MatZeroRows#

Zeros all entries (except possibly the main diagonal) of a set of rows of a matrix.

Synopsis#

#include "petscmat.h" 
PetscErrorCode MatZeroRows(Mat mat, PetscInt numRows, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)

Collective

Input Parameters#

mat - the matrix
numRows - the number of rows to zero
rows - the global row indices
diag - value put in the diagonal of the zeroed rows
x - optional vector of solutions for zeroed rows (other entries in vector are not used), these must be set before this call
b - optional vector of right hand side, that will be adjusted by provided solution entries

Notes#

This routine, along with MatZeroRowsColumns(), is typically used to eliminate known Dirichlet boundary conditions from a linear system.

For each zeroed row, the value of the corresponding b is set to diag times the value of the corresponding x.

If the resulting linear system is to be solved with KSP then one can (but does not have to) call KSPSetInitialGuessNonzero() to allow the Krylov method to take advantage of the known solution on the zeroed rows.

May be followed by using a PC of type PCREDISTRIBUTE to solve the reduced problem (PCDISTRIBUTE completely eliminates the zeroed rows and their corresponding columns) from the matrix.

Unlike MatZeroRowsColumns() for the MATAIJ and MATBAIJ matrix formats this removes the old nonzero structure, from the eliminated rows of the matrix but does not release memory. Because of this removal matrix-vector products with the adjusted matrix will be a bit faster. For the dense and block diagonal formats this does not alter the nonzero structure.

If the option MatSetOption(mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE) the nonzero structure of the matrix is not changed the values are merely zeroed.

The user can set a value in the diagonal entry (or for the MATAIJ format formats can optionally remove the main diagonal entry from the nonzero structure as well, by passing 0.0 as the final argument).

For the parallel case, all processes that share the matrix (i.e., those in the communicator used for matrix creation) MUST call this routine, regardless of whether any rows being zeroed are owned by them.

Each processor can indicate any rows in the entire matrix to be zeroed (i.e. each process does NOT have to list only rows local to itself).

You can call MatSetOption(mat,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE) if each process indicates only rows it owns that are to be zeroed. This saves a global synchronization in the implementation.