[SciPy-dev] fft benchmarks for new routines

Pearu Peterson pearu at cens.ioc.ee
Mon May 12 04:11:37 CDT 2003


Hi Chuck,

On 11 May 2003, Charles R Harris wrote:

> I ran a pared down version of Pearu's benchmark script on
> my new fft routines. I also scaled the times by the
> factor 1e8/(repeat*size*log(size)) so as to obtain a
> more informative number. There are three regions of
> interest in the results:
> 
> 1) small transforms are dominated by python calling overhead
> 2) medium transforms run in cache and perform best
> 3) long transforms do out of cache references, performance sucks
> 
> I think my routines are better, not that I'm biased or anything.
> They are faster, require only one work array for all transforms
> below a given size, and the code is readable - not to say
> understandable. 8)

That's nice!
Where one can see your routines?

> To cut down on out-of-cache overhead, I unrolled a loop, which
> made the code about 3x longer and somewhat ugly. The result is
> a sort of bastard split radix 2,4 fft.

What fft backend are you using in scipy? FFTPACK, FFTW, or DJBFFT? 

Note that if your routines are power-two only then it would be still
easy to include (if you wish) them to scipy.fftpack using the same
hooks as for wrapping DJBFFT.

Thanks,
	Pearu




More information about the Scipy-dev mailing list