dtfft

Python bindings for dtFFT — a high-performance library for parallel data transpositions and Fast Fourier Transforms via MPI.

This package is distributed as a source distribution (sdist) and is compiled locally during installation. It gives you full control over all build options via CMake flags.

If you need a pre-built wheel (no compiler required), use one of the variant packages: dtfft-openmpi, dtfft-mpich, dtfft-fftw-openmpi, dtfft-cuda12x-openmpi, etc. See the full list on PyPI or the documentation.

Installation

Building from source requires a Fortran compiler (GCC ≥ 10, Intel, or NVHPC), CMake ≥ 3.25, and an MPI implementation with development headers. Build-time Python dependencies (scikit-build, cmake, ninja, pybind11) are installed automatically by pip.

Important: mpi4py in your environment must be compiled against the same MPI implementation that dtFFT will be linked with. Install it from source:

pip install --no-binary mpi4py mpi4py

Transpose-only (no FFT backend):

pip install dtfft

With FFTW3:

CMAKE_ARGS="-DDTFFT_WITH_FFTW=ON" pip install dtfft

With cuFFT (CUDA):

Also install cupy matching your CUDA toolkit version first, e.g. pip install cupy-cuda12x.

CMAKE_ARGS="-DDTFFT_WITH_CUDA=ON -DDTFFT_WITH_CUFFT=ON" pip install dtfft

Any CMake option documented on the build page can be passed via CMAKE_ARGS.

Quick start

import numpy as np
from mpi4py import MPI
import dtfft

# Create a 3-D complex-to-complex plan
plan = dtfft.PlanC2C([256, 256, 256], comm=MPI.COMM_WORLD)

# Allocate MPI-decomposed buffers
x = plan.get_ndarray(plan.alloc_size)
y = plan.get_ndarray(plan.alloc_size)

x[...] = np.random.random(x.shape) + 1j * np.random.random(x.shape)

# Forward transform  (pencil decomposition applied automatically)
plan.execute(x, y, dtfft.Execute.FORWARD)

# Backward transform
plan.execute(y, x, dtfft.Execute.BACKWARD)

Documentation

• API reference
• Building from source
• Examples

License

GPL v3 — see LICENSE.