[Trilinos-Users] EpetraExt mult_A_B

[Williams, Alan B]

Burkhard,

Thank you for the optimization, it does provide a substantial
performance gain in 3 of the kernels (mult_A_B, mult_Atrans_B and
mult_Atrans_Btrans). I will incorporate the change into the Trilinos
code-base, and it will be included in the upcoming 7.0 release.

The remaining kernel, mult_A_Btrans can't be modified in the same way as
the others, but I found a couple of other optimizations for that routine
that improved its performance also. It is still the slowest one, so more
work is needed there.

As I mentioned, our plan is to replace at least some of these kernels
with a much faster outer-product formulation. I just haven't gotten to
it yet.

Alan
 

> -----Original Message-----
> From: trilinos-users-bounces at software.sandia.gov 
> [mailto:trilinos-users-bounces at software.sandia.gov] On Behalf 
> Of Burkhard Doliwa
> Sent: Thursday, July 27, 2006 7:44 AM
> To: trilinos-users at software.sandia.gov
> Subject: [Trilinos-Users] EpetraExt mult_A_B
> 
> Dear EpetraExt authors,
> 
> looking at the EpetraExt matrix-matrix multiplication routine 
> (mult_A_B), one sees that its performance scales as n*n with 
> the size n 
> of square   n x n   matrices (assuming constant # nnz per 
> row). With a 
> VERY small modification (s.App.), one can reach the scaling ~n.
> 
> Perhaps similar modifications can be done in the 
> transpose-mult routines.
> 
> Cheers,
> Burkhard
> 
>