[petsc-users] MatZeroRows and MATELEMENTAL

[Francesco Ballarin]

Thanks Jack, Barry and Matt.

I patched my PETSc 3.5.2 as follows, but I am not really sure that this is
the correct way (for example, I see in MatSetValues_Elemental two calls to
functions P2RO and RO2E for the computation of the local row index but I do
not understand what is their purpose, nor if I should apply them also to
rowInd[s]).

Thanks

======

diff --git a/petsc/src/mat/impls/elemental/matelem.cxx
b/petsc/src/mat/impls/elemental/matelem.cxx
--- a/petsc/src/mat/impls/elemental/matelem.cxx
+++ b/petsc/src/mat/impls/elemental/matelem.cxx
@@ -772,6 +772,36 @@ static PetscErrorCode MatZeroEntries_Elemental(Mat A)
 }

 #undef __FUNCT__
+#define __FUNCT__ "MatZeroRows_Elemental"
+PetscErrorCode MatZeroRows_Elemental(Mat AA,PetscInt N,const PetscInt
rows[],PetscScalar diag,Vec /*x*/,Vec /*b*/)
+{
+  // TODO: consider also the optional arguments x and b.
+  Mat_Elemental * a( (Mat_Elemental*)AA->data );
+  elem::DistMatrix<PetscElemScalar> * A( a->emat );
+
+  PetscFunctionBegin;
+  std::vector<PetscInt> rowInd(rows,rows+N);
+  const elem::Int localWidth( A->LocalWidth() );
+  for( elem::Int s(0); s < N; ++s )
+  {
+    const elem::Int row = rowInd[s];
+    if( A->IsLocalRow(row) )
+    {
+      const elem::Int iLoc(A->LocalRow(row));
+      for( elem::Int jLoc(0); jLoc < localWidth; ++jLoc )
+      {
+        const elem::Int j(A->GlobalCol(jLoc));
+        if( row == j )
+          A->SetLocal( iLoc, jLoc, diag );
+        else
+          A->SetLocal( iLoc, jLoc, 0. );
+       }
+    }
+  }
+  PetscFunctionReturn(0);
+}
+
+#undef __FUNCT__
 #define __FUNCT__ "MatGetOwnershipIS_Elemental"
 static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
 {
@@ -957,7 +987,7 @@ static struct _MatOps MatOps_Values = {
        MatAssemblyEnd_Elemental,
        0, //MatSetOption_Elemental,
        MatZeroEntries_Elemental,
-/*24*/ 0,
+/*24*/ MatZeroRows_Elemental,
        MatLUFactorSymbolic_Elemental,
        MatLUFactorNumeric_Elemental,
        MatCholeskyFactorSymbolic_Elemental,

======



On Mon, Oct 27, 2014 at 11:00 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:

>
> > On Oct 26, 2014, at 4:32 AM, Francesco Ballarin <
> francesco.ballarin at polimi.it> wrote:
> >
> > Dear PETSc users,
> > I have a code for PETSc 3.3 in which the PLAPACK interface, now
> discontinued, was used. I am trying to update it to PETSc 3.5.2, using the
> Elemental interface.
> >
> > I see that MatZeroRows is not implemented for MATELEMENTAL. What
> function should I use instead? Does it also support the version with
> optional arguments x and b?
>
>    Unfortunately due to Elemental’s complicated parallel storage of matrix
> entries we don’t have any routines to do this. (We don’t even have routines
> to generically set values INSERT_VALUES) on other processes. cc:ing
> Elemental’s author to see what he suggests.
>
>   Barry
>
> >
> > Thanks,
> > best regards,
> > Francesco
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20141028/71ffc814/attachment.html>