Hi Jim,
in addition to what Matt already said, keep in mind is that you usually
won't see a two-fold performance gain in iterative solvers anyway, as
the various integers used for storing the nonzeros in the sparse matrix
don't change their size. I once played with an implementation of an
non-preconditioned mixed-precision CG solver, and I only obtained about
a 40 percent overall performance gain for well-conditioned systems. For
less well-conditioned systems you may not get any better overall
performance at all (or worse, fail to converge).
Best regards,
Karli
On 08/12/2013 12:32 PM, Matthew Knepley wrote:
On Mon, Aug 12, 2013 at 12:24 PM, Jim Fonseca <[email protected]
<mailto:[email protected]>> wrote:
Hi,
We are curious about the mixed-precision capabilities in NEMO5. I
see that there is a newish configure option to allow single
precision for linear solve. Other than that, I found this old post:
https://lists.mcs.anl.gov/mailman/htdig/petsc-users/2012-August/014842.html
Is there any other information about to see if we can take advantage
of this capability?
Mixed-precision is hard, and especially hard in PETSc because the C type
system is limited.
However, it also needs to be embedded in an algorithm that can take
advantage of it. I would
always start out with a clear motivation:
- What would mixed precision accomplish in your code?
- What is the most possible benefit you would see?
and decide if that is worth a large time investment.
Thanks,
Jim
--
Jim Fonseca, PhD
Research Scientist
Network for Computational Nanotechnology
Purdue University
765-496-6495 <tel:765-496-6495>
www.jimfonseca.com <http://www.jimfonseca.com>
--
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which
their experiments lead.
-- Norbert Wiener
