Small C++ library wrapping various FFT implementations for some common audio processing use cases
d41a117b8cbe — Chris Cannam default tip 2 years ago
Update CI for SLEEF
bc573d93d1ce — Chris Cannam 2 years ago
Add SLEEF support
87de39e4d7ab — Chris Cannam sleef 2 years ago
Add SLEEF support. Seems pretty good
993dac232e8a — Chris Cannam 2 years ago
Remove unsupported libraries from tests

heads

tip
browse log
v1.0.0
browse .tar.gz

clone

read-only
https://hg.sr.ht/~breakfastquay/bqfft
read/write
ssh://hg@hg.sr.ht/~breakfastquay/bqfft

#bqfft

A small library wrapping various FFT implementations for some common audio processing use cases. Contains a built-in implementation and wrappers for FFTW3, SLEEF, KissFFT, Intel IPP, and Apple vDSP. Suitable for Windows, Mac, Linux, and mobile platforms.

Note this is not a general FFT interface, as it handles only real signals on the time-domain side.

Transforms of any length are supported, but if you request a length that bqfft does not know how to calculate using any of the libraries that have been compiled in, a simple slow DFT will be used instead. A warning will be printed to stderr if this happens.

Of the available libraries, vDSP, IPP, SLEEF, and the built-in implementation support power-of-two FFT lengths only, KissFFT supports any multiple of two, and FFTW supports any length. You can compile in more than one library, so for example if you compile in Accelerate and KissFFT, the former will be used for powers of two and the latter for other even lengths.

Here are some other pros and cons of the supported libraries:

  • Intel IPP - The fastest on actual Intel hardware. Of uncertain benefit with other manufacturers. Not available beyond x86/amd64, not open source.

  • Apple vDSP - Generally the fastest on all Apple hardware, and provided with the OS. There is seldom any good reason not to use this on Apple platforms.

  • SLEEF - Typically very fast, unencumbered, portable, open source vector library; complex and (at the time of writing) rather new.

  • FFTW3 - Fast, open source, and portable, but its bulk and GPL licence may be an issue.

  • KissFFT - As used here it is single-precision throughout, so it may be a good choice for platforms on which double-precision arithmetic is slow. Disadvantage is that it does not preserve the full float range of precision (i.e. forward-inverse transform pairs on float data do not produce identical results to the input). Not especially fast on desktop or modern mobile hardware.

  • Built-in implementation - Double precision, so more precise than KissFFT, and faster on typical 64-bit desktop and modern mobile hardware. Slower than IPP, vDSP, SLEEF, and FFTW3.

Requires the bqvec library.

This code originated as part of the Rubber Band Library written by the same authors (see https://hg.sr.ht/~breakfastquay/rubberband/). It has been pulled out into a separate library and relicensed under a more permissive licence.

C++ standard required: C++98 (does not use C++11 or newer features)

  • To compile on Linux: Edit Makefile to select implementation, then make. Do read the notes in the Makefile

  • To compile on macOS: make -f build/Makefile.osx

  • To build and run tests: as above, but add the "test" target - requires Boost test headers installed

  • Depends on: bqvec

  • See also: bqresample bqaudioio bqthingfactory bqaudiostream

Build status

Copyright 2007-2022 Particular Programs Ltd. See the file COPYING for (BSD/MIT-style) licence terms.