[petsc-dev] SEGV using PETSc-GPU ViennaCL

[Karl Rupp]

Hi Lucas,

 > I am trying to use PETSc-dev for a while in a project at my university.
> I need to use the GPU for matrix-vector operations and the code that I'm
> working was already implemented using PETSc, obviously I choose the
> PETSc ViennaCL-GPU solution. The original code works perfectly, but when
> I use the GPU parameters, a SEGV occurs in some test cases.
> Valgrind says that is a problem with PetscMallocAlign at memalign but,
> some times, it says that the problem is MatGetDiagonal.
> I'm sending the configure.log because I don't know what are the cause of
> the problem (hope the problem is the configuration), and I would like to
> know if some similar problem were reported by PETSc-ViennaCL users and
> PETSc-CUDA users too.

There are some known issues with GPU support with multiple processes, 
which Dominic just created pull requests for. As Dominic said, it would 
help us a lot if you can provide instructions to reproduce the problem.


> PS: My computer has a top AMD GPU and I'm not sure that I can find a
> NVIDIA GPU computer to test my code, so OpenCL solutions are better to
> me then CUDA.

FYI: AMD GPUs usually have slightly higher memory bandwidth than NVIDIA 
GPUs, which is the limiting factor for almost all operations inside 
PETSc. There is no significant difference in performance between OpenCL 
and CUDA on NVIDIA GPUs anymore.

Best regards,
Karli