Re: [ublas] Performance with SSE2

From: James N. Knight (nate_at_[hidden])
Date: 2007-03-12 13:10:53

> 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
>   
It is.  The implementation, however, is not necessarily focused on 
efficiency (more on correctness).   In my experience if your code only 
uses vector operations (level 1 blas) and your vectors are not extremely 
large then there isn't much of a performance gain from linking to a 
tuned library like atlas or acml.  For Level 2/3 blas operations and 
matrices larger than say 30x30 using a library that takes the cache and 
architecture into account makes a large difference.

James