[SciPy-dev] Experience with Goto's BLAS

[David Cournapeau]

Nils Wagner wrote:
> Hi all,
>
> Is there anyone who can report about the usage of Goto's 
> BLAS with respect to numpy/scipy ?
> What can be said about advantages of ACML, MKL, ATLAS and 
> Goto's BLAS ?
> Who can shed light on the differences between the 
> different libraries ?
The big difference to me: one is open source and freely available 
(ATLAS), not the others. Goto BLAS is supposed to be faster than ATLAS, 
but is available only through registration, and is definitely not open 
source; it is based on another approach than ATLAS for optimization 
(ATLAS tries hard to optimize depending on the CPU cache, GOTO focused 
more on minimizing TLB misses: the principle is for example explained in 
"On Reducing TLB Misses in Matrix Multiplication" (2002) by Kazushige 
Goto et al.)

ACML and MKL are binary only, and available only on some platforms: MKL 
is faster than ATLAS on the platforms I have tested (windows and linux) 
for some test cases. Maybe ACML is, too.

cheers,

David