[petsc-users] In-place MUMPS call?

[Jed Brown]

On Sun, Mar 6, 2011 at 20:36, Jack Poulson <jack.poulson at gmail.com> wrote:

> The documentation seems to suggest that I can choose a sparse-direct solver
> and factor a MATMPISBAIJ matrix A out-of-place like this:
>
>     Mat F;
>     MatGetFactor( A, MAT_SOLVER_MUMPS, MAT_FACTOR_CHOLESKY, &F );
>     MatFactorInfo cholInfo;
>     cholInfo.fill = 3.0; // arbitrary choice
>     cholInfo.dtcol = 0.5; // arbitrary choice
>     IS perm;
>     ISCreateStride( comm, size, 0, 1, &perm );
>     MatCholeskyFactorSymbolic( F, A, perm, &cholInfo );
>     MatCholeskyFactorNumeric( F, A, &cholInfo );
>
> but the in-place alternative, MatCholeskyFactor
>

I don't know of any sparse direct solver that actually does in-place
factorization. An example of what MatCholeskyFactor does to make itself look
in-place:

PetscErrorCode MatCholeskyFactor_SeqSBAIJ(Mat A,IS perm,const MatFactorInfo
*info)
{
  PetscErrorCode ierr;
  Mat            C;

  PetscFunctionBegin;
  ierr = MatGetFactor(A,"petsc",MAT_FACTOR_CHOLESKY,&C);CHKERRQ(ierr);
  ierr = MatCholeskyFactorSymbolic(C,A,perm,info);CHKERRQ(ierr);
  ierr = MatCholeskyFactorNumeric(C,A,info);CHKERRQ(ierr);
  A->ops->solve            = C->ops->solve;
  A->ops->solvetranspose   = C->ops->solvetranspose;
  ierr = MatHeaderMerge(A,C);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

Is there a way to tell MatCholeskyFactor which approach to use (i.e.,
> MAT_SOLVER_MUMPS),
>

No, maybe this should  be added.


> or, failing that, can I destroy my original matrix 'A' and keep the factor
> 'F' around for solves?
>

Do this, there is no performance penalty.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20110306/f49dddb7/attachment.htm>