[SciPy-dev] scipy.fft module slow for complex inputs when linked to fftw3
david at ar.media.kyoto-u.ac.jp
Sun Aug 27 21:37:57 CDT 2006
David M. Cooke wrote:
> The patch format is fine, btw. Although, it looks like you used diff;
> it's much easier to just do 'svn diff' (add the file or directory
> names to the end if you're only interested in some of what you've
I of course discovered svn diff a few minutes after having sent the
patch... I am no really used to subversion, will do this way next time.
> I'm working on splitting the _fttpack module into separate extensions
> for each fft library, so it's clearer. Then, you could link in both
> fftw2 and fftw3, and use either (or fftpack). Along the way, I'd like
> to add more of their API functions (such as the wisdom functions), or
> have routines that may have better performance (fftw3, for instance,
> can be more efficient [or not] if it's allowed to destroy its input
> array, or if the input and output array are different).
I was thinking about doing it myself; maybe we can share our effort ?
> Also, it looks like using the guru interface, plans could be cached
> (but for each length, there would be possibly two plans: for aligned
> and unaligned data).
When I used this way, it didn't work, but I didn't investigate more than
a few minutes. I was wondering if this could come from alignment
problems, or some other issues. I don't know much about the C numpy
interface: how is the memory allocated for arrays ? Is the memory always
16 bytes aligned ? Is it an option ? Can we check it ?
More information about the Scipy-dev