eigen vectors per radial surface (the _o half-grid quantities and the avg/wavg surface averages). Move them to preallocated per-thread ThreadLocalStorage scratch and assign in place, so the radial loop allocates nothing. Two benefits: it removes per-surface heap churn from the hot force loop, and it makes the kernel differentiable by Enzyme, which cannot trace dynamic Eigen temporaries (forward and reverse mode both abort on them). This is the allocation-free prerequisite for an exact autodiff Hessian. Pure refactor, identical arithmetic. Verified bit-for-bit: vmec_standalone MHD energy unchanged on solovev (2.548352e+00) and cth_like_fixed_bdy (5.057191e-02). * examples: globalize HVP Newton with a backtracking line search The full Newton step overshoots on stiff 3D equilibria (cth_like stalled at the iteration cap with ||F|| ~ 5e-2). Add a backtracking line search on ||F|| so each step is damped to a decrease. With it the HVP-Newton converges on cth_like in 9 outer iterations (||F|| = 1.8e-10) and still converges solovev in 8. * dft_toroidal: make ForcesToFourier allocation-free The forces transform materialized two per-(surface,m,zeta) Eigen temporaries (tempR_seg, tempZ_seg) inside the inner loop. Reuse per-thread scratch instead, so the whole FFTX-off force path (geometryFromFourier, computeJacobian/Metric/BContra/BCo, pressureAndEnergies, computeMHDForces, forcesToFourier) is now allocation-free end to end. Same arithmetic as the previous .eval(); verified bit-for-bit: solovev 2.548352e+00, cth_like_fixed_bdy 5.057191e-02. * enzyme: exact autodiff of the VMEC Jacobian kernel (forward vs reverse) Demonstrate exact automatic differentiation of a real VMEC nonlinear kernel. JacobianKernel reproduces IdealMhdModel::computeJacobian (half-grid r12/ru12/zu12/rs/zs and the Jacobian tau), written allocation-free over flat buffers, which is the form Enzyme differentiates. For L = 0.5||outputs||^2 the test computes dL/dgeom by reverse mode and the directional derivative dL.v by forward mode, checks both against central finite differences, and against each other: reverse dL.v vs FD : 1.9e-9 forward dL.v vs FD : 1.9e-9 forward vs reverse : 2.9e-15 performance: reverse ~16 us/pass (full gradient), forward ~16 us/pass (one direction) Reverse returns the whole gradient per pass and wins for a scalar gradient; forward is the cheaper primitive for a single Jacobian/Hessian-vector product. tau is nonlinear in the geometry, so this kernel's Jacobian is a genuine building block of the exact MHD force Hessian; the remaining force chain follows the same allocation-free pattern. * ideal_mhd_model: share the Jacobian kernel between solver and autodiff Move the half-grid Jacobian arithmetic into jacobian_kernel.h (ComputeHalfGridJacobian), allocation-free over flat buffers. Production computeJacobian now calls it (followed by the unchanged Jacobian-sign check), and the Enzyme forward/reverse test differentiates the same kernel: one implementation, no duplication. Bit-exact: vmec_standalone MHD energy unchanged on solovev (2.548352e+00) and cth_like_fixed_bdy (5.057191e-02). Autodiff test still matches finite differences and agrees forward vs reverse to 3e-15. * ideal_mhd_model: share the metric kernel (gsqrt, guu, guv, gvv) Extract computeMetricElements into the shared, allocation-free kernel ComputeMetricElements (metric_kernel.h), over flat buffers, and call it from the solver. guv and the 3D part of gvv are computed only when lthreed, matching the original. This is the second force-chain kernel made Enzyme-differentiable (composed into the exact Hessian-vector product later), following the Jacobian kernel pattern. Bit-exact: vmec_standalone MHD energy unchanged on solovev (2.548352e+00, 2D) and cth_like_fixed_bdy (5.057191e-02, 3D path with guv/gvv). * ideal_mhd_model: share the contravariant-field kernel (bsupu, bsupv) Factor the bsupu/bsupv arithmetic out of computeBContra into the shared, allocation-free kernel ComputeBsupContra (bcontra_kernel.h). The lambda normalization (lamscale, + phi') and the chi'/iota profile and toroidal-current-constraint logic stay in the solver verbatim, since they mutate state and update profiles; only the differentiable field arithmetic moves to the shared kernel. Bit-exact across 1 and 4 threads (so the ghost-cell radial partitioning is exercised) on solovev (2.548352e+00, 2D) and cth_like_fixed_bdy (5.057191e-02, 3D). * ideal_mhd_model: share the covariant-field kernel (bsubu, bsubv) Extract the metric index-lowering (bsubu = guu B^u + guv B^v, bsubv = guv B^u + gvv B^v; guv absent in 2D) from computeBCo into the shared, allocation-free kernel ComputeBCo (bco_kernel.h). Bit-exact across 1 and 4 threads on solovev (2.548352e+00) and cth_like_fixed_bdy (5.057191e-02). * ideal_mhd_model: share the magnetic-pressure kernel Extract the field-dependent magnetic pressure |B|^2/2 = 0.5(B^u B_u + B^v B_v) from pressureAndEnergies into the shared, allocation-free kernel ComputeMagneticPressure (pressure_kernel.h). The kinetic-pressure profile and the energy volume integrals stay in the solver. Bit-exact across 1 and 4 threads on solovev (2.548352e+00) and cth_like_fixed_bdy (5.057191e-02). Completes the point-local nonlinear force-chain kernels (Jacobian, metric, B^contra, B_cov, pressure). * ideal_mhd_model: share the MHD force-density kernel Extract computeMHDForces' real-space force-density assembly (armn/azmn/ brmn/bzmn, and crmn/czmn in 3D, even+odd) into the shared, allocation-free kernel ComputeMHDForceDensity (mhdforce_kernel.h). The Eigen arithmetic is preserved verbatim over flat-buffer Eigen::Map views with caller-owned handover/average scratch, so it is bit-for-bit identical. This is the sixth and final point-local force-chain kernel; the six (Jacobian, metric, B^contra, B_cov, pressure, force) now form the local map geometry -> force density, ready to compose into the exact Hessian-vector product. (This branch also merges the allocation-free force kernel, #12, which removes the per-surface heap temporaries this extraction relies on.) Bit-exact across 1 and 4 threads on solovev (2.548352e+00) and cth_like_fixed_bdy (5.057191e-02). * enzyme: exact Hessian of the composed local force map 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. * ideal_mhd_model: share the hybrid lambda-force kernel Extract hybridLambdaForce's full-grid lambda force (blmn, and clmn in 3D) into lambda_force_kernel.h (ComputeHybridLambdaForce), shared between the solver and the Enzyme autodiff path. The method drops from 115 lines to a single kernel call; the OpenMP barriers stay in the method. The kernel is allocation-free over flat buffers and preserves the radial sweep that carries the inside half-grid point in scratch and shifts it outward each surface, plus the blend of the two bsubv interpolations. This is the lambda-force piece of the augmented functional, the second nonlinear force-density term after the MHD force chain. * ideal_mhd_model: share the constraint-force kernels Extract the two local (non-transform) pieces of the spectral-condensation constraint force into constraint_force_kernel.h, shared between the solver and the Enzyme autodiff path: - ComputeEffectiveConstraintForce: gConEff = (rCon-rCon0) ru + (zCon-zCon0) zu (effectiveConstraintForce), skipping the axis surface. - AddConstraintForces: add the bandpass-filtered gCon back into the MHD R/Z forces and write frcon/fzcon (the constraint part of assembleTotalForces). The Fourier-space bandpass between them stays the shared free function deAliasConstraintForce; the free-boundary rBSq contribution stays in assembleTotalForces. Allocation-free over flat buffers. This completes the local force-density terms of the augmented functional (MHD + lambda + constraint), the nonlinear core of the exact Hessian. * enzyme: extend the composed-force Hessian test with the lambda force Add the hybrid lambda force (lambda_force_kernel.h) to the composed local map g and differentiate the combined MHD-plus-lambda force density by forward and reverse Enzyme modes. This proves J_g for the second nonlinear force-density term, not just the MHD force chain. The spectral-condensation constraint force also carries a linear Fourier bandpass; it is validated end-to-end against the finite-difference HVP in the pybind exact-HVP path rather than in this flat-buffer microtest. * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * apply pre-commit formatting (ruff, docformatter, clang-format) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * bazel: declare force-chain kernel headers in ideal_mhd_model (sandbox fix) * test: docformatter-format test_internal_gradient docstrings Satisfies the docformatter pre-commit hook (was failing CI). * test: docformatter-format external/internal optimizer test docstrings Satisfies the docformatter pre-commit hook (was failing CI). * ci: re-trigger (transient apt-403 on packages.microsoft.com) * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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. * 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: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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 for Clang >= 21. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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 for Clang >= 21. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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 for Clang >= 21. * ci: sync VMEC2000-from-source build, benchmark fork guard, abseil commit pin Bring this stack branch up to the corrected CI baseline (from #583/#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 for Clang >= 21. * 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. * ideal_mhd_model: mark Jacobian kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: hoist ForcesToFourier scratch out of the inner loop 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. * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: hoist ForcesToFourier scratch out of the inner loop 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. * ideal_mhd_model: hoist ForcesToFourier scratch out of the inner loop 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. * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: hoist ForcesToFourier scratch out of the inner loop 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. * ideal_mhd_model: hoist ForcesToFourier scratch out of the inner loop 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. * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * ideal_mhd_model: mark Jacobian metric kernel buffers __restrict 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). * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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. * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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 27d36d21e1dd8ea6f73127b95bdc81d529f81672) * output_quantities: compare jcuru/jcurv at a looser opt-in tolerance 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…
…sion#619) moved vmecpp.run_continuation() logic directly into vmecpp.run() Co-authored-by: Philipp Jurašić <166746189+jurasic-pf@users.noreply.github.com>
…n#621) The Eigen3 migration (proximafusion#410) and hot-loop rework (proximafusion#454) replaced the fused scalar poloidal accumulation in the toroidal transforms with per-quantity Eigen .dot() calls, and the FFT path additionally allocates two Eigen vectors per innermost (m,k) iteration via .eval(). On the short theta axis (nThetaReduced ~9-16) and small ntor+1 this is a pessimization: benchmark-runs history shows ToroidalForcesToFourier regressed ~2x from the pre-proximafusion#410 baseline and never recovered, including at the flagship 12x12 FFT size. Two changes: - dft_toroidal.cc: restore the pre-proximafusion#410 fused-scalar-loop DFT code verbatim (single pass over theta reading each basis value once; the original "auto-vectorize was a pessimization" note is retained). Fixes the DFT-fallback resolutions and the fftx-disabled build. - fft_toroidal.cc: the FFT path only replaces the toroidal direction; its poloidal fill kept the .dot()+.eval() pattern. Fuse it into one allocation-free scalar pass. Measured (same-machine A/B, --config=opt, OMP=1): 12x12 FFT forces 1.57x faster (1.00e-3 -> 6.35e-4), 6x8 neutral. The c2r FourierToReal output scatter already uses plain segment += and is left unchanged. fft_toroidal_test and vmec_test pass. CI benchmarks will confirm the recovery and inform whether any FFT shapes should still fall back to DFT. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fix typos in docs
Run every solver once on explicit Solov'ev and CTH-like inputs. Keep the strict internal-state comparison on the reproducible 2D case, and verify force balance and energy in 3D without optional diagnostics or repeated evaluations. Co-authored-by: Philipp Jurašić <166746189+jurasic-pf@users.noreply.github.com>
krystophny
force-pushed
the
simsopt-adjoint
branch
from
839a8e6 to
eb39c01
Compare
Add the implicit-function adjoint that turns VMEC++ into a gradient-providing equilibrium component for SIMSOPT, the original goal. vmecpp_adjoint.py: for a converged fixed-boundary equilibrium F_I(x)=0, the boundary sensitivity of a scalar objective J follows from H_II lambda = dJ/dx_I, dJ/dx_B = dJ/dx_B - (dF_I/dx_B)^T lambda, with H the symmetric Hessian of the augmented functional. It is matrix-free via hessian_vector_product and apply_preconditioner (the SPD interior system is solved with preconditioned CG). One Hessian solve gives the whole boundary gradient, versus one equilibrium re-solve per boundary DOF for finite differences. simsopt_vmec_gradient.py: VmecEnergy wraps this as a SIMSOPT Optimizable whose dJ is the adjoint gradient, plus a gradient-cost benchmark. Verified: the adjoint gradient matches brute-force re-solve finite differences (rel 2.4e-4) and the SIMSOPT Optimizable's dJ matches finite differences of J (rel ~1e-6). On solovev (ns=11, 18 boundary DOFs) the adjoint boundary gradient costs 762 force evaluations versus 9112 for finite differences (12x), and the gap grows with the boundary DOF count.
Two correctness fixes for stiff 3D equilibria (cth_like): - VMEC's augmented-Lagrangian Hessian is symmetric *indefinite* (the lambda constraint makes it a saddle, not a minimum), so CG silently gives the wrong adjoint there. Use GMRES, which handles indefinite systems, for the H_II solve and the interior Newton solve. With a loose, restarted tolerance the adjoint solve stays cheap. - Add a backtracking line search to solve_interior so the interior re-solve (used by the SIMSOPT wrapper and the finite-difference reference) converges on 3D instead of overshooting. Verified with a directional-derivative check against a re-converged finite-difference reference: solovev 1.5e-4, cth_like 2.2e-2 relative; both previously agreed only in 2D. Boundary-gradient cost on solovev: 626 force evaluations (analytic adjoint) versus 10460 (finite differences).
krystophny
force-pushed
the
simsopt-adjoint
branch
from
eb39c01 to
2f258ed
Compare
Fix formatting in docstring for clarity.
* Make raw VMEC forces history independent Separate the legacy previous-residual m=1 projection used by native iteration from the exact constrained projection used by external force evaluations. Cover both directions across the residual threshold. * Use compact storage for constraint policy * Name the m=1 gauge policy explicitly
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
4 participants
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
Use VMEC++ as a differentiable component in an external optimizer: the
boundary-shape gradient
dJ/dx_Bby the implicit-function adjoint instead offinite-differencing over boundary DOFs.
examples/vmecpp_adjoint.py: partition the state into interior/boundary,converge the interior to force balance (
solve_interior), then one adjointsolve
H_II lambda = dJ/dx_I(GMRES preconditioned byM^-1) gives the fullboundary gradient. The interior Hessian is symmetric indefinite (the lambda
constraint is a saddle), so GMRES is used, not CG.
examples/simsopt_vmec_gradient.py: a SIMSOPTOptimizablewrapping it.This PR uses the finite-difference HVP. With the exact autodiff HVP (#23) the
same adjoint gets cheaper still (numbers below).
Verification (force evals counted in VMEC++, ns=11)
The adjoint computes the same gradient as finite-differencing over the boundary
but at a cost independent of the number of boundary DOFs: 7x fewer evals on
solovev (18 DOFs), 93x on cth_like (150 DOFs) with the FD HVP, and 25x / 263x
with the exact HVP (#23). Gradients agree with the FD reference to 3.9e-4 /
4.0e-2.
Stacked on #10 (HVP).