Native 64-bit Windows conda packages for FEniCSx, including DOLFINx, PETSc, petsc4py, HDF5, and MPI support.
These packages are built primarily for use with FEATool Multiphysics, providing its native Windows FEniCSx/PETSc backend, but they can also be installed and used directly from Python without WSL, Docker, or a Linux virtual machine.
Packages are published on the precise-simulation Anaconda channel.
- Windows 10 or Windows 11, 64-bit
- Miniforge, Miniconda, or Anaconda
- CPython 3.12, 3.13, 3.14, or 3.15
Create a dedicated conda environment and install DOLFINx:
conda create -n fenics-windows python=3.12
conda activate fenics-windows
conda install -c precise-simulation -c conda-forge "libblas=*=*openblas" fenics-dolfinxPython 3.12, 3.13, 3.14, and 3.15 are supported. The packages use Python's stable ABI, so the same DOLFINx binary can serve all supported interpreter versions.
The install pulls in the required Windows builds of PETSc, petsc4py, HDF5, Intel MPI, mpi4py, Basix, FFCx, UFL, PT-Scotch, MUMPS, OpenBLAS, and their dependencies.
Check that DOLFINx and PETSc can be imported:
python -c "import dolfinx, petsc4py; print('DOLFINx', dolfinx.__version__); print('petsc4py', petsc4py.__version__)"Then create a small mesh:
python -c "from mpi4py import MPI; from dolfinx import mesh; mesh.create_unit_square(MPI.COMM_SELF, 8, 8); print('DOLFINx installation OK')"If both commands complete successfully, the environment is ready for normal serial FEniCSx use.
A minimal example:
from mpi4py import MPI
from dolfinx import mesh
domain = mesh.create_unit_square(MPI.COMM_WORLD, 16, 16)
print(
f"Rank {MPI.COMM_WORLD.rank}: "
f"{domain.topology.index_map(domain.topology.dim).size_local} cells"
)Save this as test.py and run it serially:
python test.pyA more complete Poisson example is included in this repository as scripts/test-poisson.py:
python scripts\test-poisson.pyThe packages include Intel MPI and support multi-process DOLFINx runs.
Run the example above with two local MPI ranks:
mpiexec -localonly -n 2 python test.pyor run the bundled Poisson example:
mpiexec -localonly -n 2 python scripts\test-poisson.py-localonly restricts process launching to the current computer.
Serial DOLFINx use does not start mpiexec and normally requires no Windows Firewall changes.
Multi-process runs use Intel MPI's Hydra launcher and may trigger a Windows Firewall prompt for mpiexec.exe or one of the Hydra proxy executables. On managed systems, an administrator or IT deployment may need to allow these executables through the firewall.
Users who only need serial FEniCSx can use the packages without configuring MPI launcher firewall rules.
Do not set:
I_MPI_FABRICS=shm
Current Intel MPI versions select local shared-memory transport automatically. If an older configuration has this variable set, clear it with:
set I_MPI_FABRICS=The Windows stack includes:
- DOLFINx
- PETSc and petsc4py
- Intel MPI and mpi4py
- parallel HDF5 with local HDF5/XDMF support
- Basix
- FFCx
- UFL
- PT-Scotch mesh partitioning
- MUMPS sparse direct solver
- OpenBLAS
Parallel HDF5/XDMF files and distributed DOLFINx computations are supported.
PETSc's MUMPS direct solver is available, for example:
ksp.setType("preonly")
pc = ksp.getPC()
pc.setType("lu")
pc.setFactorSolverType("mumps")The packages use the 32-bit-integer PT-Scotch build, which has been validated for DOLFINx mesh partitioning on Windows.
When constructing distributed meshes manually, do not provide an identical complete copy of the global cell array independently on every MPI rank. Supply the mesh data on the root rank or distribute the input appropriately.
For a strict rank-per-core MPI execution model, set:
set OPENBLAS_NUM_THREADS=1This is optional for normal serial use.
The repository also contains a bridge-oriented packaging path for embedding the qualified Windows FEniCS stack into a separate CPython 3.12 host runtime.
.\scripts\package-embedded-runtime.ps1 -Prefix <fenics-conda-prefix> -Output .\dist
.\scripts\test-embedded-runtime.ps1 -Zip .\dist\fenics-embed-runtime-v0.1.0-win-x64-cp312.zip -Python <python-3.12.10.exe>
The resulting ZIP is an overlay, not a Python installation. It preserves the package-owned Lib/site-packages, native Library/bin closure, Library/fenics-jit runtime/backend, CPython headers, runtime data and provenance needed by DOLFINx/PETSc/FFCx. The fenics_embed_runtime.activate() bootstrap relocates DLL, PETSc and JIT roots against the embedding runtime's sys.prefix.
Hosted qualification merges the exact overlay into a clean non-conda CPython 3.12.10 runtime, removes ambient compiler/SDK state, then performs a fresh micro-Clang FFCx Poisson JIT and PETSc solve. This payload is intended for the optional FEniCS runtime profile of the Python MATLAB/Octave bridge; normal conda installation remains the primary direct-Python distribution path.
The historical embedded-runtime-macos-prototype workflow is retained as a
manual diagnostic for the first qualification stage of macOS arm64 and Rosetta
x86_64 overlays. It
creates a FEniCS 0.11 / NumPy 2.5.2 source environment for each architecture
and retains the solver's exact package transaction for subsequent pinning. It
downloads the exact official Zig 0.15.2 toolchain for that
architecture, poisons ambient compiler/SDK discovery, and requires a fresh FFCx
JIT plus serial PETSc Poisson solve through a zig cc wrapper targeting
macOS 14.0.
The successful #27 transaction is frozen in the retained package locks and Zig
checksums. Fresh conda solves can drift as the channel changes, so this
historical workflow is not a pull-request acceptance gate. The
embedded-runtime-macos-overlay workflow below replays the immutable
transactions without solving and owns current producer acceptance.
The embedded-runtime-macos-overlay workflow replays those checked-in locks as
solver-free @EXPLICIT transactions using the exact URL ordering retained by
the qualified prototype and micromamba 2.10.0. Each exact URL is SHA-256
anchored by the checked-in lock, and the installed package records are then
compared field-for-field with that lock. It packages the FEniCS/PETSc/MPI closure in a
private fenics-runtime/ namespace, bundles the exact Zig toolchain, and
excludes producer CPython and NumPy. It performs complete Mach-O closure and
ad-hoc-signature audits, records package ownership and license inventory, then
builds the overlay twice and requires byte-identical ZIP/summary/checksum
outputs. The resulting artifact hashes are retained as qualification evidence
rather than treated as public release assets.
The same public workflow now performs a self-contained consumer qualification
after deleting the producer conda prefix/package cache and extracted Zig source
tree. It constructs a clean, hash-pinned official CPython 3.12.10
Python.framework plus NumPy 2.5.2 runtime under a path containing spaces,
rejects overlay overwrites, poisons ambient compiler/native discovery, activates
the overlay twice, verifies provider ownership, runs the full import stack and a
fresh Zig/FFCx + PETSc Poisson solve, audits generated Mach-O modules, and
rechecks all inherited consumer bytes. No private repository or cross-repository
credential participates in producer acceptance.
Update the environment with conda:
conda activate fenics-windows
conda update -c precise-simulation -c conda-forge fenics-dolfinxPackages are published at:
https://anaconda.org/precise-simulation
This repository contains the recipes and CI infrastructure used to build and publish the native Windows packages.
The current supported Python/JIT backend matrix is defined in config/support-policy.json. The maintenance and release procedure for new Python minors and JIT backend revisions is documented in docs/jit-backend-maintenance.md.
For problems specific to these Windows packages, please open a GitHub issue and include:
- Windows version
- Python version
- output of
conda list - the command used to install the packages
- the complete error message
For general DOLFINx/FEniCSx usage questions that are not Windows-package specific, refer to the upstream FEniCSx documentation.