[petsc-users] Hypre runtime switch CPU/GPU

[Milan Pelletier]

Ok thanks for the answers, this would definitely be a super useful feature to have.

Milan

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
Le jeudi 7 octobre 2021 à 1:22 PM, Stefano Zampini <stefano.zampini at gmail.com> a écrit :

> We have discussed full runtime switch in HYPRE with Ruipeng few weeks ago, I'm not sure what's the status. cc'ing him
>
> Il giorno gio 7 ott 2021 alle ore 14:10 Mark Adams <mfadams at lbl.gov> ha scritto:
>
>> I'm not sure, but I suspect that Hypre does not support runtime switching and our model is that you can switch at runtime. This leads to an inconsistency.
>>
>> If we remove -mat_type hypre then your issue would go away but 1) we would have to add it back if hypre supports runtime switching in the future, and break everyone's input decks, and 2) it would be inconsistent with the PETSc model.
>>
>> I could see throwing an error if you do not use -mat_type hypre and are configured for GPUs.
>>
>> Mark
>>
>> On Wed, Oct 6, 2021 at 3:31 PM Milan Pelletier via petsc-users <petsc-users at mcs.anl.gov> wrote:
>>
>>> Dear PETSc users,
>>>
>>> Is there a way to switch a runtime setting for PETSc+Hypre to run on CPU, even when it has been compiled to allow for GPU support?
>>> I looks like setting the matrix and vector types to respectively "seqaij" and "seq" results in GPU computation when Hypre is used as a preconditioner. I thought GPU would be used only when mat_type is set to "hypre", following the examples provided with the last release.
>>>
>>> Thanks for the help,
>>> Best regards,
>>>
>>> Milan
>
> --
>
> Stefano
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20211007/1b783a42/attachment.html>