OONumerics User : |
From: Toon Knapen (toon.knapen_at_[hidden])
Date: 2007-06-14 08:53:14
We have benchmarked netlib-blas, ATLAS (3.6), ACML and GotoBLAS.
GotoBLAS came out as the clear winner. Currently we're also benchmarking
also (Intel) MKL.
PS: we're mainly performing dgemm and zgemm.
toon
Evgenii Rudnyi wrote:
>> I would like to speed up matrix computations, any suggestion to do that?
>> My conclusion, after some web surfing, is that I should use ATLAS. In
>> "Automated Empirical Optimization of Software and the ATLAS project" by
>> Whaley, Petitet and Dongarra, it seems that computational cost can be
>> reduced ten times or more. Has anybody experienced with this library and
>> obtained this results?
>>
>> Anyway, this paper is 7 years old, and I can't find a newer reference. I
>> know that in http://math-atlas.sourceforge.net I can find some timings,
>> but they are not compared with the timings obtained if you do not use
>> ATLAS... Does anybody know recent comparisons about the computational
>> saving obtained using ATLAS?
>
> I have using ATLAS for my tool MOR for ANSYS (http://ModelReduction.com)
> for quite awhile. I can confirm that with pure BLAS from LAPACK the
> performance is much slower. It is related with the reuse of the
> processor cache that is achieved by using the optimized BLAS. As far as
> I understand, nothing has changed since them. Well, there is new
> research, you can look for example at FLAME
> (http://www.cs.utexas.edu/users/flame/), but ATLAS seems to be the most
> robust tool. Well, if you find an optimized BLAS manually for your
> architecture, it could be faster.
>
> Best wishes,
>
> Evgenii
> _______________________________________________
> oon-list mailing list
> oon-list_at_[hidden]
> http://www.oonumerics.org/mailman/listinfo.cgi/oon-list