On Mon, Oct 13, 2014 at 2:16 PM, Karl Rupp <[email protected]> wrote:
> Hi David, > > > when using the snes/ex12 example > >> https://bitbucket.org/petsc/petsc/src/f5476a2b443e0eedb22d31cb1b2d85 >> b9b918f73f/src/snes/examples/tutorials/ex12.c?at=master >> with >> >> -variable_coefficient field -petscfe_type opencl -mat_petscfe_type opencl >> >> and >> >> -dm_vec_type viennacl -dm_mat_type seqaijviennacl -mat_type >> seqaijviennacl >> >> will PetscFEIntegrateResidual_OpenCL transfer the residual vector >> directly to the >> viennacl datastructures on the GPU for the solve ? >> > > > No, unfortunately this doesn't work yet. The residual implementation is > 'matrix free', so there is no matrix set up and *mat_type without effect. > Setting > -dm_vec_type viennacl > will be supported in the future, but currently the OpenCL buffer is only > set up to compute the residual. > > > is the output residual of PetscFEIntegrateResidual_OpenCL a vector on >> the CPU ? >> > > yes, it is read back to elemVec[], which is passed to > PetscFEIntegrateResidual_OpenCL() > I forgot that assembly is still being done on the CPU, so elemVec[] is read back. However, the vector is OpenCL, so the assembled results will be on the GPU. I wrote code a long time ago to assemble on the GPU, but there was no speedup at all, so I threw it away. I am not sure it will ever make sense to do that. Matt > Best regards, > Karli > > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
