[petsc-users] Speed of MatGetRowSum

[Smith, Barry F.]

  Here is the current code: ;-)

PetscErrorCode MatGetRowSum(Mat mat, Vec v)
{
  Vec            ones;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(mat,MAT_CLASSID,1);
  PetscValidType(mat,1);
  PetscValidHeaderSpecific(v,VEC_CLASSID,2);
  if (!mat->assembled) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
  MatCheckPreallocated(mat,1);
  ierr = MatCreateVecs(mat,&ones,NULL);CHKERRQ(ierr);
  ierr = VecSet(ones,1.);CHKERRQ(ierr);
  ierr = MatMult(mat,ones,v);CHKERRQ(ierr);
  ierr = VecDestroy(&ones);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

An optimized for AIJ one would avoid creating the matrix and avoid the multiples by 1 in the summation and avoid the communication.

I wouldn't worry about its performance; it won't be the bottle neck generally. Writing a custom routine for AIJ (seq and MPI) would be easy, just structure it like the MatMult_xxx() routine.

   Barry


> On Apr 27, 2018, at 11:58 AM, Derek Gaston <friedmud at gmail.com> wrote:
> 
> I was wondering about the comment on MatGetRowSum that says:
> 
> "This code is slow since it is not currently specialized for different formats"
> 
> How slow are we talking about here?  Is it actually going to loop over i,j from 0 to N on a sparse matrix?
> 
> I need to get a vector that represents the "lumped" mass matrix: but I would like to get it fairly quickly.  I would just compute the lumped matrix directly and then use MatGetDiag - but I actually want both (unlumped/consistent and lumped) and MatGetRowSum seemed like the best direction.  Willing to take advice though!
> 
> Derek