Re: interesting FFT benchmarking!

[Dmitri A. Sergatskov]

address@hidden wrote:


> octave> A = rand(2048,2048); t0=cputime; B=fft(A); cputime;ans-t0
> ans = 0.85000
> octave> A = rand(2048,2048)+i*rand(2048,2048); t0=cputime;B=fft(A); cputime; 
> ans-t0
> ans = 0.55000

I get absolutely identical timing between real and complex:

octave:69> A = rand(2048,2048); t0=cputime; B=fft(A); cputime;ans-t0
ans = 0.74000
octave:70> A = rand(2048,2048)+i*rand(2048,2048); t0=cputime;B=fft(A); 
cputime;ans-t0
ans = 0.74000

> > > A = rand(2048,2048)+i*rand(2048,2048); t0=cputime; B=fft(A); cputime; ans-t0
>
> ans =
>     0.8700
>
> > > A = rand(2048,2048); t0=cputime; B=fft(A); cputime; ans-t0
>
> ans =
>     0.5600

I can reproduce that in Matlab:

>> A = rand(2048,2048); t0=cputime; B=fft(A); cputime; ans-t0

ans =

    0.5100

>> A = rand(2048,2048)+i*rand(2048,2048); t0=cputime; B=fft(A); cputime; ans-t0

ans =

    0.6500

> I think I'll start adding small complex noises to my signals to get them
> analysed faster! :) No kidding!!!
>
> octave-2.1.50:130> A = rand(2048,2048);
> octave-2.1.50:131> t0=cputime; B=fft(A); cputime; ans-t0
> ans = 0.85000
> octave-2.1.50:136> Anoise = A+i*e-100*ones(2048,2048);
> octave-2.1.50:137> t0=cputime; B=fft(Anoise); cputime; ans-t0
> octave-2.1.50:138> t0=cputime; B=fft(Anoise); cputime; ans-t0
> ans = 0.54000

Cannot reproduce that:

octave:71> A = rand(2048,2048);
octave:72> t0=cputime; B=fft(A); cputime; ans-t0
ans = 0.71000
octave:73> Anoise = A+i*e-100*ones(2048,2048);
octave:74> t0=cputime; B=fft(Anoise); cputime; ans-t0
ans = 0.75000

This is octave-2.1.50 with included fft. Compiled by gcc-3.2.2 with -O2 
-mcpu=athlon-mp
(I did notice that compiling with -O3 produces slightly slower code for fft 
then with -O2,
I am wondering if smaller cache on Athlons to blame...).

Regards,

Dmitri.



-------------------------------------------------------------
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
-------------------------------------------------------------