Re: [Help-gsl] Slow BLAS

[Gong Yi Liao]

On Sun, 2006-10-22 at 12:17 +0200, Jochen Küpper wrote:
> On 22.10.2006, at 10:53, Gong Yi Liao wrote:
> 
> >     I find that the Linear algebra subroutines bundled (more specific,
> > the BLAS) with GSL1.8 are quite slow, for example, For computing  
> > SVD of

> 
> Well, then use the fast BLAS with GSL, it's a simple matter of  
> specifying it in the link-step.
> 
> Greetings,
> Jochen

I have to say sorry first since I did not specify my problems in 
detail, I have tried to linked with variant BLASs, including GotoBLAS
with cblas interface compiled from NetLib's source code, ATLAS packaged
by debian maintainer (already provided cblas interface with debian
patch) and gslcblas provided by GSL developing team.
Something interesting is that the gslcblas is the fastest one.
In my experiment, 
ATLAS takes 56.676 seconds (all the cases including the time for
generating gaussian random numbers) to compute SVD of a 1000 by 1000
matrix (all the cases are computed with the same sizes of the matrix);

GotoBLAS (with multi-threading support )takes 56.587 seconds to compute;
 
gslcblas takes 56.127 seconds to compute.

Note that the sys time measurements taken by ATLAS and GotoBLAS are both
0.020 second and gslcblas takes 0.012 second

I guess that optimized BLAS does not help GSL to perform better with
large scale matrices.

It may be some problems about allocating the elements of the large scale
matrix in the memory while doing such a large scale linear algebra
computation; another solution may be building a comprehensive LAPACK
interface for GSL, but it seems not have good idea for
implementation(according to the archives of the mailing list).

Any advice is appreciated.

Best regards, Gong Yi Liao   
> _______________________________________________
> Help-gsl mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/help-gsl