Rocket-FFT

Rocket-FFT makes Numba aware of numpy.fft and scipy.fft. It takes its name from the PocketFFT Fast Fourier Transformation library that powers it, and Numba's goal of making your scientific Python code blazingly fast - like a rocket.

Getting Started

The easiest way to get Rocket-FFT is to:

$ pip install rocket-fft

Alternatively, you can build it from source:

$ git clone https://github.com/styfenschaer/rocket-fft.git
$ cd rocket-fft
$ python setup.py install

The latter requires a C++ compiler compatible with your Python installation.

Once installed successfully, the following will work (no import required):

import numba as nb
import numpy as np

@nb.njit
def jit_fft(x):
    return np.fft.fft(x)

a = np.array([1, 6, 1, 8, 0, 3, 3, 9])
jit_fft(a)

Performance Tip

Rocket-FFT makes extensive use of Numba's polymorphic dispatching to achieve both flexible function signatures similar to SciPy and NumPy, and low compilation times. Compilation takes only a few hundred milliseconds in most cases. Calls with default arguments follow a fast path and compile fastest.

NumPy-like and SciPy-like interfaces

NumPy and SciPy show a subtle difference in how they handle the axes argument in some functions¹. Rocket-FFT implements both ways and lets its users choose between them.

You can set the interface by using the scipy_like or numpy_like function from the rocket_fft namespace:

from rocket_fft import numpy_like, scipy_like

numpy_like()

Both functions can be used regardless of whether SciPy is installed². By default, Rocket-FFT uses the SciPy-like interface if SciPy is installed, and the NumPy-like interface otherwise. Note that the interface cannot be changed after the compilation of Rocket-FFT's internals.

¹NumPy allows duplicate axes in fft2, ifft2, fftn and ifftn, whereas SciPy doesn't
²SciPy is an optional runtime dependency

Known limitations

• Rocket-FFT implements NumPy's FFT interface and behavior as introduced in NumPy 2.0. Note, however, that the out argument is not currently supported. Attempting to specify it will result in a compile-time error.

• There is a known issue in the implementations of scipy.fft.dst, scipy.fft.idst, scipy.fft.dstn, and scipy.fft.idstn that may produce incorrect results when norm and/or orthogonalize are set to non-default values. Specifying these arguments with values other than their defaults emits a warning, and it is the user's responsibility to verify the results. Using the default values is safe.

Low-Level Interface

Rocket-FFT also provides a low-level interface to the PocketFFT library. Using the low-level interface can significantly reduce compile time, minimize overhead and give more flexibility to the user. It also provides some functions that are not available through the SciPy-like and NumPy-like interfaces. You can import its functions from the rocket_fft namespace:

from rocket_fft import c2c, dct, ...

The low-level interface includes the following functions:

def c2c(ain: NDArray[c8] | NDArray[c16], aout: NDArray[c8] | NDArray[c16], axes: NDArray[i8], forward: b1, fct: f4 | f8, nthreads: i8) -> None: ...
def r2c(ain: NDArray[f4] | NDArray[f8], aout: NDArray[c8] | NDArray[c16], axes: NDArray[i8], forward: b1, fct: f4 | f8, nthreads: i8) -> None: ...
def c2r(ain: NDArray[c8] | NDArray[c16], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], forward: b1, fct: f4 | f8, nthreads: i8) -> None: ...
def c2c_sym(ain: NDArray[f4] | NDArray[f8], aout: NDArray[c8] | NDArray[c16], axes: NDArray[i8], forward: b1, fct: f4 | f8, nthreads: i8) -> None: ...
def dst(ain: NDArray[f4] | NDArray[f8], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], type: i8, fct: f4 | f8, ortho: b1, nthreads: i8) -> None: ...
def dct(ain: NDArray[f4] | NDArray[f8], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], type: i8, fct: f4 | f8, ortho: b1, nthreads: i8) -> None: ...
def r2r_separable_hartley(ain: NDArray[f4] | NDArray[f8], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], fct: f4 | f8, nthreads: i8) -> None: ...
def r2r_genuine_hartley(ain: NDArray[f4] | NDArray[f8], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], fct: f4 | f8, nthreads: i8) -> None: ...
def r2r_fftpack(ain: NDArray[f4] | NDArray[f8], aout: NDArray[f4] | NDArray[f8], axes: NDArray[i8], real2hermitian: b1, forward: b1, fct: f4 | f8, nthreads: i8) -> None: ...
def good_size(target: i8, real: b1) -> i8: ...

Note that the low-level interface provides a lower level of safety and convenience compared to the SciPy-like and NumPy-like interfaces. There is almost no safety net, and it is up to the user to ensure proper usage. You may want to consult the original PocketFFT C++ implementation before using it.