On Tue, Oct 28, 2014 at 5:28 AM, Nathaniel Smith <njs@pobox.com> wrote:

On 28 Oct 2014 04:07, "Matthew Brett" <matthew.brett@gmail.com> wrote:
>
> Hi,
>
> On Mon, Oct 27, 2014 at 8:07 PM, Sturla Molden <sturla.molden@gmail.com> wrote:
> > Sturla Molden <sturla.molden@gmail.com> wrote:
> >
> >> If we really need a
> >> kick-ass fast FFT we need to go to libraries like FFTW, Intel MKL or
> >> Apple's Accelerate Framework,
> >
> > I should perhaps also mention FFTS here, which claim to be faster than FFTW
> > and has a BSD licence:
> >
> > http://anthonix.com/ffts/index.html
>
> Nice.  And a funny New Zealand name too.
>
> Is this an option for us?  Aren't we a little behind the performance
> curve on FFT after we lost FFTW?

It's definitely attractive. Some potential issues that might need dealing with, based on a quick skim:

- seems to have a hard requirement for a processor supporting SSE, AVX, or NEON. No fallback for old CPUs or other architectures. (I'm not even sure whether it has x86-32 support.)

-  no runtime CPU detection, e.g. SSE vs AVX appears to be a compile time decision

- not sure if it can handle non-power-of-two problems at all, or at all efficiently. (FFTPACK isn't great here either but major regressions would be bad.)

- not sure if it supports all the modes we care about (e.g. rfft)

This stuff is all probably solveable though, so if someone has a hankering to make numpy (or scipy) fft dramatically faster then you should get in touch with the author and see what they think.

-n

I recently became aware of another C-library for doing FFTs (and other things):

https://github.com/arrayfire/arrayfire

They claim to have comparable FFT performance to MKL when run on a CPU (they also support running on the GPU but that is probably outside the scope of numpy or scipy).  It used to be proprietary but now it is under a BSD-3-Clause license.  It seems it supports non-power-of-2 FFT operations as well (although those are slower).  I don't know much beyond that, but it is probably worth looking in