[petsc-users] MatMatMult involving MPIAIJ and MATNEST

[Bikash Kanungo]

So I guess, in terms of performance creating M^{-1} through
MatGetLocalSubMatrix and then explicitly build M^{-1}H is a more efficient
way.
On May 11, 2015 1:58 PM, "Jed Brown" <jed at jedbrown.org> wrote:

> Bikash Kanungo <bikash at umich.edu> writes:
>
> > The physics of the problem requires evaluation of M^{-1} from four
> > sub-matrices. That's why I'm constrained to use it as MATNEST. The other
> > option seems to be MatGetLocalSubMatrix but it seemed easier for me to go
> > with MATNEST (particularly to avoid the local index sets).
>
> If you need a matrix-matrix multiply, then artificial intermediate
> scales will tend to reduce performance rather than improving it.
>
> We can implement MatMatMult_Nest_Nest, but based on what you've said, I
> doubt it's a good fit for your problem.  MatMatMult_Nest_MPIAIJ would
> likely perform worse than MatMatMult_MPIAIJ_MPIAIJ.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150511/bfcf811a/attachment.html>