Re: Octave and threaded ATLAS and FFTW

[Dmitri A. Sergatskov]

John W. Eaton wrote:

> I think it would be worth an attempt just to see what happens, but I
> don't have time (or the need) to do it myself.  If you do it, please
> post your results to the list.

I did compile octave 2.1.50 with pthreaded atlas. Though it mostly work
and some benchmarks show improvements. E.g. from (in)famous Octave2 benchmark
(which was discussed on octave-maint. list)

"normal" ATLAS:
700x700 cross-product matrix (b = a' * a)___________ (sec): 0.2713

"pthreded" ATLAS

700x700 cross-product matrix (b = a' * a)___________ (sec): 0.1727

I am not quite sure yet those are real numbers since I discovered that
"cputime" does not work as expected any more (at least not as I expected):

octave:2> tic; t1=cputime ; a=rand(2000); t2=cputime ; toc
ans = 0.16871
octave:3> t2-t1
ans = 0.16000
(That is probably close enough)

octave:4> tic; t1=cputime ; b=inv(a); t2=cputime ; toc
ans = 7.3011
octave:5> t2-t1
ans = 1.8600
(7.3 seems more real here)

octave:9> tic; t1=cputime ; c=b*a; t2=cputime ; toc
ans = 2.9955
octave:10> t2-t1
ans = 0
(Well...)

octave:12> tic; t1=cputime ; ifft(fft(a)); t2=cputime ; toc
ans = 1.5286
octave:13> t2-t1
ans = 1.5000

(Looks OK again)

So all BLAS/LAPACK benchmarks which use "cputime" are greatly improved :)
(Octave2 uses tic; toc method.)

Frankly, by reading ATLAS docs, I do not understand why would I get any 
improvements
since ATLAS by itself will not spawn separate tasks...

> Thanks,
>
> jwe

Sincerely,
Dmitri.

p.s.

The tests done on AthlonMPx2 2000 MHz. Octave2 is from
http://www.sciviews.org/other/benchmark.htm
ATLAS 2.6.0 which I compiled myself.



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