[SciPy-dev] fft benchmarks for new routines
pearu at cens.ioc.ee
Mon May 12 04:11:37 CDT 2003
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)
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.
More information about the Scipy-dev