[petsc-users] Using CUDA Graphs with MatComputeOperator

[Junchao Zhang]

I haven't used CUDA graph with PETSc.  Do you happen to have a working
example so we can debug?

--Junchao Zhang


On Tue, May 14, 2024 at 6:08 PM Sreeram R Venkat <srvenkat at utexas.edu>
wrote:

> I have a MatShell object that I want to convert to a MATDENSECUDA.
> Normally, I use MatComputeOperator for this. However, I would now also like
> to use a CUDA Graph so that all the calls to MatMult are captured. I can
> wrap a code like for (int
> ZjQcmQRYFpfptBannerStart
> This Message Is From an External Sender
> This message came from outside your organization.
>
> ZjQcmQRYFpfptBannerEnd
> I have a MatShell object that I want to convert to a MATDENSECUDA.
> Normally, I use MatComputeOperator for this. However, I would now also like
> to use a CUDA Graph so that all the calls to MatMult are captured. I can
> wrap a code like
>
> for (int i = 0; i < N; i++)
>     MatMult(A, x,y);
>
> in a CUDA Graph, and it runs fine. If I try to wrap MatComputeOperator in
> a graph, I get runtime errors like
>  cuda error 906 (cudaErrorStreamCaptureImplicit) : operation would make
> the legacy stream depend on a capturing blocking stream
>
> I tried modifying the MatConvert_Shell routine to only put the graph
> around the main for loop, but that still gives the same errors. Is there a
> way to use CUDA Graphs here (either through a modified MatConvert_Shell or
> otherwise)?
>
> Thanks,
> Sreeram
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20240514/0f2002a3/attachment-0001.html>