[petsc-users] MatMult inside a for loop

[Barry Smith]

> On Feb 8, 2015, at 5:41 PM, Ronal Celaya <ronalcelayavzla at gmail.com> wrote:
> 
> Hello
> If I have a MatMult operation inside a for loop (e. g. CG algorithm), and the matrix A is MPIAIJ, vector x is gathered to local process in every loop?

  Yes, internal to MatMult() it calls MatMult_MPIAIJ() which is in src/mat/impls/aij/mpi/mpiaij,c which has the following code:

PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
{
  Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
  PetscErrorCode ierr;
  PetscInt       nt;

  PetscFunctionBegin;
  ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr);
  if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
  ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr);
  ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr);
  PetscFunctionReturn(0);

The needed values of x are communicated in the VecScatterBegin() to VecScatterEnd(). Note only exactly those values needed by each process are communicated in the scatter so not all values are communicated to all processes. Since the matrix is very sparse (normally) only a small percentage of the values need to be communicated.

  Barry

> 
> I'm sorry for my English.
> 
> Regards, 
> 
> -- 
> Ronal Celaya