anyfft

anyfft is a thin compatibility layer that wraps various python FFT implementations in the scipy.fft API (if they do not already provide it). my motiviation is to have a package that I can install anywhere that will take advantage of the available hardware (CUDA, OpenCL, CPU fallback) without refactoring.

import anyfft
import numpy as np

array = np.random.rand(128, 128)
anyfft.fft(array, plugin='reikna')
# alternative, use namespace directly:
anyfft.reikna.fft(array)

current backends:

• numpy (np.fft)
• scipy (scipy.fft)
• fftpack (scipy.fftpack)
• pyfftw
• reikna (OpenCL FFT)
• cupy (CUDA FFT)

available functions:

fft, fft2, fftn, ifft, ifft2, ifftn, fftshift, ifftshift, irfft (reikna WIP), irfft2 (reikna WIP), irfftn (reikna WIP), rfft, rfft2, rfftn

... functions are tested to match the output of the corresponding scipy.fft function

scipy backend

Scipy supports pluggable fft backends

This is going to be the preferred way to use this library:

In [1]: from anyfft import ReiknaBackend

In [2]: from scipy import fft

In [3]: d = np.random.random((2048,2048))

In [4]: %%timeit
   ...: with fft.set_backend('scipy'):
   ...:     x = fft.fftn(d)
   ...:
49.2 ms ± 188 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [5]: %%timeit
   ...: with fft.set_backend(ReiknaBackend):
   ...:     x = fft.fftn(d)
   ...:
14.9 ms ± 2.66 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)