I have finally created an MR that moves the Jacobi accessing of the
diagonal to the GPU, which should improve the GPU performance of your code.
https://urldefense.us/v3/__https://gitlab.com/petsc/petsc/-/merge_requests/8756__;!!G_uCfscf7eWS!YRnufR4azDdpfJngalubC7i6AYidmMy5jzcDkedwqFk6a7Mrf60Lo_nAXsczsIY6DHwZzmPwpii1ARUQLJiC6Q8$
Please give it a try and let us know if it causes any difficulties or,
hopefully, improves your code's performance significantly.
Sorry for the long delay, NVIDIA is hiring too many PETSc developers away
from us.
Barry
> On Jul 31, 2025, at 6:46 AM, LEDAC Pierre <[email protected]> wrote:
>
> Thanks Barry, I agree but didn't dare asking for that.
>
> Pierre LEDAC
> Commissariat à l’énergie atomique et aux énergies alternatives
> Centre de SACLAY
> DES/ISAS/DM2S/SGLS/LCAN
> Bâtiment 451 – point courrier n°41
> F-91191 Gif-sur-Yvette
> +33 1 69 08 04 03
> +33 6 83 42 05 79
> De : Barry Smith <[email protected] <mailto:[email protected]>>
> Envoyé : mercredi 30 juillet 2025 20:34:26
> À : Junchao Zhang
> Cc : LEDAC Pierre; [email protected] <mailto:[email protected]>
> Objet : Re: [petsc-users] [GPU] Jacobi preconditioner
>
>
> We absolutely should have a MatGetDiagonal_SeqAIJCUSPARSE(). It's somewhat
> embarrassing that we don't provide this.
>
> I have found some potential code at
> https://urldefense.us/v3/__https://stackoverflow.com/questions/60311408/how-to-get-the-diagonal-of-a-sparse-matrix-in-cusparse__;!!G_uCfscf7eWS!YRnufR4azDdpfJngalubC7i6AYidmMy5jzcDkedwqFk6a7Mrf60Lo_nAXsczsIY6DHwZzmPwpii1ARUQ1vbEovg$
>
>
> Barry
>
>
>
>
>> On Jul 28, 2025, at 11:43 AM, Junchao Zhang <[email protected]
>> <mailto:[email protected]>> wrote:
>>
>> Yes, MatGetDiagonal_SeqAIJCUSPARSE hasn't been implemented. petsc/cuda and
>> petsc/kokkos backends are separate code.
>> If petsc/kokkos meet your needs, then just use them. For petsc users, we
>> hope it will be just a difference of extra --download-kokkos
>> --download-kokkos-kernels in configuration.
>>
>> --Junchao Zhang
>>
>>
>> On Mon, Jul 28, 2025 at 2:51 AM LEDAC Pierre <[email protected]
>> <mailto:[email protected]>> wrote:
>>> Hello all,
>>>
>>> We are solving with PETSc a linear system updated every time step (constant
>>> stencil but coefficients changing).
>>>
>>> The matrix is preallocated once with MatSetPreallocationCOO() then filled
>>> each time step with MatSetValuesCOO() and we use device pointers for coo_i,
>>> coo_j, and coefficients values.
>>>
>>> It is working fine with a GMRES Ksp solver and PC Jacobi but we are
>>> surprised to see that every time step, during PCSetUp,
>>> MatGetDiagonal_SeqAIJ is called whereas the matrix is on the device.
>>> Looking at the API, it seems there is no MatGetDiagonal_SeqAIJCUSPARSE()
>>> but a MatGetDiagonal_SeqAIJKOKKOS().
>>>
>>> Does it mean we should use Kokkos backend in PETSc to have Jacobi
>>> preconditioner built directly on device ? Or I am doing something wrong ?
>>> NB: Gmres is running well on device.
>>>
>>> I could use -ksp_reuse_preconditioner to avoid Jacobi being recreated each
>>> solve on host but it increases significantly the number of iterations.
>>>
>>> Thanks,
>>>
>>> <pastedImage.png>
>>>
>>>
>>> Pierre LEDAC
>>> Commissariat à l’énergie atomique et aux énergies alternatives
>>> Centre de SACLAY
>>> DES/ISAS/DM2S/SGLS/LCAN
>>> Bâtiment 451 – point courrier n°41
>>> F-91191 Gif-sur-Yvette
>>> +33 1 69 08 04 03
>>> +33 6 83 42 05 79
