Re: zero out columns

["Dave May" <dave.mayhem23@xxxxxxxxx>]

For some finite element codes I've written in the past I have wanted to leave the constraints in the matrix. Since the global FE stiffness is assembled from element (local) contributions, a masking matrix used to zero out entries can also be applied on the element level as well. I would assemble the full matrix, then apply the mask. I found this approach efficient enough (provided you used MAT_FLUSH_ASSEMBLY when switching from ADD_VALUES to INSERT_VALUES) as you are not generating any new non entires within the matrix sparsity pattern. 

Cheers,
    Dave.

On Jan 18, 2008 9:32 PM, Toby D. Young <tyoung@xxxxxxxxxxx> wrote:

Hello users,

The procedure I wish to program with petsc requires setting a series of
matrix elements to zero, and for a given number of rows and columns.
The operation is intended to run for large parallel sparse matrices.

Using MatZeroRowsIS() I can easily achieve zeroing out the rows I want
and now I am wondering how to zero out the columns. I imagine there are
a few other ways of doing this, I've come up with two:


Method 1
MatGetColumnIJ()     // get a specified column
VecSetValues()       // set wanted indices to zero
MatRestoreColumnIJ() // restore vector to the matrix

The Petsc manual warns "Since PETSc matrices are usually stored in
compressed row format, this routine will generally be
slow." (of MatGetColumnIJ).

Method 2
MatTranspose()       // Take the transpose of the matrix
MatZeroRowsIS()      // Zeroi out the row which was the target column

This requires creating a new matrix, i.e. the transpose and then
releasing memory from the old matrix. The advantage is that it is
relatively straightforward to code up.

The rather obvious third way is to simply force the column elements to
be zero with MatSetValues() which, I am guessing, is not likely to be
efficient and I am not really considering this.

I am wondering if anyone has a clue as to which of the methods above is
likely to be more efficient for large parallel sparse matrices. Perhaps
someone may have a suggesttion for an alternative approach.

Thanks in advance.

Best,
       Toby




--

Toby D. Young - Adiunkt (Assistant Professor)
Department of Computational Science
Institute of Fundamental Technological Research
Polish Academy of Sciences
Room 206, Swietokrzyska 21
00-049 Warsaw, POLAND