// B here is the "reduced" matrix. Square matrices w/ Row=Domain=Range only.
double test_with_matvec_reduced_maps(const Epetra_CrsMatrix &A, const Epetra_CrsMatrix &B, const Epetra_Map & Bfullmap){
const Epetra_Map & Amap = A.DomainMap();
Epetra_Vector Xa(Amap), Ya(Amap), Diff(Amap);
const Epetra_Map *Bmap = Bfullmap.NumMyElements() > 0 ? &B.DomainMap() : 0;
Epetra_Vector *Xb = Bmap ? new Epetra_Vector(*Bmap) : 0;
Epetra_Vector *Yb = Bmap ? new Epetra_Vector(*Bmap) : 0;
Epetra_Vector Xb_alias(View,Bfullmap, Bmap ? Xb->Values(): 0);
Epetra_Vector Yb_alias(View,Bfullmap, Bmap ? Yb->Values(): 0);
Epetra_Import Ximport(Bfullmap,Amap);
// Set the input vector
Xa.SetSeed(24601);
Xa.Random();
Xb_alias.Import(Xa,Ximport,Insert);
// Do the multiplies
A.Apply(Xa,Ya);
if(Bmap) B.Apply(*Xb,*Yb);
// Check solution
Epetra_Import Yimport(Amap,Bfullmap);
Diff.Import(Yb_alias,Yimport,Insert);
Diff.Update(-1.0,Ya,1.0);
double norm;
Diff.Norm2(&norm);
delete Xb; delete Yb;
return norm;
}
void EpetraLinearOp::computeAbsRowSum(Epetra_Vector & x) const
{
TEUCHOS_ASSERT(!is_null(rowMatrix_));
RCP<Epetra_CrsMatrix> crsMatrix
= Teuchos::rcp_dynamic_cast<Epetra_CrsMatrix>(rowMatrix_);
TEUCHOS_TEST_FOR_EXCEPTION(is_null(crsMatrix),
Exceptions::OpNotSupported,
"EpetraLinearOp::computeAbsRowSum(...): wrapped matrix must be of type "
"Epetra_CrsMatrix for this method. Other operator types are not supported."
);
//
// Put inverse of the sum of absolute values of the ith row of A in x[i].
// (this is a modified copy of Epetra_CrsMatrix::InvRowSums)
//
if (crsMatrix->Filled()) {
TEUCHOS_TEST_FOR_EXCEPTION(is_null(crsMatrix),
std::invalid_argument,
"EpetraLinearOp::computeAbsRowSum(...): Epetra_CrsMatrix must be filled"
);
}
int i, j;
x.PutScalar(0.0); // Make sure we sum into a vector of zeros.
double * xp = (double*)x.Values();
if (crsMatrix->Graph().RangeMap().SameAs(x.Map()) && crsMatrix->Exporter() != 0) {
Epetra_Vector x_tmp(crsMatrix->RowMap());
x_tmp.PutScalar(0.0);
double * x_tmp_p = (double*)x_tmp.Values();
for (i=0; i < crsMatrix->NumMyRows(); i++) {
int NumEntries = 0;
double * RowValues = 0;
crsMatrix->ExtractMyRowView(i,NumEntries,RowValues);
for (j=0; j < NumEntries; j++) x_tmp_p[i] += std::abs(RowValues[j]);
}
TEUCHOS_TEST_FOR_EXCEPT(0!=x.Export(x_tmp, *crsMatrix->Exporter(), Add)); //Export partial row sums to x.
}
else if (crsMatrix->Graph().RowMap().SameAs(x.Map())) {
for (i=0; i < crsMatrix->NumMyRows(); i++) {
int NumEntries = 0;
double * RowValues = 0;
crsMatrix->ExtractMyRowView(i,NumEntries,RowValues);
double scale = 0.0;
for (j=0; j < NumEntries; j++) scale += std::abs(RowValues[j]);
xp[i] = scale;
}
}
else { // x.Map different than both crsMatrix->Graph().RowMap() and crsMatrix->Graph().RangeMap()
TEUCHOS_TEST_FOR_EXCEPT(true); // The map of x must be the RowMap or RangeMap of A.
}
}
//=============================================================================
int Epetra_FastCrsMatrix::Solve(bool Upper, bool Trans, bool UnitDiagonal, const Epetra_Vector& x, Epetra_Vector& y) const {
//
// This function find y such that Ly = x or Uy = x or the transpose cases.
//
if (!Filled()) EPETRA_CHK_ERR(-1); // Matrix must be filled.
if ((Upper) && (!UpperTriangular())) EPETRA_CHK_ERR(-2);
if ((!Upper) && (!LowerTriangular())) EPETRA_CHK_ERR(-3);
if ((!UnitDiagonal) && (NoDiagonal())) EPETRA_CHK_ERR(-4); // If matrix has no diagonal, we must use UnitDiagonal
if ((!UnitDiagonal) && (NumMyDiagonals()<NumMyRows_)) EPETRA_CHK_ERR(-5); // Need each row to have a diagonal
int i, j, j0;
int * NumEntriesPerRow = NumEntriesPerRow_;
int ** Indices = Indices_;
double ** Values = Values_;
int NumMyCols_ = NumMyCols();
// If upper, point to last row
if ((Upper && !Trans) || (!Upper && Trans)) {
NumEntriesPerRow += NumMyRows_-1;
Indices += NumMyRows_-1;
Values += NumMyRows_-1;
}
double *xp = (double*)x.Values();
double *yp = (double*)y.Values();
if (!Trans) {
if (Upper) {
j0 = 1;
if (NoDiagonal()) j0--; // Include first term if no diagonal
for (i=NumMyRows_-1; i >=0; i--) {
int NumEntries = *NumEntriesPerRow--;
int * RowIndices = *Indices--;
double * RowValues = *Values--;
double sum = 0.0;
for (j=j0; j < NumEntries; j++) sum += RowValues[j] * yp[RowIndices[j]];
if (UnitDiagonal) yp[i] = xp[i] - sum;
else yp[i] = (xp[i] - sum)/RowValues[0];
}
}
else {
j0 = 1;
if (NoDiagonal()) j0--; // Include first term if no diagonal
for (i=0; i < NumMyRows_; i++) {
int NumEntries = *NumEntriesPerRow++ - j0;
int * RowIndices = *Indices++;
double * RowValues = *Values++;
double sum = 0.0;
for (j=0; j < NumEntries; j++) sum += RowValues[j] * yp[RowIndices[j]];
if (UnitDiagonal) yp[i] = xp[i] - sum;
else yp[i] = (xp[i] - sum)/RowValues[NumEntries];
}
}
}
// *********** Transpose case *******************************
else {
if (xp!=yp) for (i=0; i < NumMyCols_; i++) yp[i] = xp[i]; // Initialize y for transpose solve
if (Upper) {
j0 = 1;
if (NoDiagonal()) j0--; // Include first term if no diagonal
for (i=0; i < NumMyRows_; i++) {
int NumEntries = *NumEntriesPerRow++;
int * RowIndices = *Indices++;
double * RowValues = *Values++;
if (!UnitDiagonal) yp[i] = yp[i]/RowValues[0];
for (j=j0; j < NumEntries; j++) yp[RowIndices[j]] -= RowValues[j] * yp[i];
}
}
else {
j0 = 1;
if (NoDiagonal()) j0--; // Include first term if no diagonal
for (i=NumMyRows_-1; i >= 0; i--) {
int NumEntries = *NumEntriesPerRow-- - j0;
int * RowIndices = *Indices--;
double * RowValues = *Values--;
if (!UnitDiagonal) yp[i] = yp[i]/RowValues[NumEntries];
for (j=0; j < NumEntries; j++) yp[RowIndices[j]] -= RowValues[j] * yp[i];
}
}
}
UpdateFlops(2*NumGlobalNonzeros64());
return(0);
}
//=============================================================================
int Epetra_FastCrsMatrix::Multiply(bool TransA, const Epetra_Vector& x, Epetra_Vector& y) const {
//
// This function forms the product y = A * x or y = A' * x
//
int i, j;
double * xp = (double*)x.Values();
double *yp = (double*)y.Values();
int NumMyCols_ = NumMyCols();
if (!TransA) {
// If we have a non-trivial importer, we must import elements that are permuted or are on other processors
if (Importer()!=0) {
if (ImportVector_!=0) {
if (ImportVector_->NumVectors()!=1) { delete ImportVector_; ImportVector_= 0;}
}
if (ImportVector_==0) ImportVector_ = new Epetra_MultiVector(ColMap(),1); // Create import vector if needed
ImportVector_->Import(x, *Importer(), Insert);
xp = (double*)ImportVector_->Values();
}
// If we have a non-trivial exporter, we must export elements that are permuted or belong to other processors
if (Exporter()!=0) {
if (ExportVector_!=0) {
if (ExportVector_->NumVectors()!=1) { delete ExportVector_; ExportVector_= 0;}
}
if (ExportVector_==0) ExportVector_ = new Epetra_MultiVector(RowMap(),1); // Create Export vector if needed
yp = (double*)ExportVector_->Values();
}
// Do actual computation
for (i=0; i < NumMyRows_; i++) {
int NumEntries = *NumEntriesPerRow++;
int * RowIndices = *Indices++;
double * RowValues = *Values++;
double sum = 0.0;
for (j=0; j < NumEntries; j++) sum += RowValues[j] * xp[RowIndices[j]];
yp[i] = sum;
}
if (Exporter()!=0) y.Export(*ExportVector_, *Exporter(), Add); // Fill y with Values from export vector
}
else { // Transpose operation
// If we have a non-trivial exporter, we must import elements that are permuted or are on other processors
if (Exporter()!=0) {
if (ExportVector_!=0) {
if (ExportVector_->NumVectors()!=1) { delete ExportVector_; ExportVector_= 0;}
}
if (ExportVector_==0) ExportVector_ = new Epetra_MultiVector(RowMap(),1); // Create Export vector if needed
ExportVector_->Import(x, *Exporter(), Insert);
xp = (double*)ExportVector_->Values();
}
// If we have a non-trivial importer, we must export elements that are permuted or belong to other processors
if (Importer()!=0) {
if (ImportVector_!=0) {
if (ImportVector_->NumVectors()!=1) { delete ImportVector_; ImportVector_= 0;}
}
if (ImportVector_==0) ImportVector_ = new Epetra_MultiVector(ColMap(),1); // Create import vector if needed
yp = (double*)ImportVector_->Values();
}
// Do actual computation
for (i=0; i < NumMyCols_; i++) yp[i] = 0.0; // Initialize y for transpose multiply
for (i=0; i < NumMyRows_; i++) {
int NumEntries = *NumEntriesPerRow++;
int * RowIndices = *Indices++;
double * RowValues = *Values++;
for (j=0; j < NumEntries; j++) yp[RowIndices[j]] += RowValues[j] * xp[i];
}
if (Importer()!=0) y.Export(*ImportVector_, *Importer(), Add); // Fill y with Values from export vector
}
UpdateFlops(2*NumGlobalNonzeros64());
return(0);
}
