[petsc-dev] bad code that needs fixing
Paul Mullowney
paulm at txcorp.com
Mon Mar 18 14:54:22 CDT 2013
Should have caught that.
-Paul
> device must be cast to int before being passed to cudaSetDevice().
>
> cudaSetDevice(device);
>
> On Mar 17, 2013, at 3:36 PM, Paul Mullowney<paulm at txcorp.com> wrote:
>
>> Here's a patch following the instructions on the wiki for small patches.
>>
>> -Paul
>>> {
>>> int size;
>>> ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
>>> if (size>1) {
>>> int devCount, device, rank;
>>> ierr = cudaGetDeviceCount(&devCount);CHKERRQ(ierr);
>>> ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
>>> device = rank % devCount;
>>> ierr = cudaSetDevice(device);CHKERRQ(ierr);
>>> } else {
>>> int device;
>>> /* the code below works for serial GPU simulations */
>>> ierr = PetscOptionsGetInt(NULL,"-cuda_set_device",&device,&flg1);CHKERRQ(ierr);
>>> if (flg1) {
>>> ierr = cudaSetDevice(device);CHKERRQ(ierr);
>>> }
>>> }
>>> }
>>>
>>> size int should be PETScMPIInt and device int cannot be an int if it is passed to PetscOptionsGetInt which requires a PetscInt.
>>>
>>>
>> <0001-cuda-device-selection.patch>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: 0002-cuda-device-selection.patch
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130318/852ed9a7/attachment.ksh>
More information about the petsc-dev
mailing list