| a65300a6 | 31-Jul-2013 |
Paul Mullowney <paulm@txcorp.com> |
Fix to aijcusparse for ell/hyb matrices in Multi-GPU MatMult
The parameters of the matrix (i.e. the number of rows) were not being set correctly in a parallel (Multi-GPU) MatMult for Ell/Hyb matrix
Fix to aijcusparse for ell/hyb matrices in Multi-GPU MatMult
The parameters of the matrix (i.e. the number of rows) were not being set correctly in a parallel (Multi-GPU) MatMult for Ell/Hyb matrix storage formats. In particular, the offdiagonal component matrix used in MatMultAdd had the number of rows set incorrectly. The existing test, src/ksp/ksp/examples/tests/ex43 (runex43_5) only set the diagonal component to be in non-csr format so this wasn't caught earlier. D'oh. I have checked all combinations of storage formats on ksp/ksp/examples/tutorials/ex2.c. Results are entirely consistent in serial and parallel and are tested on CUDA 5.0(2070, Fermi) and CUDA 4.2 (1060, Tesla) architectures.
show more ...
|