James A Charles <[email protected]> writes: > For this multiplication it is a dense matrix of size Nxp (Where N >> > p) with a matrix of size pxp. If dense-sparse works my workaround > could be to store the pxp matrix as sparse since it never gets very > large anyways (and its triangular). And I don't need direct access to > its data array.
Yeah, or you could store the pxp matrix redundantly on each process and use MatDenseGetArray, then either (a) call BLAS gemm or (b) MatCreateSeqDense(... with array) and MatMult for the serial matrix multiply. > For MatMatMult in parallel sparse-dense, dense-sparse, and > sparse-sparse works? Yes > And dense-dense does not because you have switched to Elemental and > that has a different API that must be used? Well, our elemental support is intended to be simplified and currently only supports the [MC,MR] distribution, which is a 2D distribution for squarish matrices. It would be nice to add support for more Elemental distributions, especially [VC,*], but I don't have short-term plans to do this. Contributions welcome. :-)
pgpfFOXVq3Aaf.pgp
Description: PGP signature
