Re: MinGW libblas [WAS: Re: Octave 3.6.0 on Windows XP plot fails.]

[Martin Helm]

Am 20.02.2012 11:04, schrieb Tatsuro MATSUOKA:
> Hello
>
> Please tell me which is correct to measure the performance.
>
> 1. n=2000; A=randn(n); B=randn(n);tic; C=A*B; t=toc, MFLOPS=2*n^3/t*1e-6
>
> 2. n=2000; A=randn(n); B=randn(n);ta=cputime; C=A*B; t=cputime-ta, 
> MFLOPS=2*n^3/t*1e-6
>
> Case 1 gives two - four times faster than that of the case 2.
>
> Example
> octave:2> n=2000; A=randn(n); B=randn(n);tic; C=A*B; t=toc, 
> MFLOPS=2*n^3/t*1e-6
> t =  0.70904
> MFLOPS = 2.2566e+004
> octave:3> n=2000; A=randn(n); B=randn(n);ta=cputime; C=A*B; t=cputime-ta, 
> MFLOPS=2*n^3/t*1e-6
> t =  1.4040
> MFLOPS = 1.1396e+004
>
>
> Which test is correct to check the performance.
>
> Regards
>
> Tatsuro
>
The cputime sums together the time spent on all cpu's so if it runs lets
say a total of 4sec till its complete parallel on 4 cpu's then cputime
will give 16sec.
You better use tic/toc.