Re: MTL

Hi -- as published in a few papers (e.g., 
http://csdl.computer.org/dl/mags/cs/1999/06/c6070.pdf), MTL was able to 
match or exceed vendor BLAS for matrix-matrix product (i.e., vendor 
DGEMM).  We were also able to exceed the performance of ATLAS.

However.  As another message mentioned, one has to be careful in 
interpreting this.  MTL is able to match or exceed vendor BLAS when:
o) MTL is compiled with KCC on the front end and vendor C compiler on 
the backed (KCC has some important optimizations for generic 
programming that few other compilers provide)
o) MTL is compiled with BLAIS and FAST components to do various 
optimizations (such as blocking, tiling, unrolling)
o) MTL is tuned for the particular target microprocessor (RISC with 
lots of registers, e.g., powerpc or sparc)

Since MTL was not performance tuned on other architectures and 
compilers, it is unlikely that MTL would be able to have vendor-level 
performance on other platforms.  Tuning is an extremely important part 
of getting high levels of performance (this is the whole premise of 
ATLAS, for example).

Cheers,
Andrew Lumsdaine

On Mar 12, 2004, at 6:32 AM, Jean-Marie Marchal wrote:

>
>
> Dear folks,
>
> being interested in linear algebra performances in general and in C++ 
> in
> particular, I found your site by chance looking at a Stanford 
> document.  So
> I downloaded and tested version mtl-2.1.2-21 and looked at the 
> matrix-matrix
> multiply on dense matrices from your test dir, something known in Blas 
> as
> the DGEMM operation (or rank-k update). This operation sits at the 
> core of
> all real linalg, whether direct or iterative and gets huge benefit from
> caching/blocking operations.
>
> Although with a recent Intel 7.0 C++ compiler and STLPORT I could 
> easily get
> the same performances as straight C code,  this never came close to the
> performances of the Intel MKL library (factor of 10..30!!!). So I am
> questioning the reality of the performance results that you advocate 
> in your
> publications.  Did you really compare it with MKL or an equivalent, 
> such as
> the Goto or Atlas Blas?
>
> Thanks for your help,
>
> Jean-Marie Marchal
>
>
> _______________________________________________
> This list is archived at http://www.osl.iu.edu/MailArchives/mtl-devel/

_______________________________________________
This list is archived at http://www.osl.iu.edu/MailArchives/mtl-devel/