Experimental binary wheel builder

[jhale]

This introduces the capability to build relocatable DOLFINx ABI3 binary wheels for Python >= 3.12 without PETSc/petsc4py.

An MPICH-compatible MPI implementation must be installed on the system. ABI compatibility initiative will eventually remove this limitation allowing OpenMPI to be used as well.

I will post a test set of packages with instructions here in due course.

[jhale]

gdb --args python -m pytest -k test_write_point_data python/test/unit/io/test_vtkhdf.py

#0  0x0000ffff8edd7d40 in ?? () from /lib/aarch64-linux-gnu/libc.so.6
#1  0x0000ffff8ed86900 [PAC] in raise () from /lib/aarch64-linux-gnu/libc.so.6
#2  0x0000ffff8ed71a44 [PAC] in abort () from /lib/aarch64-linux-gnu/libc.so.6
#3  0x0000ffff80e50984 [PAC] in std::__glibcxx_assert_fail(char const*, int, char const*, char const*) ()
   from /usr/local/lib/python3.13/site-packages/dolfinx/../fenics_dolfinx.libs/libdolfinx-e9eb1af4.so.0.11.0.0
#4  0x0000ffff81512f98 in std::vector<unsigned long long, std::allocator<unsigned long long> >::operator[] (this=0xffffd0750bb0, __n=1)
    at /opt/rh/gcc-toolset-14/root/usr/include/c++/14/bits/stl_vector.h:1149
#5  0x0000ffff815162b8 in dolfinx::io::hdf5::write_dataset<int> (file_handle=72057594037927937, dataset_path="/VTKHDF/Steps/CellOffsets", data=0xffffd0750d7c, range=...,
    global_size=std::vector of length 1, capacity 1 = {...}, use_mpi_io=true, use_chunking=true) at /usr/local/include/dolfinx/io/HDF5Interface.h:201
#6  0x0000ffff81505a70 in dolfinx::io::VTKHDF::write_data<float>(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, dolfinx::mesh::Mesh<float> const&, std::vector<float, std::allocator<float> > const&, double)::{lambda<typename $T0>(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, $T0)#1}::operator()<int>(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, int) const (__closure=0xffffd0750ef0, dset_name="/VTKHDF/Steps/CellOffsets", value=0) at /usr/local/include/dolfinx/io/VTKHDF.h:236
#7  0x0000ffff815062f8 in dolfinx::io::VTKHDF::write_data<float> (point_or_cell="Point", filename="point_data.vtkhdf", mesh=...,
    data=std::vector of length 36, capacity 36 = {...}, time=0) at /usr/local/include/dolfinx/io/VTKHDF.h:241

Looks like an issue in DOLFINx (or the test) triggering a standard library assert.

@pytest.mark.parametrize("dtype", [np.float32, np.float64])
def test_write_point_data(dtype):
    mesh = create_unit_square(MPI.COMM_WORLD, 5, 5, dtype=dtype)
    filename = "point_data.vtkhdf"
    write_mesh(filename, mesh)
    point_data = np.arange(mesh.geometry.index_map().size_local)
    for j in range(3):
        write_point_data(filename, mesh, point_data, float(j))

[jhale]

I now have a partial test suite pass with these wheels using:

mpiexec -n 3 python -m pytest -v -k "not write_cell_data and not write_point_data and not write_mixed_topology_data" -m "not petsc4py and not adios2" python/test/unit/

[jhale]

This is now fully tested, e.g. pip install mpich fenics-dolfinx works.

https://github.com/FEniCS/dolfinx/actions/runs/21290820557/job/61283653443

A single binary wheel is compatible for all Python versions >= 3.12 and with any MPICH-compatible MPI; we're testing MPICH and Intel MPI wheels and system-provided MPICH. This is technically possible thanks to some nice work done recently in mpi4py.

When MPI5 is standardised and implemented it will be possible to provide a single wheel for all MPI5 ABI-compliant implementations.

It should be relatively straightforward to build wheels for macOS, and even Windows, as Intel provide impi-rt wheels there.

In terms of solvers, as we do not build against PETSc, an interface to MUMPS for our own MatrixCSR class would be a big quality of life improvement, and good enough for most users to get started with.

[awensaunders]

Hi there, I know this PR is closed & merged now - is there any intention of uploading these wheels to pypi so those of us using non-conda package managers such as uv can just grab a wheel?