Skip to content

Guard CPUID/rdtsc for non-x86 and add stubs - #182

Open
chemiskyy wants to merge 1 commit into
romeric:masterfrom
3MAH:master
Open

chemiskyy wants to merge 1 commit into
romeric:masterfrom
3MAH:master

Conversation

@chemiskyy

Copy link
Copy Markdown

Proposed fix for ARM:

Fastor/config/cpuid.h — The CPUID class constructor uses x86 cpuid asm in an unguarded #else. Fix: gate it with #elif defined(x86_64) || defined(i386)
and add an #else that zeros the registers on ARM.

Fastor/util/timeit.h — The rdtsc() / rdtsc_begin() / rdtsc_end() functions use x86 rdtsc/rdtscp asm in an unguarded #else. Fix: same #elif x86 gate, plus ARM
stubs that return 0.

Add explicit x86/x86_64 guards and non-x86 fallbacks for CPUID and rdtsc. In Fastor/config/cpuid.h, avoid using inline asm on non-x86 targets by checking for __x86_64__/__i386__ and provide a safe zeroed fallback for regs otherwise. In Fastor/util/timeit.h, restrict the GCC inline-asm rdtsc implementation to x86 and add ARM/non-x86 stub implementations that return 0. These changes prevent build failures on non-x86 platforms and make the codebase more portable.
mitrlk added a commit to mitrlk/mitrlk-Marmot that referenced this pull request Aug 10, 2026
Mirrors build_ubuntu, but on GitHub's Apple Silicon runner (macos-latest is
arm64 since macos-14). This exercises clang/libc++ rather than GCC/libstdc++,
which is what makes the three fixes in this PR necessary — all of them are
accepted silently by libstdc++ and rejected or trapped by libc++.

Without this job the same problems can reappear unnoticed, since nothing else
in CI compiles Marmot with libc++.

CC/CXX are pinned to clang/clang++ rather than left to CMake's default: the
macOS runners also ship Homebrew GCC, and silently picking that up would build
against libstdc++ and defeat the purpose of the job while still reporting
green. No compiler version is pinned, so the job tracks whatever Apple clang
the runner provides. A first step prints the toolchain so the log records
which compiler actually ran.

One deviation from build_ubuntu: Fastor is cloned from an arm64-compatible
fork rather than upstream V0.6.4. Upstream Fastor does not compile on arm64 —
it reaches x86 `cpuid` and `rdtsc` inline assembly on every non-Windows
platform, and its scalar fallback backend has further defects that only
surface off x86. The fork is upstream `master` plus the minimal patch fixing
exactly that, and nothing else. Those fixes are offered upstream in
romeric/Fastor#182 and #183; the step carries a comment to switch back once
either is merged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
mitrlk added a commit to mitrlk/mitrlk-Marmot that referenced this pull request Aug 10, 2026
Mirrors build_ubuntu, but on GitHub's Apple Silicon runner (macos-latest is
arm64 since macos-14). This exercises clang/libc++ rather than GCC/libstdc++,
which is what makes the three fixes in this PR necessary — all of them are
accepted silently by libstdc++ and rejected or trapped by libc++.

Without this job the same problems can reappear unnoticed, since nothing else
in CI compiles Marmot with libc++.

Deviations from build_ubuntu, each deliberate:

* The autodiff build directory is called `build-cmake`, not `build`. autodiff
  ships a Bazel `BUILD` file in its repository root, and the macOS filesystem
  is case-insensitive, so `mkdir build` fails with "File exists" and `cd build`
  with "Not a directory". On Linux the two names are distinct, which is why
  build_ubuntu is unaffected.

* `shell: bash -leo pipefail {0}` rather than `bash -l {0}`. Specifying a
  shell replaces GitHub's default `bash -e {0}`, so `-l` alone silently drops
  errexit and only the last command of a `run:` block decides the step result.
  Since the dependency steps end in `cd ../..`, the autodiff failure above was
  reported as success and only surfaced later as a confusing "autodiff include
  directory not found". The same pattern is present in build_ubuntu, where it
  is currently latent because those steps all succeed.

* `ctest --no-tests=error`. ctest exits 0 when no tests exist, so a build that
  produced no test binaries at all still reported success.

* CC/CXX pinned to clang/clang++ rather than left to CMake's default: the
  macOS runners also ship Homebrew GCC, and silently picking that up would
  build against libstdc++ and defeat the purpose of the job while still
  reporting green. No compiler version is pinned, so the job tracks whatever
  Apple clang the runner provides. A first step prints the toolchain so the
  log records which compiler actually ran.

* Fastor is cloned from an arm64-compatible fork rather than upstream V0.6.4.
  Upstream Fastor does not compile on arm64 — it reaches x86 `cpuid` and
  `rdtsc` inline assembly on every non-Windows platform, and its scalar
  fallback backend has further defects that only surface off x86. The fork is
  upstream `master` plus the minimal patch fixing exactly that, and nothing
  else. Those fixes are offered upstream in romeric/Fastor#182 and #183; the
  step carries a comment to switch back once either is merged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
alexdummer pushed a commit to MAteRialMOdelingToolbox/Marmot that referenced this pull request Aug 17, 2026
* fix: use a shift instead of non-constexpr std::pow for nNodesLinear

std::pow is not constexpr in the C++ standard. libstdc++ provides a
non-standard constexpr overload, so this compiles on the Linux/GCC CI, but
libc++ (macOS/clang, incl. arm64) rejects it:

  error: constexpr variable 'nNodesLinear' must be initialized by a
         constant expression

2^nDim is exactly a shift for the dimensions in use, so use 1 << nDim.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* test: fix two latent bugs in the test suite exposed on arm64 macOS

Both tests pass on the Ubuntu CI (GCC/libstdc++) but are defective; one of
them fails outright with clang/libc++ on Apple Silicon. The library code is
correct in both cases — only the tests are at fault.

1. TestMarmotVoigt — the four dJ2/dJ3 stress/strain derivative checks compare
   an analytic derivative against a forward-difference approximation, but omit
   the tolerance argument to checkIfEqual(), so they inherit the default
   tol = 1e-15. A first-order forward difference with a 1e-8 step cannot meet
   that; the checks only pass where the two happen to round identically.
   On arm64/clang they do not:

     Hint: a = 17.6068 !=  b = 17.6068 ( tol = 1e-15 )
     Exception: void test_dJ2_dStress() failed

   Reproduced 5/5 runs. The sibling finite-difference checks in the same file
   already pass 1e-8 explicitly; this makes these four consistent with them.

2. TestMarmotFiniteStrainPlasticity — the reference tensor DexMapexpect is a
   Fastor::Tensor<double,3,3,3,3>, which is not zero-initialized, and the test
   assigns only the non-zero components. The remaining 80 entries are read
   uninitialized, which is undefined behaviour.

   This one currently passes, because a fresh stack happens to be zeroed. It
   is luck, not correctness: dirtying the stack before the tensor is
   constructed leaves all 80 unset entries holding garbage (values up to
   1e29). Added an explicit .zeros() so the reference is well-defined.

Verified on macOS 15 / Apple M4, clang 20 (conda-forge
arm64-apple-darwin20.0.0-clang++), -std=c++20 -stdlib=libc++, Release:
before 45/46 ctests pass, after 46/46.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix: give MarmotMaterialGeneralGradientEnhancedHypoElastic a virtual destructor

The class is abstract and instances are owned and destroyed through
base-class pointers — GeneralGradientEnhancedDisplacementFiniteElement holds
one per quadrature point as

  std::unique_ptr< MarmotMaterialGeneralGradientEnhancedHypoElastic< N > > material;

Destroying a derived object through a base pointer whose destructor is not
virtual is undefined behaviour ([expr.delete]/3). GCC/libstdc++ happens to
call the wrong destructor silently, which is why the Ubuntu CI never noticed;
clang diagnoses it and, because the UB is provable, emits a trap instead of a
call, so the process dies at the point of destruction.

Reproduced on macOS 15 / Apple M4, clang 20, -std=c++20 -stdlib=libc++, in
five lines against an otherwise unmodified tree:

  double props[] = {20000., 0.2, 1.0, 0.1, 1.0, 1.0};
  auto* m = MarmotLibrary::MarmotMaterialGeneralGradientEnhancedHypoElasticFactory<1>
              ::createMaterial("AT2PHASEFIELD", props, 6, 1);
  delete m;

  warning: delete called on 'MarmotMaterialGeneralGradientEnhancedHypoElastic<1>'
           that is abstract but has non-virtual destructor
           [-Wdelete-abstract-non-virtual-dtor]
  $ ./repro
  material ptr = 0x121804320
  [SIGTRAP, exit 133]

With this change the warning disappears and the same program exits 0. Note
that the existing TestAT2PhaseField ctest does not catch this — it passes
either way, because it never destroys a material through the base pointer.

Three sibling classes are abstract with a non-virtual destructor for the same
reason and are latent rather than currently broken, since nothing in-tree
deletes them through a base pointer yet: MarmotMaterialFiniteStrainAD,
MarmotMaterialHypoElasticAD, and MarmotMaterialPointSolver{FiniteStrain,
HypoElastic}. Happy to fix those in this PR too if preferred.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* ci: add a macOS/arm64 build

Mirrors build_ubuntu, but on GitHub's Apple Silicon runner (macos-latest is
arm64 since macos-14). This exercises clang/libc++ rather than GCC/libstdc++,
which is what makes the three fixes in this PR necessary — all of them are
accepted silently by libstdc++ and rejected or trapped by libc++.

Without this job the same problems can reappear unnoticed, since nothing else
in CI compiles Marmot with libc++.

Deviations from build_ubuntu, each deliberate:

* The autodiff build directory is called `build-cmake`, not `build`. autodiff
  ships a Bazel `BUILD` file in its repository root, and the macOS filesystem
  is case-insensitive, so `mkdir build` fails with "File exists" and `cd build`
  with "Not a directory". On Linux the two names are distinct, which is why
  build_ubuntu is unaffected.

* `shell: bash -leo pipefail {0}` rather than `bash -l {0}`. Specifying a
  shell replaces GitHub's default `bash -e {0}`, so `-l` alone silently drops
  errexit and only the last command of a `run:` block decides the step result.
  Since the dependency steps end in `cd ../..`, the autodiff failure above was
  reported as success and only surfaced later as a confusing "autodiff include
  directory not found". The same pattern is present in build_ubuntu, where it
  is currently latent because those steps all succeed.

* `ctest --no-tests=error`. ctest exits 0 when no tests exist, so a build that
  produced no test binaries at all still reported success.

* CC/CXX pinned to clang/clang++ rather than left to CMake's default: the
  macOS runners also ship Homebrew GCC, and silently picking that up would
  build against libstdc++ and defeat the purpose of the job while still
  reporting green. No compiler version is pinned, so the job tracks whatever
  Apple clang the runner provides. A first step prints the toolchain so the
  log records which compiler actually ran.

* Fastor is cloned from an arm64-compatible fork rather than upstream V0.6.4.
  Upstream Fastor does not compile on arm64 — it reaches x86 `cpuid` and
  `rdtsc` inline assembly on every non-Windows platform, and its scalar
  fallback backend has further defects that only surface off x86. The fork is
  upstream `master` plus the minimal patch fixing exactly that, and nothing
  else. Those fixes are offered upstream in romeric/Fastor#182 and #183; the
  step carries a comment to switch back once either is merged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* ci: make build_ubuntu fail on errors instead of reporting green

Same hardening as the new macOS job, applied to the existing Linux one. No
functional change while everything passes — this only affects what happens
when something breaks.

Two ways the job could report success without having built anything:

* Specifying `shell:` replaces GitHub's default `bash -e {0}`, so `bash -l {0}`
  silently drops errexit. Only the last command of a `run:` block then decides
  the step result, and the three dependency steps end in `cd ../..`, which
  always succeeds. A failed clone, cmake or `make install` was therefore
  reported as a passing step, and only surfaced later as a confusing
  "include directory not found" from Marmot's own configure — or not at all.

* `ctest` exits 0 when no tests exist, so a build that produced no test
  binaries still passed.

This was not hypothetical: the macOS job hit exactly this and reported success
having compiled nothing. On Linux the steps all currently succeed, so the
problem is latent rather than active, but the failure mode is the same.

Uses `bash -leo pipefail {0}` and `ctest --no-tests=error`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
mitrlk added a commit to mitrlk/mitrlk-EdelweissFE that referenced this pull request Aug 31, 2026
Mirrors run_tests_with_marmot.yml on GitHub's Apple Silicon runner, so that
arm64 support cannot regress unnoticed. Nothing else in CI creates the conda
environment, builds Marmot, builds the Cython extensions or runs the suite on
arm64 with clang/libc++.

Four deviations from the Linux job, each unavoidable:

* Installs from conda_requirements_osx-arm64.txt; see that file for why
  conda_requirements.txt cannot be used on this platform.

* autodiff is built in `build-cmake`, not `build`. autodiff ships a Bazel
  `BUILD` file in its root, and the macOS filesystem is case-insensitive, so
  `mkdir build` fails with "File exists" and `cd build` with "Not a directory".

* Fastor is cloned from an arm64-compatible fork. Upstream Fastor reaches x86
  `cpuid` and `rdtsc` inline assembly on every non-Windows platform and does
  not compile on arm64. The fork is upstream `master` plus that minimal patch
  and nothing else; the fixes are offered upstream in romeric/Fastor#182 and
  Edelweiss-Numerics#183, and the step carries a comment to switch back once either is merged.
  Marmot's own build_macos.yml uses the same fork for the same reason.

* `shell: bash -leo pipefail {0}` rather than `bash -l {0}`. Specifying a
  shell replaces GitHub's default `bash -e {0}`, so `-l` alone drops errexit
  and only the last command of a `run:` block decides the step result. That
  matters directly in the final step: with two `run_tests_edelweissfe`
  invocations and no errexit, a completely failing Marmot suite is reported as
  success as long as the edelweiss-only suite passes. The dependency steps have
  the same problem, since each ends in `cd ../..`, which always succeeds.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
mitrlk added a commit to mitrlk/mitrlk-EdelweissFE that referenced this pull request Aug 31, 2026
Mirrors run_tests_with_marmot.yml on GitHub's Apple Silicon runner, so that
arm64 support cannot regress unnoticed. Nothing else in CI creates the conda
environment, builds Marmot, builds the Cython extensions or runs the suite on
arm64 with clang/libc++.

Five deviations from the Linux job, each unavoidable:

* Installs from conda_requirements_osx-arm64.txt; see that file for why
  conda_requirements.txt cannot be used on this platform.

* Every dependency is cloned into the parent directory rather than into the
  checkout. Marmot ships a committed symlink, _codeql_detected_source_root
  -> ., so a Marmot clone inside the checkout makes `pip install .` follow that
  symlink into itself indefinitely while collecting files for the wheel:

      error: could not create 'build/.../Marmot/_codeql_detected_source_root/
      _codeql_detected_source_root/...': File name too long

  Linux tolerates this only because its path limit is four times macOS's; the
  recursion is there either way, and the dependency sources do not belong in
  the wheel regardless. The Eigen step already used this pattern.

* autodiff is built in `build-cmake`, not `build`. autodiff ships a Bazel
  `BUILD` file in its root, and the macOS filesystem is case-insensitive, so
  `mkdir build` fails with "File exists" and `cd build` with "Not a directory".

* Fastor is cloned from an arm64-compatible fork. Upstream Fastor reaches x86
  `cpuid` and `rdtsc` inline assembly on every non-Windows platform and does
  not compile on arm64. The fork is upstream `master` plus that minimal patch
  and nothing else; the fixes are offered upstream in romeric/Fastor#182 and
  Edelweiss-Numerics#183, and the step carries a comment to switch back once either is merged.
  Marmot's own build_macos.yml uses the same fork for the same reason.

* `shell: bash -leo pipefail {0}` rather than `bash -l {0}`, for the reason
  given in the following commit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
alexdummer added a commit to Edelweiss-Numerics/EdelweissFE that referenced this pull request Sep 27, 2026
* fix: add a conda requirements file for osx-arm64

conda_requirements.txt cannot be used on Apple Silicon. Two of its pins have
no package for this platform, so environment creation fails before anything is
installed.

1. anaconda::mkl, anaconda::mkl-include and anaconda::intel-openmp have no
   osx-arm64 build:

       PackagesNotFoundError: The following packages are not available from
       current channels:
         - intel-openmp

   The only capability lost by dropping them is PARDISO, which is part of MKL
   and therefore unavailable on this platform regardless. getDefaultLinSolver()
   already falls back to SciPy's superlu when the PARDISO extension is absent,
   so the test suite and any input file that does not name a linear solver
   explicitly are unaffected.

2. python-freethreading is dropped, which is the less obvious one. VTK
   publishes free-threaded wheels for Linux only: vtk 9.7.0 ships
   cp314-cp314t-manylinux_2_28_aarch64 and cp314-cp314t-manylinux2014_x86_64,
   but there is no cp314t macOS wheel of any architecture, and conda-forge's
   vtk likewise requires python_abi *_cp314. Keeping python-freethreading would
   therefore make `pip install -r pip_requirements.txt` fail on vtk and take
   pyvista with it.

   On macOS the choice is thus between free-threading and pyvista/VTK. This
   file picks pyvista, so that the suite runs complete rather than losing the
   one active test that needs it
   (edelweiss-only/NodeToDiscreteRigidBodyContact). Free-threading currently
   buys nothing here in any case: the Cython extensions do not declare
   free-threading support, so CPython re-enables the GIL as soon as they are
   imported. Happy to invert this trade if you would rather keep the
   interpreter identical across platforms and let that test fail.

Everything else is identical to conda_requirements.txt, so the two files stay
easy to diff against each other.

Verified on macOS 15 / Apple M4: conda_requirements.txt fails to solve as
above, this file solves cleanly and resolves to the regular cp314 interpreter.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* ci: run the test suite on macOS/arm64

Mirrors run_tests_with_marmot.yml on GitHub's Apple Silicon runner, so that
arm64 support cannot regress unnoticed. Nothing else in CI creates the conda
environment, builds Marmot, builds the Cython extensions or runs the suite on
arm64 with clang/libc++.

Five deviations from the Linux job, each unavoidable:

* Installs from conda_requirements_osx-arm64.txt; see that file for why
  conda_requirements.txt cannot be used on this platform.

* Every dependency is cloned into the parent directory rather than into the
  checkout. Marmot ships a committed symlink, _codeql_detected_source_root
  -> ., so a Marmot clone inside the checkout makes `pip install .` follow that
  symlink into itself indefinitely while collecting files for the wheel:

      error: could not create 'build/.../Marmot/_codeql_detected_source_root/
      _codeql_detected_source_root/...': File name too long

  Linux tolerates this only because its path limit is four times macOS's; the
  recursion is there either way, and the dependency sources do not belong in
  the wheel regardless. The Eigen step already used this pattern.

* autodiff is built in `build-cmake`, not `build`. autodiff ships a Bazel
  `BUILD` file in its root, and the macOS filesystem is case-insensitive, so
  `mkdir build` fails with "File exists" and `cd build` with "Not a directory".

* Fastor is cloned from an arm64-compatible fork. Upstream Fastor reaches x86
  `cpuid` and `rdtsc` inline assembly on every non-Windows platform and does
  not compile on arm64. The fork is upstream `master` plus that minimal patch
  and nothing else; the fixes are offered upstream in romeric/Fastor#182 and
  #183, and the step carries a comment to switch back once either is merged.
  Marmot's own build_macos.yml uses the same fork for the same reason.

* `shell: bash -leo pipefail {0}` rather than `bash -l {0}`, for the reason
  given in the following commit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* ci: make the existing workflows fail on errors instead of reporting green

Same change Marmot took in MAteRialMOdelingToolbox/Marmot#57, applied here.
No functional change while everything passes — it only affects what happens
when a step breaks.

Specifying `shell:` replaces GitHub's default `bash -e {0}`, so `bash -l {0}`
silently drops errexit. Bash then continues past a failed command and only the
last command of a `run:` block decides the step result.

In run_tests_with_marmot.yml this is not hypothetical. The final step is:

    run_tests_edelweissfe ./testfiles/marmot/
    run_tests_edelweissfe ./testfiles/edelweiss-only/

run_tests_edelweissfe exits 1 when a test fails, but without errexit a
completely failing Marmot suite is still reported as success, as long as the
edelweiss-only suite passes afterwards. The dependency steps have the same
problem in the other direction: each ends in `cd ../..`, which always
succeeds, so a failed clone, cmake or `make install` passes and only surfaces
later as a confusing error from an unrelated step.

Switches run_tests_with_marmot.yml, run_tests_without_marmot.yml and
sphinx.yml to `bash -leo pipefail {0}`: errexit, plus pipefail so a failure in
the left-hand side of a pipe is not masked by a succeeding right-hand side.
The new macOS workflow already uses it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(csrgeneratorv2): declare mapStarts as int64_t to match the C++ header

CSRDirectAssembler::registerEntities takes `const int64_t* mapStarts`
(_csrcore.h), but csrgeneratorv2.pyx declared the parameter, and the
corresponding memoryview, as `long`.

On the Linux x86-64 CI these are the same type, since int64_t is a typedef for
long there, so the mismatch is invisible. On macOS/arm64 int64_t is a typedef
for long long instead. Both are 64 bits wide, but they are distinct types, and
a `long*` does not implicitly convert to a `const long long*`:

    csrgeneratorv2.cpp: error: cannot initialize a parameter of type
    'const int64_t *' (aka 'const long long *') with an rvalue of type 'long *'
    [FAIL] Could not build edelweissfe.numerics.csrgeneratorv2

setup.py treats a failed extension as non-fatal, so this did not fail the
build. It failed at run time instead, and quietly: every single test in both
suites skipped with

    SKIPPED: No module named 'edelweissfe.numerics.csrgeneratorv2'

while the runner still reported "Tests failed: 0".

Declaring the parameter as int64_t matches the header exactly and is a no-op on
platforms where int64_t is long. It also matches what callers already pass,
np.int64 arrays, and removes a latent problem on Windows, where `long` is only
32 bits.

Found by the macOS/arm64 job added in this PR, on its first run.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(solvers): make the existing PARDISO-to-SuperLU fallback reachable

`getDefaultLinSolver()` already falls back to SuperLU when the PARDISO extension is
absent, but nothing ever called it: the schema seeds `"linsolver": "pardiso"`, so the
`if "linsolver" in self.options` guard was always true. On a build without PARDISO
(any non-x86 platform, since it needs MKL) every test using the default solver raised
ImportError and was reported as skipped.

Default to `""` and test the value instead of the key, matching the convention
`linsolverConfigFile` already uses next to it. NEST had a second hard-coded call site
and needed the same treatment.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q9cV34Yd7t5gR6EYEWghkh

* fix(mpc): keep diagonal slots after the MPC transformation

SciPy's sparse product drops stored zeros, so Dirichlet DOFs without stiffness
lost their diagonal slot and the matrix became singular.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GeddoUWB6H3gmRug1mj5dN

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Alexander Dummer <alexander.dummer@uibk.ac.at>
Co-authored-by: Alexander Dummer <62308023+alexdummer@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant