Re: FFTW3.0.1 wisdom

[David Bateman]

Daprès Dmitri A. Sergatskov <address@hidden> (le 19/02/2004):
> David Bateman wrote:
> >Be careful here, because it might not be just processor dependent, but
> >also dependent on the amount of memory available. FFTW might use some
> >technique where it allocates large blocks of memeory to speed up FFT's
> >but if you use say wisdom generated on a machine with 512MB on one with
> >128MB uou might, even with the same processor you might be surprised...
> >These are just my speculations, I don't know what FFTW really does
> >internally .....
> 
> In Steve Johnson paper you can see a plot where they tested
> PIV with wisdom file generated on PPC970 and PPC970 with a wisdom
> file generated on PIV -- the relative worst drop in performance
> was ~30% and in few cases it did not change at all.
> I did some simple tests of fftw with and without wisdom file
> and for _large_ problem difference is quite small. And fftw3
> without any wisdom outperforms fftw2.
> For some small cases I get up to ~50% improvement in fftw3.
> Without it fftw3 is ~30% slower than fftw2 and with it
> fftw3 30% faster than fftw2. The tests were not exhaustive, but
> it so far it looks that fftw3 is robust,  even "wrong" wisdom will
> not degrade performance too much.

Ok, you have the evidence I was only speculating on worst cases.

Cheers
David

-- 
David Bateman                                address@hidden
Motorola CRM                                 +33 1 69 35 48 04 (Ph) 
Parc Les Algorithmes, Commune de St Aubin    +33 1 69 35 77 01 (Fax) 
91193 Gif-Sur-Yvette FRANCE

The information contained in this communication has been classified as: 

[x] General Business Information 
[ ] Motorola Internal Use Only 
[ ] Motorola Confidential Proprietary



-------------------------------------------------------------
Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:  http://www.octave.org
How to fund new projects:  http://www.octave.org/funding.html
Subscription information:  http://www.octave.org/archive.html
-------------------------------------------------------------