MatGetMultiProcBlock#

Create multiple ‘parallel submatrices’ from a given Mat. Each submatrix can span multiple procs.

Synopsis#

#include "petscmat.h" 
PetscErrorCode MatGetMultiProcBlock(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)

Collective

Input Parameters#

mat - the matrix
subcomm - the sub communicator obtained as if by MPI_Comm_split(PetscObjectComm((PetscObject)mat))
scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

Output Parameter#

subMat - parallel sub-matrices each spanning a given subcomm

Notes#

The submatrix partition across processors is dictated by subComm a communicator obtained by MPI_comm_split() or via PetscSubcommCreate(). The subComm is not restricted to be grouped with consecutive original ranks.

Due the MPI_Comm_split() usage, the parallel layout of the submatrices map directly to the layout of the original matrix [wrt the local row,col partitioning]. So the original ‘DiagonalMat’ naturally maps into the ‘DiagonalMat’ of the subMat, hence it is used directly from the subMat. However the offDiagMat looses some columns - and this is reconstructed with MatSetValues()

This is used by PCBJACOBI when a single block spans multiple MPI ranks