Re: [Help-gsl] gsl_blas speedup

[Riccardo Lucchese]

Sorry for the late reply.!

Ok, I understand. Thanks for the explanation !

Riccardo

On Jan 3, 2008 12:57 PM, Brian Gough <address@hidden> wrote:
> At Wed, 2 Jan 2008 15:44:40 +0100,
> Riccardo Lucchese wrote:
> > Anyway on calls like ddot this can be 2 times faster than now, and
> > as in my tests simple functions like ddot don't take any advantage
> > in using sse ecc..  (especially due to the calls overhead).
> >
> > Maybe I'm all wrong :) ? Any other ideas?
>
> To do the equivalent of 'NOSIZECHECK' I would call the cblas_ routines
> directly, instead of the gsl_blas routines.  This will eliminate the
> size-checking overhead.
>
> If function call overhead is still a problem (i.e. 'FORCEINLINE') then
> your vectors must be quite small, otherwise the computation of the
> BLAS routine itself will dominate.
>
> BLAS was really designed for large vectors. In the small vector case
> some equivalent inline functions could help.  James Bergstra wrote
> some sample inline BLAS routines a while back but I believe he decided
> they did not give enough performance advantage unless the vector was
> extremely small (like a few elements).  I think there could be some
> cases (like DDOT as you say) where it could be worthwhile though.  It
> would make sense to write such functions as an alternative to any
> cblas_ library.
>
> With small vectors one is usually working with a fixed rather than
> variable length, which is another distinction from the usual BLAS
> routines.
>
> --
> Brian Gough
>
> Network Theory Ltd,
> Publishing Free Software Manuals --- http://www.network-theory.co.uk/
>
>