Re: Working patch for FFTW 3.0.x and Nd FFT's

[Daniel J Sebald]

David Bateman wrote:

> According to Daniel J Sebald <address@hidden> (on 02/18/04):
>  
>
> > That was sort of my original concern, about the user getting stuck in 
> > some situation he or she doesn't know what is going on.  (I'm wondering 
> > what could possibly be taking 12 hours, but that's a different matter.)
> >    
>
> FFTW tests all possible ways it knows of to do a particular FFT and this
> stores the information on the best one for the current architecture.
> With the command "fftw-wisdom -c -v -o /etc/fftw/wisdom" it tests about
> 8 different ways FFTW might be called to do a particular FFT for the
> following FFT sizes
>
>
>     "1", "2", "4", "8", "16", "32", "64", "128", "256", "512", "1024",
>     "2048", "4096", "8192", "16384", "32768", "65536", "131072",
>     "262144", "524288", "1048576",
>
>     "10", "100", "1000", "10000", "100000", "1000000",
>
>     "2x2", "4x4", "8x8", "10x10", "16x16", "32x32", "64x64", "100x100",
>     "128x128", "256x256", "512x512", "1000x1000", "1024x1024",
>
>     "2x2x2", "4x4x4", "8x8x8", "10x10x10", "16x16x16", "32x32x32",
>     "64x64x64", "100x100x100"
>
> So its a lot of work.

That is what I suspected.  The following points aren't critically 
important:  I'm guessing that with the typical nonlinear dependence of 
computations based on FFT length, it is those very long FFT tests that 
are taking the longest time.  (I also question the utility of a million 
point FFT, but let's say someone has a valid scenario for requiring such 
resolution.)  One would think that some logic could be applied to the 
trends of algorithm performance for the shorter FFTs to deduce what 
would be the most efficient algorithm for the very long FFTs.  Anyway, 
one often gets a feeling of the kind of information that is trying to be 
accumulated versus the amount of computer power that should be required, 
and somehow 12 hours seems like a long time to build a "wisdom table". 
This isn't Octave's concern though, I guess.

> > Why not the name fft_wisdom as opposed to fftw_wisdom?  I realize that 
> > FFTW is being used, but if later some other FFT package is used then 
> > fft_wisdom could still apply.  (I may have come in late on this discussion.)
> >    
>
> My original patch had fft_wisdom, John preferred fftw_wisdom. He has CVS
> write access, so he has the last say :-)
>
> D.

:-)

D.