Perf/optimization v2

CHANGELOG.md

@@ -0,0 +1,195 @@
+# Changelog
+
+All notable changes to simcoon are listed here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and simcoon uses
+[semantic versioning](https://semver.org/). Each pull request adds its entry
+under **Unreleased**; the section takes the version number at the release.
+
+## In progress
+
+Known gaps and planned work, not yet in a release. An item moves to
+**Unreleased** with the pull request that closes it.
+
+- `ncp="min"` serves the mechanical EPICP, EPKCP and MODUL only: every
+  thermomechanical law (`_T`, `sim.umat_T`, thermomechanical blocks) is refused
+  by `check_ncp_serves` and keeps Fischer-Burmeister.
+- `ncp="min"` is refused with Armstrong-Frederick back stresses under the
+  cutting plane (path dependent); whether the closest point, which restarts
+  from the start of the step, lifts this is not verified.
+- `integrator="closest_point"` is provided with its algorithmic tangent only
+  (`tangent_mode="algorithmic"`): its continuum operator is not implemented.
+- Other complementarity functions to measure against Fischer-Burmeister and
+  `min`, under both integrators: active set, scaled Fischer-Burmeister.
+- The numerical options of the mean-field schemes (MIMTN/MISCN closed form,
+  rule, Anderson, tolerance) are still props slots: they are per phase (nested
+  schemes), so they need a named builder rather than a run option.
+- fedoo does not pass `ncp` and `integrator` through yet.
+- The external plugins (UMEXT, UMABA, UMANS) serve the small-strain route only: a finite-strain
+  entry in the plugin API (law fed the logarithmic strain, returning tau, as the internal
+  kernels) would let them run under NLGEOM.
+- Migration of the laws to the modular engine (MODUL), state in
+  `benchmark/README.md`. Routed through the adapter (code 201): ELISO, ELIST,
+  ELORT, EPKCP, EPHIL, EPTRI, EPHAC, EPANI, EPDFA, EPCHG, EPHIN. To do:
+  - adapters for the hyperelastic laws whose MODUL equivalent exists (NEOHC,
+    MOORI, YEOHH, ISHAH, GETHH, SWANH, HOLZA; NEOHI to confirm);
+  - EPICP, EPCHA and PRONK keep their legacy kernel while it is faster;
+  - no MODUL equivalent yet: OGDEN, SNTVE, HYPOO, EPJCK, LLDM0, ZENER, ZENNK,
+    the SMA family, the micromechanics schemes;
+  - the thermomechanical dispatch (`umat_T`) has no adapter: every `_T` law
+    runs its legacy kernel.
+
+## Planned for 3.0
+
+Removals and renames scheduled for the next major version (breaking, kept in 2.x).
+
+- The migration message for `tangent_mode=3` / `"closest_point"`: the check
+  `tangent_mode in ("closest_point", 3)` of `solve` (`simcoon/solver/core.py`)
+  and its C++ counterpart in `kernel_tangent_mode` (`umat_smart.cpp`). The
+  value then gets the generic error of an unknown tangent mode; the internal
+  kernel code 3 of the closest point is unchanged.
+- Possibly rename `ZENER` / `ZENNK` after their rheology: they are the Poynting-Thomson
+  (Kelvin) form of the standard linear solid and its generalized Kelvin chain, not the Zener
+  (Maxwell) form; the 2.x names stay for compatibility.
+
+## [Unreleased]
+
+### Breaking changes
+
+- `copy_parameters`, `apply_parameters`, `copy_constants` and `apply_constants`
+  are removed. `KeyTemplate(src, params, consts).write(dst)` copies the
+  templates and substitutes the keys in one call, about 5 times faster
+  (templates read once, each working file written once).
+- The C++ kernel `umat_damage_weibull` is removed: no UMAT name dispatched to it. The
+  Weibull damage function `damage_weibull` (Functions/damage.hpp, `sim.damage_weibull`) stays.
+- The C++ kernel `umat_prony_Nfast_plastic` (no header, no UMAT name) is removed: Prony
+  viscoelasticity with J2 plasticity is a `ModularMaterial` (MODUL).
+- `tangent_mode=3` / `"closest_point"` is refused with a `ValueError`: the local
+  integrator is now chosen apart, with `integrator="closest_point"`, and
+  `tangent_mode` only takes `"none"`, `"continuum"` or `"algorithmic"`.
+- `integrator="closest_point"` is refused with a `ValueError` for a law without a
+  closest-point update instead of running its cutting plane silently: it serves
+  EPICP, EPCHA, EPJCK, MODUL and its adapters and the SMA kernels (mechanical),
+  EPICP and EPKCP (thermomechanical). A Python law (PYEXT) receives
+  `tangent_mode` 0, 1 or 2 only.
+- The numerical option slots of the props are removed: EPICP `props[6]`
+  (min-NCP) and `props[7]` (closed-form tangent), EPKCP `props[7]`, and
+  `ModularMaterial(ncp=...)`. The complementarity function is the run option
+  `ncp`, and the EPICP closed-form consistent tangent is always used.
+- `condense="approximate"` / `"exact"` are renamed `"spatial_face"` /
+  `"material_face"` (#145).
+- The thermomechanical UMATs return the heat of the increment `R` (storage
+  included), not a rate (#140).
+
+### Added
+
+- The Ansys USERMAT plugin (`UMANS`, `umat_plugin_ans_api`) is dispatched on the small-strain
+  route, as UMABA, and tested by `TUMABA` on the same case. The example plugin is rebuilt on the
+  `fea_transfer` component maps (now public: `fea_components`, `ansys_components`, `fe2vec`,
+  `fe2mat`...): it read the tangent and wrote the rotation row-major (transposed), hid four Wm
+  slots in the USERMAT's statev, and sent condensed states as 3D. `TUMANS` checks the data layout
+  on a USERMAT with a non-symmetric tangent and a non-trivial rotation. The Fortran example did
+  not compile in fixed form. `sim.umat` names the plugins in its error (`is_external_plugin`).
+- `KeyTemplate.write(dst, test=i)`: a constant holds one value per test
+  (`input_values`, e.g. the fibre angle of each specimen) and the value of test
+  `i` is substituted. A constant with a single value warns: it belongs in the
+  template, not in a key.
+- `identification(..., method="least_squares")`: bounded trust-region
+  reflective least squares, parameters scaled to [0, 1] (logarithmically for
+  positive bounds), `n_starts` starting points, `n_threads` for the Jacobian
+  columns or the generations of differential evolution. `calc_residuals`
+  returns the residuals whose sum of squares is `calc_cost`. Differential
+  evolution remains the default.
+- Run options `ncp="fischer_burmeister" | "min"` and
+  `integrator="cutting_plane" | "closest_point"` in `solve`, `sim.umat` and
+  `sim.umat_T`. `min` applies to EPICP, EPKCP and MODUL (also next to Prony and
+  damage mechanisms), not to Armstrong-Frederick back stresses.
+- `simcoon.rotate_voigt`: batch rotation of Voigt stresses, strains and
+  stiffness or concentration tensors by one rotation matrix per point, without
+  the GIL.
+- Native dS/dE of the finite-strain laws built on F, on demand.
+- C++: `cubic_coefficients` and `is_isotropic(L, Tensor4Type)` (constitutive.hpp; the isotropy
+  relation follows the Voigt convention of the type), `spectral_congruence` (eig_sym3.hpp),
+  `add_linearop_eigsym` (contimech.hpp), replacing private copies.
+- `isochoric_invariants` takes a batch, (3,3,N) tensors or (6,N) Voigt columns, as
+  `isochoric_pstretch` does; both share one implementation (parallel, without the GIL).
+- Condensed stress states (`nshr`, plane stress, shells, beams, trusses) in
+  `sim.umat`, the solver and the Abaqus bridges, with `condensation_serves`
+  (#139, #141, #145).
+- Finite strain (NLGEOM) and thermomechanical finite strain in the Abaqus
+  bridges and the solver, `tangent_output` for `umat_T` (#143).
+- LLDM0 ply damage model (Ladevèze–Le Dantec) (#144).
+- STATEV size checks and typed exceptions in the Abaqus UMATs: numerical
+  failures request a step cut, input errors abort (#146).
+- `slice_view`: lock-free aliases of cube slices in parallel loops (#147).
+- GOH 2006 tube-inflation example for HOLZA (#142).
+- `benchmark/`: batch benchmarks and bit-for-bit dumps of the solver outputs.
+
+### Changed
+
+- EPKCP thermomechanical: the cutting plane evaluates the power-law hardening through
+  `PowerLawHardening` (C1 blend below p = 1e-6), as EPICP and its own closest point; for
+  m < 1 the first plastic increment no longer meets an infinite slope.
+- Faster kernels, results bitwise unchanged unless stated: reuse of the
+  ModularUMAT instances, fixed-size 3x3 spectral solves and small dense LU,
+  Eshelby integration and cached homogenization, closed-form
+  Fischer-Burmeister for one and two mechanisms.
+- The Chaboche, composite and MOORI identification examples use the least
+  squares; the JSON-keys and analytical hyperelastic examples keep
+  differential evolution.
+
+### Fixed
+
+- An invalid elastic convention ended the process with `exit(0)` (`M_iso`, `L_cubic`, `M_cubic`,
+  `L_ortho`, `M_ortho`, `L_isotrans`, `M_isotrans`; from Python the interpreter quit with a success
+  code) or, in `L_iso`, returned a zero stiffness: they now throw `std::invalid_argument`
+  (`ValueError` in Python).
+- Small code defects without effect on results: `assert(Y_crit(n))` out of bounds in
+  `Fischer_Burmeister(_limits)` (a Debug build aborted), `variant::build` did not store `g`,
+  the `state_variables_T` default constructor left `Wm`/`Wm_start` to the Armadillo default,
+  an unreachable `HfR = 0.5*a2* -a3` branch of the thermomechanical unified_T closest point.
+- Copies of `step_meca` / `step_thermomeca` lost their loading: the copy constructor set
+  `BC_mecas` from `mecas`, the assignment skipped `BC_mecas`, `BC_Ts`, `control_type` and
+  `tab_data`. The copy operations of `step`, `step_meca` and `step_thermomeca` are now the
+  memberwise defaults (test `Tstep`). Nothing in simcoon copied a step.
+- Under finite strain the solver ran UMEXT as a no-op (unchanged, zero stress) and failed on
+  UMABA with "Unknown umat name": the plugins are now refused with an explicit error.
+- `Eshelby_penny` set the in-plane shear S_2323 (Voigt 23) to 1 and S_1212 to 0: now
+  S_1212 = S_1313 = 1/2 (1 in engineering Voigt) and S_2323 = 0, the flat oblate limit.
+  No homogenization scheme called it.
+- `d2lagrange_pow_1` placed `p0` wrongly (exact only for p0 = 1); its test now checks it
+  against a finite difference of `dlagrange_pow_1`. No law called it.
+- Documentation: .cpp comments cut to short implementation notes, contracts (props,
+  statev, theory) moved to the header Doxygen, dead code and wrong `@file` banners removed;
+  the EPCHA header example read the back-strains as the backstresses (statev 20-31).
+- `isochoric_pstretch` ignored `J` on a single point (it recomputed it from the
+  tensor), unlike its batch form.
+- Race on the Eshelby quadrature tables between concurrent solves.
+- MIHEN crash (null pointer) and MISCN failure on the second block of a
+  loading path.
+- `n_threads` only took effect for 1 thread: it is now honoured under GCD and
+  OpenMP, and `Lt_convert` takes it too.
+
+## [2.2.0] - 2026-10-14
+
+Tensor bases (#126, #127, #128), cached rotation operators (#130), Eshelby memo
+and lazy fields (#129), closest-point return mapping (#131, #133), Johnson-Cook
+(#85), analytic SMA tangents (#136), logarithmic U (#137). See the
+[release notes](https://github.com/3MAH/simcoon/releases/tag/2.2.0).
+
+## [2.1.0] - 2026-10-01
+
+See the [release notes](https://github.com/3MAH/simcoon/releases/tag/2.1.0).
+
+## [2.0.1] - 2026-09-09
+
+See the [release notes](https://github.com/3MAH/simcoon/releases/tag/2.0.1).
+
+## [2.0.0] - 2026-09-07
+
+See the [release notes](https://github.com/3MAH/simcoon/releases/tag/2.0.0).
+
+[Unreleased]: https://github.com/3MAH/simcoon/compare/2.2.0...HEAD
+[2.2.0]: https://github.com/3MAH/simcoon/compare/2.1.0...2.2.0
+[2.1.0]: https://github.com/3MAH/simcoon/compare/2.0.1...2.1.0
+[2.0.1]: https://github.com/3MAH/simcoon/compare/2.0.0...2.0.1
+[2.0.0]: https://github.com/3MAH/simcoon/compare/2.0.0b1...2.0.0

CMakeLists.txt

@@ -371,14 +371,29 @@ if(SIMCOON_BUILD_TESTS)
   if(NOT MSVC)
     # Fortran is only needed for Abaqus UMAT plugin tests
     enable_language(Fortran)
-    set_source_files_properties(testBin/Umats/UMABA/external/UMAT_ABAQUS_ELASTIC.f PROPERTIES Fortran_PREPROCESS OFF)
-    add_library(umat_plugin_aba SHARED testBin/Umats/UMABA/external/UMAT_ABAQUS_ELASTIC.f testBin/Umats/UMABA/external/umat_plugin_aba.cpp)
+    set_source_files_properties(external/UMAT_ABAQUS_ELASTIC.f PROPERTIES Fortran_PREPROCESS OFF)
+    add_library(umat_plugin_aba SHARED external/UMAT_ABAQUS_ELASTIC.f external/umat_plugin_aba.cpp)
     # Use default PREFIX (lib) and SUFFIX (.so on Linux, .dylib on macOS)
     # to match dylib::decorations::os_default() expectations
     set_target_properties(umat_plugin_aba PROPERTIES
       LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/testBin/Umats/UMABA/external
     )
     target_link_libraries(umat_plugin_aba PRIVATE simcoon ${ARMADILLO_LIBRARIES})
+
+    # Ansys USERMAT plugin, built from the external/ example: TUMABA runs it on the same case
+    set_source_files_properties(external/USERMAT_ANSYS_ELASTIC.for PROPERTIES Fortran_PREPROCESS OFF)
+    add_library(umat_plugin_ans SHARED external/USERMAT_ANSYS_ELASTIC.for external/umat_plugin_ans.cpp)
+    set_target_properties(umat_plugin_ans PROPERTIES
+      LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/testBin/Umats/UMABA/external
+    )
+    target_link_libraries(umat_plugin_ans PRIVATE simcoon ${ARMADILLO_LIBRARIES})
+
+    # the same plugin on a USERMAT that exposes what it receives (TUMANS: data layout)
+    add_library(umat_plugin_ans_layout SHARED external/umat_plugin_ans.cpp test_extern/Umats/UMANS/usermat_layout.cpp)
+    set_target_properties(umat_plugin_ans_layout PROPERTIES OUTPUT_NAME umat_plugin_ans
+      LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/testBin/Umats/UMANS/external
+    )
+    target_link_libraries(umat_plugin_ans_layout PRIVATE simcoon ${ARMADILLO_LIBRARIES})
   endif()
 
   add_library(umat_plugin_ext SHARED