[petsc-users] [Ext] Re: [SLEPc] Computing Smallest Eigenvalue+Eigenvector of Many Small Matrices

[Adam Denchfield]

syevjBatched from cuSolver is quite good once it's configured fine. It's a
direct solve for all eigenpairs, works for batches of small matrices with
sizes up to (I believe) 32x32. The default CUDA example using it works
except if you have "too many" small matrices, in which case you'll overload
the GPU memory and need to further batch the calls. I found it to be fast
enough for my needs.

Regards,
*Adam Denchfield*

*Ph.D Student, Physics*
University of Illinois in Chicago
B.S.  Applied Physics (2018)
Illinois Institute of Technology
Email: adenchfi at hawk.iit.edu


On Wed, Jul 7, 2021 at 2:31 AM Jose E. Roman <jroman at dsic.upv.es> wrote:

> cuSolver has syevjBatched, which seems to fit your purpose. But I have
> never used it.
>
> Lanczos is not competitive for such small matrices.
>
> Jose
>
>
> > El 6 jul 2021, a las 21:56, Jed Brown <jed at jedbrown.org> escribió:
> >
> > Have you tried just calling LAPACK directly? (You could try dsyevx to
> see if there's something to gain by computing less than all the
> eigenvalues.) I'm not aware of a batched interface at this time, but that's
> what you'd want for performance.
> >
> > Jacob Faibussowitsch <jacob.fai at gmail.com> writes:
> >
> >> Hello PETSc/SLEPc users,
> >>
> >> Similar to a recent question I am looking for an algorithm to compute
> the smallest eigenvalue and eigenvector for a bunch of matrices however I
> have a few extra “restrictions”. All matrices have the following properties:
> >>
> >> - All matrices are the same size
> >> - All matrices are small (perhaps no larger than 12x12)
> >> - All matrices are SPD
> >> - I only need the smallest eigenpair
> >>
> >> So far my best bet seems to be Lanczos but I’m wondering if there is
> some wunder method I’ve overlooked.
> >>
> >> Best regards,
> >>
> >> Jacob Faibussowitsch
> >> (Jacob Fai - booss - oh - vitch)
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210707/1d087f4e/attachment.html>