zaft 0.3.10

High-performance FFT library

zaft — Fast Fourier Transform for Python

zaft is a drop-in replacement for numpy.fft and scipy.fft backed by a hand-tuned Rust engine with SIMD acceleration:

Platform	Instruction set
x86-64	AVX2 + FMA
AArch64	NEON / SVE

Installation

pip install zaft

Pre-built wheels are available for Linux (x86-64, aarch64), macOS (Intel + Apple Silicon), and Windows (x86-64).

Usage

import numpy as np
import zaft

x = np.random.randn(1024).astype(np.complex128)

# 1-D transforms — identical signatures to numpy.fft / scipy.fft
X = zaft.fft(x)
x2 = zaft.ifft(X)

# Real transforms
y = np.random.randn(1024)
Y = zaft.rfft(y)
y2 = zaft.irfft(Y, n=len(y))

# 2-D transforms
img = np.random.randn(512, 512).astype(np.complex128)
IMG = zaft.fft2(img)

# Frequency helpers
freqs = zaft.fftfreq(1024, d=1/44100)

Planner

Create a Plan when the same transform length is reused many times — planning cost is paid once and subsequent calls are essentially free:

plan = zaft.Plan(n=1024, workers=4)

for frame in audio_frames:
    spectrum = plan.execute_forward(frame)
    # … process …

Normalisation

The norm keyword matches scipy.fft / numpy.fft:

norm	Forward scale	Inverse scale
"backward" (default)	1	1/N
"ortho"	1/√N	1/√N
"forward"	1/N	1

API reference

Function	Equivalent in NumPy/SciPy
fft(a, n, norm)	numpy.fft.fft
ifft(...)	numpy.fft.ifft
rfft(...)	numpy.fft.rfft
irfft(...)	numpy.fft.irfft
fft2(...)	numpy.fft.fft2
ifft2(...)	numpy.fft.ifft2
fftn(...)	numpy.fft.fftn
ifftn(...)	numpy.fft.ifftn
fftfreq(n, d)	numpy.fft.fftfreq
rfftfreq(n, d)	numpy.fft.rfftfreq
fftshift(a)	numpy.fft.fftshift
ifftshift(a)	numpy.fft.ifftshift
next_fast_len(n)	scipy.fft.next_fast_len
Plan(n, workers)	scipy.fft.get_backend / planning

Algorithm selection

The library automatically chooses the fastest algorithm for each transform length:

• Powers of 2 / 4 / 8 — split-radix butterfly with AVX2/NEON kernels
• Radix 3, 5, 6, 7, 10, 11, 13 — dedicated radix implementations
• Composite lengths — Good-Thomas or mixed-radix decomposition
• Primes — Rader's algorithm (convolution-based) or Bluestein's chirp-Z

License

BSD 3-Clause — see LICENSE.