[petsc-users] SUPERLU_DIST in single precision

[Xiaoye S. Li]

Barry,

We have a macro-ized basefile to produced all 4 precisions for
SuperLU_DIST. In the past, we didn't generate single precision code, since
there was no request.

Since now we started working on mixed precision algorithms.  I generated
single precision code a few weeks ago, and fixed some bugs on the way. A
few days ago, I nailed down a final bug (hopefully) related to single
precision SGEMM or STRSM BLAS problems on GPU (i.e., single precision
cuBLAS error).  All I know is that CPU code is working on Summit, but it
has computation errors after I enable GPU.    I plan to do some more
investigation on this before asking Nvidia folks.

I am busy this week, but should get to it next week.  I suggest you wait
for me to clear this, instead of duplicating the work.

Sherry


On Tue, Jul 21, 2020 at 5:22 PM Barry Smith <bsmith at petsc.dev> wrote:

>
>    Felix,
>
>     What are your needs, do you want this for CPUs or for GPUs?  Do you
> wish to run all your code in single precision or just the SuperLU_Dist
> solver while the rest of your code double?
>
>     If you want to run everything on CPUs using single precision then
> adding the support is very easy, we can provide that for you any time. The
> other cases will require more thought.
>
>     Barry
>
>
> > On Jul 21, 2020, at 8:58 AM, flw at rzg.mpg.de wrote:
> >
> > Dear PETSc support team,
> > some time ago you told me that you are planning on releasing a version
> that supports SUPERLU_DIST in single-precision soon. Can you tell me
> roughly what time frame you had in mind?
> >
> > Best regards,
> > Felix
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20200721/2eb3887e/attachment-0001.html>