|
|
Ublas : |
From: Preben Hagh Strunge Holm (preben_at_[hidden])
Date: 2007-03-12 12:57:54
> AFAIR you only gain performance benefits if the compiler is able to
> automatically vectorize computations. Unfortunatly are the expression
> template constructions quite complicated. So most compilers do not recognize
> possible optimizations. Second, there were (some years ago) some explicit
> loop unrolling tests. Look for BOOST_UBLAS_USE_DUFF_DEVICE inside the code.
> (Tell me if you manage successful compilation with
> BOOST_UBLAS_USE_DUFF_DEVICE defined.)
Okay... it seems like it's not anything I wan't to do now. But I'll keep
it as a thing to do after my master thesis if the summer get's to cold
and wet :-)
> I think, if you look for maximal performance, you have to use the
> BLAS-bindings inside some critical routines. All other access patters are to
> complicated to benefit from SIMD anyway.
All of my routines are "critical" but I guess not more critical than the
computation itself.
Do you really mean BLAS-bindings? Not the lapack bindings? I actually
thought that ublas was a BLAS-implementation for C++?
Thanks
Best regards,
Preben Holm