[SciPy-Dev] [Numpy-discussion] BSD C port of FFTPACK incl. bluestein algorithm

[David Cournapeau]

On Fri, Nov 18, 2011 at 11:42 AM, Robert Kern <robert.kern at gmail.com> wrote:
> On Fri, Nov 18, 2011 at 11:19, Dag Sverre Seljebotn
> <d.s.seljebotn at astro.uio.no> wrote:
>> I've been in touch with Martin Reinecke, author of the libpsht code for
>> spherical harmonic transforms, about licensing issues.
>>
>> libpsht itself will remain under the GPL, but he is likely to release
>> his C port of FFTPACK under BSD in the near future, as it is based on
>> the public domain FFTPACK.
>>
>> I'm grateful for this change for my own purposes (allows releasing my
>> own competing SHT library under the BSD) -- but it could perhaps be
>> useful for NumPy or SciPy as well, depending on how complete the port
>> is? E.g., perhaps make numpy.fft more complete (is the
>> numpy.fft/scipy.fftpack split simply because of the Fortran dependency?).
>
> It used to be the case that scipy.fftpack allowed one to build against
> multiple different, usually faster, FFT libraries like FFTW. I think
> we have backed away from that since the cost of maintaining the build
> configuration for all of those different backends was so high. It's
> worth noting that numpy.fft is already using a C translation of
> FFTPACK. I'm not sure what the differences are between this
> translation and Martin's.

Having a Bluestein transformation alone would be worthwhile, as it
would avoid the N^2 penalty for prime sizes.

I am wondering about precision issues, though (when I tried
implementing bluestein transforms on top of fftpack, it gave very bad
results numerically-wise). A comparison with fftw would be good here.

regards,

David