void CSRMatrix::create(CSRMatrix::ValueType type, unsigned m, const CSRMap & elms)
{
m_valType = type;
m_dimension = m;
unsigned nnz = elms.size();
m_numNonZero = nnz;
if(type > tIndexOnly) m_value->create(nnz * type);
m_rowPtr->create((m+1) * 4);
m_colInd->create(nnz * 4);
unsigned * I = rowPtr();
unsigned * J = colInd();
unsigned lastRow = m+2;
unsigned row, col;
unsigned ielm = 0;
CSRMap::const_iterator it = elms.begin();
for(; it !=elms.end(); ++it) {
row = it->first/m;
col = it->first%m;
if(row != lastRow) {
I[row] = ielm;
lastRow = row;
}
J[it->second] = col;
ielm++;
}
I[m] = ielm;
}
void CudaCSRMatrix::initOnDevice()
{
m_deviceValue->create(numNonZero() * valueType());
m_deviceRowPtr->create((dimension() + 1) * 4);
m_deviceColInd->create(numNonZero() * 4);
m_deviceRowPtr->hostToDevice(rowPtr());
m_deviceColInd->hostToDevice(colInd());
}
GraphType* FinleyDomain::createTrilinosGraph(bool reducedOrder) const
{
index_t myNumTargets;
index_t numTargets;
const index_t* target;
const_TrilinosMap_ptr rowMap;
const_TrilinosMap_ptr colMap;
if (reducedOrder) {
myNumTargets = m_nodes->getNumReducedDegreesOfFreedom();
numTargets = m_nodes->getNumReducedDegreesOfFreedomTargets();
target = m_nodes->borrowTargetReducedDegreesOfFreedom();
rowMap = m_nodes->trilinosReducedRowMap;
colMap = m_nodes->trilinosReducedColMap;
} else {
myNumTargets = m_nodes->getNumDegreesOfFreedom();
numTargets = m_nodes->getNumDegreesOfFreedomTargets();
target = m_nodes->borrowTargetDegreesOfFreedom();
rowMap = m_nodes->trilinosRowMap;
colMap = m_nodes->trilinosColMap;
}
boost::scoped_array<IndexList> indexList(new IndexList[numTargets]);
#pragma omp parallel
{
// insert contributions from element matrices into columns in
// index list
IndexList_insertElements(indexList.get(), m_elements, reducedOrder,
target, reducedOrder, target);
IndexList_insertElements(indexList.get(), m_faceElements,
reducedOrder, target, reducedOrder, target);
IndexList_insertElements(indexList.get(), m_contactElements,
reducedOrder, target, reducedOrder, target);
IndexList_insertElements(indexList.get(), m_points, reducedOrder,
target, reducedOrder, target);
}
Teuchos::ArrayRCP<size_t> rowPtr(myNumTargets + 1);
for (size_t i = 0; i < myNumTargets; i++) {
rowPtr[i+1] = rowPtr[i] + indexList[i].count(0, numTargets);
}
Teuchos::ArrayRCP<LO> colInd(rowPtr[myNumTargets]);
#pragma omp parallel for
for (index_t i = 0; i < myNumTargets; i++) {
indexList[i].toArray(&colInd[rowPtr[i]], 0, numTargets, 0);
std::sort(&colInd[rowPtr[i]], &colInd[rowPtr[i+1]]);
}
GraphType* graph = new GraphType(rowMap, colMap, rowPtr, colInd);
Teuchos::RCP<Teuchos::ParameterList> params = Teuchos::parameterList();
params->set("Optimize Storage", true);
graph->fillComplete(rowMap, rowMap, params);
return graph;
}
void CSRMatrix::verbose()
{
unsigned lastRow = m_dimension + 2;
unsigned i, j;
unsigned * row = rowPtr();
j = 0;
for(i=0; i< m_numNonZero; i++) {
if(i == *row) {
std::cout<<"\n row"<<j<<" ("<<*row<<") ";
j++;
row++;
}
std::cout<<" "<<colInd()[i];
}
std::cout<<"\nn non-zero "<<*row<<"\n";
std::cout<<"dimension "<<m_dimension<<"\n";
std::cout<<"value size "<<m_valType<<"\n";
}
double Ifpack_FrobeniusNorm(const Epetra_RowMatrix& A)
{
double MyNorm = 0.0, GlobalNorm;
std::vector<int> colInd(A.MaxNumEntries());
std::vector<double> colVal(A.MaxNumEntries());
for (int i = 0 ; i < A.NumMyRows() ; ++i) {
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
MyNorm += colVal[j] * colVal[j];
}
}
A.Comm().SumAll(&MyNorm,&GlobalNorm,1);
return(sqrt(GlobalNorm));
}
// ============================================================================
int ML_Epetra::MultiLevelPreconditioner::
CreateAuxiliaryMatrixCrs(Epetra_FECrsMatrix* &FakeMatrix)
{
int NumMyRows = RowMatrix_->NumMyRows();
const Epetra_Map& RowMap = RowMatrix_->RowMatrixRowMap();
const Epetra_Map& ColMap = RowMatrix_->RowMatrixColMap();
FakeMatrix = new Epetra_FECrsMatrix(Copy,RowMap,
2*RowMatrix_->MaxNumEntries());
if (FakeMatrix == 0)
ML_CHK_ERR(-1); // something went wrong
int NumDimensions = 0;
double* x_coord = List_.get("x-coordinates", (double *)0);
if (x_coord != 0) ++NumDimensions;
// at least x-coordinates must be not null
if( NumDimensions == 0 ) {
std::cerr << ErrorMsg_ << "Option `aggregation: use auxiliary matrix' == true" << std::endl
<< ErrorMsg_ << "requires x-, y-, or z-coordinates." << std::endl
<< ErrorMsg_ << "You must specify them using options" << std::endl
<< ErrorMsg_ << "`x-coordinates' (and equivalently for" << std::endl
<< ErrorMsg_ << "y- and z-." << std::endl;
ML_CHK_ERR(-2); // wrong parameters
}
double* y_coord = List_.get("y-coordinates", (double *)0);
if (y_coord != 0) ++NumDimensions;
double* z_coord = List_.get("z-coordinates", (double *)0);
if (z_coord != 0) ++NumDimensions;
// small check to avoid strange behavior
if( z_coord != 0 && y_coord == 0 ) {
std::cerr << ErrorMsg_ << "Something wrong: `y-coordinates'" << std::endl
<< ErrorMsg_ << "is null, while `z-coordinates' is null" << std::endl;
ML_CHK_ERR(-3); // something went wrong
}
double theta = List_.get("aggregation: theta",0.0);
bool SymmetricPattern = List_.get("aggregation: use symmetric pattern",false);
// usual crap to clutter the output
if( verbose_ ) {
std::cout << std::endl;
std::cout << PrintMsg_ << "*** Using auxiliary matrix to create the aggregates" << std::endl
<< PrintMsg_ << "*** Number of dimensions = " << NumDimensions << std::endl
<< PrintMsg_ << "*** theta = " << theta;
if( SymmetricPattern ) std::cout << ", using symmetric pattern" << std::endl;
else std::cout << ", using original pattern" << std::endl;
std::cout << std::endl;
}
// create vectors containing coordinates, replicated for all unknonws
// FIXME: I don't really need Z in all cases
// for west
// I am over-allocating, for large number of equations per node
// this is not optimal. However, it is a only-once importing
// of some more data. It should harm too much...
//
// The following will work for constant number of equations per
// node only. The fix should be easy, though.
Epetra_Vector RowX(RowMap); RowX.PutScalar(0.0);
Epetra_Vector RowY(RowMap); RowY.PutScalar(0.0);
Epetra_Vector RowZ(RowMap); RowZ.PutScalar(0.0);
for (int i = 0 ; i < NumMyRows ; i += NumPDEEqns_) {
RowX[i] = x_coord[i / NumPDEEqns_];
if (NumDimensions > 1) RowY[i] = y_coord[i / NumPDEEqns_];
if (NumDimensions > 2) RowZ[i] = z_coord[i / NumPDEEqns_];
}
// create vectors containing coordinates for columns
// (this is useful only if MIS/ParMETIS are used)
Epetra_Vector ColX(ColMap); ColX.PutScalar(0.0);
Epetra_Vector ColY(ColMap); ColY.PutScalar(0.0);
Epetra_Vector ColZ(ColMap); ColZ.PutScalar(0.0);
// get coordinates for non-local nodes (in column map)
Epetra_Import Importer(ColMap,RowMap);
ColX.Import(RowX,Importer,Insert);
if (NumDimensions > 1) ColY.Import(RowY,Importer,Insert);
if (NumDimensions > 2) ColZ.Import(RowZ,Importer,Insert);
// global row and column numbering
int* MyGlobalRowElements = RowMap.MyGlobalElements();
int* MyGlobalColElements = ColMap.MyGlobalElements();
// room for getrow()
int MaxNnz = RowMatrix_->MaxNumEntries();
std::vector<int> colInd(MaxNnz);
std::vector<double> colVal(MaxNnz);
std::vector<double> coord_i(3);
std::vector<double> coord_j(3);
// =================== //
// cycle over all rows //
// =================== //
for (int i = 0; i < NumMyRows ; i += NumPDEEqns_) {
int GlobalRow = MyGlobalRowElements[i];
if( i%NumPDEEqns_ == 0 ) { // do it just once for each block row
switch( NumDimensions ) {
case 3:
coord_i[2] = RowZ[i];
case 2:
coord_i[1] = RowY[i];
case 1:
coord_i[0] = RowX[i];
}
int NumEntries;
ML_CHK_ERR(RowMatrix_->ExtractMyRowCopy(i,MaxNnz,NumEntries,
&colVal[0],&colInd[0]));
// NOTE: for VBR matrices, the "real" value that will be used in
// the subsequent part of the code is only the one for the first
// equations. For each block, I replace values with the sum of
// the std::abs of each block entry.
for (int j = 0 ; j < NumEntries ; j += NumPDEEqns_) {
colVal[j] = std::fabs(colVal[j]);
for (int k = 1 ; k < NumPDEEqns_ ; ++k) {
colVal[j] += std::fabs(colVal[j+k]);
}
}
// work only on the first equations. Theta will blend the
// coordinate part with the sub of std::abs of row elements.
int GlobalCol;
double total = 0.0;
for (int j = 0 ; j < NumEntries ; j += NumPDEEqns_) {
if (colInd[j] >= NumMyRows)
continue;
if (colInd[j]%NumPDEEqns_ == 0) {
// insert diagonal later
if (colInd[j] != i) {
// get coordinates of this node
switch (NumDimensions) {
case 3:
coord_j[2] = ColZ[colInd[j]];
case 2:
coord_j[1] = ColY[colInd[j]];
case 1:
coord_j[0] = ColX[colInd[j]];
}
// d2 is the square of the distance between node `i' and
// node `j'
double d2 = (coord_i[0] - coord_j[0]) * (coord_i[0] - coord_j[0]) +
(coord_i[1] - coord_j[1]) * (coord_i[1] - coord_j[1]) +
(coord_i[2] - coord_j[2]) * (coord_i[2] - coord_j[2]);
if (d2 == 0.0) {
std::cerr << std::endl;
std::cerr << ErrorMsg_ << "distance between node " << i/NumPDEEqns_ << " and node "
<< colInd[j]/NumPDEEqns_ << std::endl
<< ErrorMsg_ << "is zero. Coordinates of these nodes are" << std::endl
<< ErrorMsg_ << "x_i = " << coord_i[0] << ", x_j = " << coord_j[0] << std::endl
<< ErrorMsg_ << "y_i = " << coord_i[1] << ", y_j = " << coord_j[1] << std::endl
<< ErrorMsg_ << "z_i = " << coord_i[2] << ", z_j = " << coord_j[2] << std::endl
<< ErrorMsg_ << "Now proceeding with distance = 1.0" << std::endl;
std::cerr << std::endl;
d2 = 1.0;
}
// blend d2 with the actual values of the matrix
// FIXME: am I useful?
double val = -(1.0 - theta) * (1.0 / d2) + theta * (colVal[j]);
GlobalCol = MyGlobalColElements[colInd[j]];
// insert this value on all rows
for (int k = 0 ; k < NumPDEEqns_ ; ++k) {
int row = GlobalRow + k;
int col = GlobalCol + k;
if (FakeMatrix->SumIntoGlobalValues(1,&row,1,&col,&val) != 0) {
ML_CHK_ERR(FakeMatrix->InsertGlobalValues(1,&row,1,&col,&val));
}
}
total -= val;
// put (j,i) element as well, only for in-process stuff.
// I have some problems with off-processor elements.
// It is here that I need the FE matrix.
if (SymmetricPattern == true && colInd[j] < NumMyRows ) {
for( int k=0 ; k<NumPDEEqns_ ; ++k ) {
int row = GlobalCol+k;
int col = GlobalRow+k;
if( FakeMatrix->SumIntoGlobalValues(1,&row,1,&col,&val) != 0 ) {
ML_CHK_ERR(FakeMatrix->InsertGlobalValues(1,&row,1,&col,&val));
}
}
total -= val;
}
}
}
}
// create lines with zero-row sum
for (int k = 0 ; k < NumPDEEqns_ ; ++k) {
int row = GlobalRow + k;
if (FakeMatrix->SumIntoGlobalValues(1,&row,1,&row,&total) != 0) {
if (FakeMatrix->InsertGlobalValues(1,&row,1,&row,&total) != 0)
ML_CHK_ERR(-9); // something went wrong
}
}
}
}
if (FakeMatrix->FillComplete())
ML_CHK_ERR(-5); // something went wrong
// stick pointer in Amat for level 0 (finest level)
ml_->Amat[LevelID_[0]].data = (void *)FakeMatrix;
// tell ML to keep the tentative prolongator
ML_Aggregate_Set_Reuse(agg_);
// pray that no bugs will tease us
return(0);
}
// ============================================================================
int ML_Epetra::MultiLevelPreconditioner::
CreateAuxiliaryMatrixVbr(Epetra_VbrMatrix* &FakeMatrix)
{
// FakeMatrix has already been created before
if (FakeMatrix == 0)
ML_CHK_ERR(-1); // something went wrong
int NumDimensions = 0;
double* x_coord = List_.get("x-coordinates", (double *)0);
if (x_coord != 0) ++NumDimensions;
// at least x-coordinates must be not null
if( NumDimensions == 0 ) {
std::cerr << ErrorMsg_ << "Option `aggregation: use auxiliary matrix' == true" << std::endl
<< ErrorMsg_ << "requires x-, y-, or z-coordinates." << std::endl
<< ErrorMsg_ << "You must specify them using options" << std::endl
<< ErrorMsg_ << "`x-coordinates' (and equivalently for" << std::endl
<< ErrorMsg_ << "y- and z-)." << std::endl;
ML_CHK_ERR(-2); // wrong parameters
}
double* y_coord = List_.get("y-coordinates", (double *)0);
if (y_coord != 0) ++NumDimensions;
double* z_coord = List_.get("z-coordinates", (double *)0);
if (z_coord != 0) ++NumDimensions;
// small check to avoid strange behavior
if( z_coord != 0 && y_coord == 0 ) {
std::cerr << ErrorMsg_ << "Something wrong: `y-coordinates'" << std::endl
<< ErrorMsg_ << "is null, while `z-coordinates' is not null" << std::endl;
ML_CHK_ERR(-3); // something went wrong
}
// usual crap to clutter the output
if( verbose_ ) {
std::cout << std::endl;
std::cout << PrintMsg_ << "*** Using auxiliary matrix to create the aggregates" << std::endl
<< PrintMsg_ << "*** Number of dimensions = " << NumDimensions << std::endl
<< PrintMsg_ << "*** (the version for Epetra_VbrMatrix is currently used)" << std::endl;
std::cout << std::endl;
}
const Epetra_BlockMap& RowMap = FakeMatrix->RowMap();
const Epetra_BlockMap& ColMap = FakeMatrix->ColMap();
int NumMyRowElements = RowMap.NumMyElements();
int NumMyColElements = ColMap.NumMyElements();
int* MyGlobalRowElements = RowMap.MyGlobalElements();
int* MyGlobalColElements = ColMap.MyGlobalElements();
// use point map to exchange coordinates
Epetra_Map PointRowMap(-1,NumMyRowElements,MyGlobalRowElements,0,Comm());
Epetra_Map PointColMap(-1,NumMyColElements,MyGlobalColElements,0,Comm());
Epetra_Vector RowX(PointRowMap);
Epetra_Vector RowY(PointRowMap);
Epetra_Vector RowZ(PointRowMap);
for (int i = 0 ; i < NumMyRowElements ; ++i) {
RowX[i] = x_coord[i];
if (NumDimensions > 1) RowY[i] = y_coord[i];
if (NumDimensions > 2) RowZ[i] = z_coord[i];
}
// create vectors containing coordinates for columns
// (this is useful only if MIS/ParMETIS are used)
Epetra_Vector ColX(PointColMap);
Epetra_Vector ColY(PointColMap);
Epetra_Vector ColZ(PointColMap);
// get coordinates for non-local nodes (in column map)
Epetra_Import Importer(PointColMap,PointRowMap);
ColX.Import(RowX,Importer,Insert);
if (NumDimensions > 1) ColY.Import(RowY,Importer,Insert);
if (NumDimensions > 2) ColZ.Import(RowZ,Importer,Insert);
// room for getrow()
int MaxNnz = FakeMatrix->MaxNumEntries();
std::vector<int> colInd(MaxNnz);
std::vector<double> colVal(MaxNnz);
std::vector<double> coord_i(3);
std::vector<double> coord_j(3);
// change the entries of FakeMatrix so that it corresponds to a discrete
// Laplacian. Note: This is not exactly the same as in the Crs case.
FakeMatrix->PutScalar(0.0);
for (int LocalRow = 0; LocalRow < NumMyRowElements ; ++LocalRow) {
int RowDim, NumBlockEntries;
int* BlockIndices;
Epetra_SerialDenseMatrix** RowValues;
switch (NumDimensions) {
case 3:
coord_i[2] = RowZ[LocalRow];
case 2:
coord_i[1] = RowY[LocalRow];
case 1:
coord_i[0] = RowX[LocalRow];
}
FakeMatrix->ExtractMyBlockRowView(LocalRow,RowDim,NumBlockEntries,
BlockIndices,RowValues);
// accumulator for zero row-sum
double total = 0.0;
for (int j = 0 ; j < NumBlockEntries ; ++j) {
int LocalCol = BlockIndices[j];
// insert diagonal later
if (LocalCol != LocalRow) {
// get coordinates of this node
switch (NumDimensions) {
case 3:
coord_j[2] = ColZ[LocalCol];
case 2:
coord_j[1] = ColY[LocalCol];
case 1:
coord_j[0] = ColX[LocalCol];
}
// d2 is the square of the distance between node `i' and
// node `j'
double d2 = (coord_i[0] - coord_j[0]) * (coord_i[0] - coord_j[0]) +
(coord_i[1] - coord_j[1]) * (coord_i[1] - coord_j[1]) +
(coord_i[2] - coord_j[2]) * (coord_i[2] - coord_j[2]);
if (d2 == 0.0) {
std::cerr << std::endl;
std::cerr << ErrorMsg_ << "distance between node " << LocalRow << " and node "
<< LocalCol << std::endl
<< ErrorMsg_ << "is zero. Coordinates of these nodes are" << std::endl
<< ErrorMsg_ << "x_i = " << coord_i[0] << ", x_j = " << coord_j[0] << std::endl
<< ErrorMsg_ << "y_i = " << coord_i[1] << ", y_j = " << coord_j[1] << std::endl
<< ErrorMsg_ << "z_i = " << coord_i[2] << ", z_j = " << coord_j[2] << std::endl
<< ErrorMsg_ << "Now proceeding with distance = 1.0" << std::endl;
std::cerr << std::endl;
d2 = 1.0;
}
for (int k = 0 ; k < RowValues[j]->M() ; ++k) {
for (int h = 0 ; h < RowValues[j]->N() ; ++h) {
if (k == h)
(*RowValues[j])(k,h) = - 1.0 / d2;
}
}
total += 1.0 /d2;
}
}
// check that the diagonal block exists
bool ok = false;
int DiagonalBlock = 0;
for (int j = 0 ; j < NumBlockEntries ; ++j) {
if (BlockIndices[j] == LocalRow) {
DiagonalBlock = j;
ok = true;
break;
}
}
assert (ok == true);
for (int k = 0 ; k < RowValues[DiagonalBlock]->N() ; ++k) {
(*RowValues[DiagonalBlock])(k,k) = total;
}
}
// tell ML to keep the tentative prolongator
ML_Aggregate_Set_Reuse(agg_);
// pray that no bugs will tease us
return(0);
}
int Ifpack_AnalyzeMatrixElements(const Epetra_RowMatrix& A,
const bool abs, const int steps)
{
bool verbose = (A.Comm().MyPID() == 0);
double min_val = DBL_MAX;
double max_val = -DBL_MAX;
std::vector<int> colInd(A.MaxNumEntries());
std::vector<double> colVal(A.MaxNumEntries());
for (int i = 0 ; i < A.NumMyRows() ; ++i) {
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
double v = colVal[j];
if (abs)
if (v < 0) v = -v;
if (v < min_val)
min_val = v;
if (v > max_val)
max_val = v;
}
}
if (verbose) {
cout << endl;
Ifpack_PrintLine();
cout << "Label of matrix = " << A.Label() << endl;
cout << endl;
}
double delta = (max_val - min_val) / steps;
for (int k = 0 ; k < steps ; ++k) {
double below = delta * k + min_val;
double above = below + delta;
int MyBelow = 0, GlobalBelow;
for (int i = 0 ; i < A.NumMyRows() ; ++i) {
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
double v = colVal[j];
if (abs)
if (v < 0) v = -v;
if (v >= below && v < above) MyBelow++;
}
}
A.Comm().SumAll(&MyBelow, &GlobalBelow, 1);
if (verbose) {
printf("Elements in [%+7e, %+7e) = %10d ( = %5.2f %%)\n",
below, above, GlobalBelow,
100.0 * GlobalBelow / A.NumGlobalNonzeros64());
}
}
if (verbose) {
Ifpack_PrintLine();
cout << endl;
}
return(0);
}
int Ifpack_Analyze(const Epetra_RowMatrix& A, const bool Cheap,
const int NumPDEEqns)
{
int NumMyRows = A.NumMyRows();
long long NumGlobalRows = A.NumGlobalRows64();
long long NumGlobalCols = A.NumGlobalCols64();
long long MyBandwidth = 0, GlobalBandwidth;
long long MyLowerNonzeros = 0, MyUpperNonzeros = 0;
long long GlobalLowerNonzeros, GlobalUpperNonzeros;
long long MyDiagonallyDominant = 0, GlobalDiagonallyDominant;
long long MyWeaklyDiagonallyDominant = 0, GlobalWeaklyDiagonallyDominant;
double MyMin, MyAvg, MyMax;
double GlobalMin, GlobalAvg, GlobalMax;
long long GlobalStorage;
bool verbose = (A.Comm().MyPID() == 0);
GlobalStorage = sizeof(int*) * NumGlobalRows +
sizeof(int) * A.NumGlobalNonzeros64() +
sizeof(double) * A.NumGlobalNonzeros64();
if (verbose) {
print();
Ifpack_PrintLine();
print<const char*>("Label", A.Label());
print<long long>("Global rows", NumGlobalRows);
print<long long>("Global columns", NumGlobalCols);
print<long long>("Stored nonzeros", A.NumGlobalNonzeros64());
print<long long>("Nonzeros / row", A.NumGlobalNonzeros64() / NumGlobalRows);
print<double>("Estimated storage (Mbytes)", 1.0e-6 * GlobalStorage);
}
long long NumMyActualNonzeros = 0, NumGlobalActualNonzeros;
long long NumMyEmptyRows = 0, NumGlobalEmptyRows;
long long NumMyDirichletRows = 0, NumGlobalDirichletRows;
std::vector<int> colInd(A.MaxNumEntries());
std::vector<double> colVal(A.MaxNumEntries());
Epetra_Vector Diag(A.RowMatrixRowMap());
Epetra_Vector RowSum(A.RowMatrixRowMap());
Diag.PutScalar(0.0);
RowSum.PutScalar(0.0);
for (int i = 0 ; i < NumMyRows ; ++i) {
long long GRID = A.RowMatrixRowMap().GID64(i);
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
if (Nnz == 0)
NumMyEmptyRows++;
if (Nnz == 1)
NumMyDirichletRows++;
for (int j = 0 ; j < Nnz ; ++j) {
double v = colVal[j];
if (v < 0) v = -v;
if (colVal[j] != 0.0)
NumMyActualNonzeros++;
long long GCID = A.RowMatrixColMap().GID64(colInd[j]);
if (GCID != GRID)
RowSum[i] += v;
else
Diag[i] = v;
if (GCID < GRID)
MyLowerNonzeros++;
else if (GCID > GRID)
MyUpperNonzeros++;
long long b = GCID - GRID;
if (b < 0) b = -b;
if (b > MyBandwidth)
MyBandwidth = b;
}
if (Diag[i] > RowSum[i])
MyDiagonallyDominant++;
if (Diag[i] >= RowSum[i])
MyWeaklyDiagonallyDominant++;
RowSum[i] += Diag[i];
}
// ======================== //
// summing up global values //
// ======================== //
A.Comm().SumAll(&MyDiagonallyDominant,&GlobalDiagonallyDominant,1);
A.Comm().SumAll(&MyWeaklyDiagonallyDominant,&GlobalWeaklyDiagonallyDominant,1);
A.Comm().SumAll(&NumMyActualNonzeros, &NumGlobalActualNonzeros, 1);
A.Comm().SumAll(&NumMyEmptyRows, &NumGlobalEmptyRows, 1);
A.Comm().SumAll(&NumMyDirichletRows, &NumGlobalDirichletRows, 1);
A.Comm().SumAll(&MyBandwidth, &GlobalBandwidth, 1);
A.Comm().SumAll(&MyLowerNonzeros, &GlobalLowerNonzeros, 1);
A.Comm().SumAll(&MyUpperNonzeros, &GlobalUpperNonzeros, 1);
A.Comm().SumAll(&MyDiagonallyDominant, &GlobalDiagonallyDominant, 1);
A.Comm().SumAll(&MyWeaklyDiagonallyDominant, &GlobalWeaklyDiagonallyDominant, 1);
double NormOne = A.NormOne();
double NormInf = A.NormInf();
double NormF = Ifpack_FrobeniusNorm(A);
if (verbose) {
print();
print<long long>("Actual nonzeros", NumGlobalActualNonzeros);
print<long long>("Nonzeros in strict lower part", GlobalLowerNonzeros);
print<long long>("Nonzeros in strict upper part", GlobalUpperNonzeros);
print();
print<long long>("Empty rows", NumGlobalEmptyRows,
100.0 * NumGlobalEmptyRows / NumGlobalRows);
print<long long>("Dirichlet rows", NumGlobalDirichletRows,
100.0 * NumGlobalDirichletRows / NumGlobalRows);
print<long long>("Diagonally dominant rows", GlobalDiagonallyDominant,
100.0 * GlobalDiagonallyDominant / NumGlobalRows);
print<long long>("Weakly diag. dominant rows",
GlobalWeaklyDiagonallyDominant,
100.0 * GlobalWeaklyDiagonallyDominant / NumGlobalRows);
print();
print<long long>("Maximum bandwidth", GlobalBandwidth);
print();
print("", "one-norm", "inf-norm", "Frobenius", false);
print("", "========", "========", "=========", false);
print();
print<double>("A", NormOne, NormInf, NormF);
}
if (Cheap == false) {
// create A + A^T and A - A^T
Epetra_FECrsMatrix AplusAT(Copy, A.RowMatrixRowMap(), 0);
Epetra_FECrsMatrix AminusAT(Copy, A.RowMatrixRowMap(), 0);
#ifndef EPETRA_NO_32BIT_GLOBAL_INDICES
if(A.RowMatrixRowMap().GlobalIndicesInt()) {
for (int i = 0 ; i < NumMyRows ; ++i) {
int GRID = A.RowMatrixRowMap().GID(i);
assert (GRID != -1);
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
int GCID = A.RowMatrixColMap().GID(colInd[j]);
assert (GCID != -1);
double plus_val = colVal[j];
double minus_val = -colVal[j];
if (AplusAT.SumIntoGlobalValues(1,&GRID,1,&GCID,&plus_val) != 0) {
IFPACK_CHK_ERR(AplusAT.InsertGlobalValues(1,&GRID,1,&GCID,&plus_val));
}
if (AplusAT.SumIntoGlobalValues(1,&GCID,1,&GRID,&plus_val) != 0) {
IFPACK_CHK_ERR(AplusAT.InsertGlobalValues(1,&GCID,1,&GRID,&plus_val));
}
if (AminusAT.SumIntoGlobalValues(1,&GRID,1,&GCID,&plus_val) != 0) {
IFPACK_CHK_ERR(AminusAT.InsertGlobalValues(1,&GRID,1,&GCID,&plus_val));
}
if (AminusAT.SumIntoGlobalValues(1,&GCID,1,&GRID,&minus_val) != 0) {
IFPACK_CHK_ERR(AminusAT.InsertGlobalValues(1,&GCID,1,&GRID,&minus_val));
}
}
}
}
else
#endif
#ifndef EPETRA_NO_64BIT_GLOBAL_INDICES
if(A.RowMatrixRowMap().GlobalIndicesLongLong()) {
for (int i = 0 ; i < NumMyRows ; ++i) {
long long GRID = A.RowMatrixRowMap().GID64(i);
assert (GRID != -1);
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
long long GCID = A.RowMatrixColMap().GID64(colInd[j]);
assert (GCID != -1);
double plus_val = colVal[j];
double minus_val = -colVal[j];
if (AplusAT.SumIntoGlobalValues(1,&GRID,1,&GCID,&plus_val) != 0) {
IFPACK_CHK_ERR(AplusAT.InsertGlobalValues(1,&GRID,1,&GCID,&plus_val));
}
if (AplusAT.SumIntoGlobalValues(1,&GCID,1,&GRID,&plus_val) != 0) {
IFPACK_CHK_ERR(AplusAT.InsertGlobalValues(1,&GCID,1,&GRID,&plus_val));
}
if (AminusAT.SumIntoGlobalValues(1,&GRID,1,&GCID,&plus_val) != 0) {
IFPACK_CHK_ERR(AminusAT.InsertGlobalValues(1,&GRID,1,&GCID,&plus_val));
}
if (AminusAT.SumIntoGlobalValues(1,&GCID,1,&GRID,&minus_val) != 0) {
IFPACK_CHK_ERR(AminusAT.InsertGlobalValues(1,&GCID,1,&GRID,&minus_val));
}
}
}
}
else
#endif
throw "Ifpack_Analyze: GlobalIndices type unknown";
AplusAT.FillComplete();
AminusAT.FillComplete();
AplusAT.Scale(0.5);
AminusAT.Scale(0.5);
NormOne = AplusAT.NormOne();
NormInf = AplusAT.NormInf();
NormF = Ifpack_FrobeniusNorm(AplusAT);
if (verbose) {
print<double>("A + A^T", NormOne, NormInf, NormF);
}
NormOne = AminusAT.NormOne();
NormInf = AminusAT.NormInf();
NormF = Ifpack_FrobeniusNorm(AminusAT);
if (verbose) {
print<double>("A - A^T", NormOne, NormInf, NormF);
}
}
if (verbose) {
print();
print<const char*>("", "min", "avg", "max", false);
print<const char*>("", "===", "===", "===", false);
}
MyMax = -DBL_MAX;
MyMin = DBL_MAX;
MyAvg = 0.0;
for (int i = 0 ; i < NumMyRows ; ++i) {
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
MyAvg += colVal[j];
if (colVal[j] > MyMax) MyMax = colVal[j];
if (colVal[j] < MyMin) MyMin = colVal[j];
}
}
A.Comm().MaxAll(&MyMax, &GlobalMax, 1);
A.Comm().MinAll(&MyMin, &GlobalMin, 1);
A.Comm().SumAll(&MyAvg, &GlobalAvg, 1);
GlobalAvg /= A.NumGlobalNonzeros64();
if (verbose) {
print();
print<double>(" A(i,j)", GlobalMin, GlobalAvg, GlobalMax);
}
MyMax = 0.0;
MyMin = DBL_MAX;
MyAvg = 0.0;
for (int i = 0 ; i < NumMyRows ; ++i) {
int Nnz;
IFPACK_CHK_ERR(A.ExtractMyRowCopy(i,A.MaxNumEntries(),Nnz,
&colVal[0],&colInd[0]));
for (int j = 0 ; j < Nnz ; ++j) {
double v = colVal[j];
if (v < 0) v = -v;
MyAvg += v;
if (colVal[j] > MyMax) MyMax = v;
if (colVal[j] < MyMin) MyMin = v;
}
}
A.Comm().MaxAll(&MyMax, &GlobalMax, 1);
A.Comm().MinAll(&MyMin, &GlobalMin, 1);
A.Comm().SumAll(&MyAvg, &GlobalAvg, 1);
GlobalAvg /= A.NumGlobalNonzeros64();
if (verbose) {
print<double>("|A(i,j)|", GlobalMin, GlobalAvg, GlobalMax);
}
// ================= //
// diagonal elements //
// ================= //
Diag.MinValue(&GlobalMin);
Diag.MaxValue(&GlobalMax);
Diag.MeanValue(&GlobalAvg);
if (verbose) {
print();
print<double>(" A(k,k)", GlobalMin, GlobalAvg, GlobalMax);
}
Diag.Abs(Diag);
Diag.MinValue(&GlobalMin);
Diag.MaxValue(&GlobalMax);
Diag.MeanValue(&GlobalAvg);
if (verbose) {
print<double>("|A(k,k)|", GlobalMin, GlobalAvg, GlobalMax);
}
// ============================================== //
// cycle over all equations for diagonal elements //
// ============================================== //
if (NumPDEEqns > 1 ) {
if (verbose) print();
for (int ie = 0 ; ie < NumPDEEqns ; ie++) {
MyMin = DBL_MAX;
MyMax = -DBL_MAX;
MyAvg = 0.0;
for (int i = ie ; i < Diag.MyLength() ; i += NumPDEEqns) {
double d = Diag[i];
MyAvg += d;
if (d < MyMin)
MyMin = d;
if (d > MyMax)
MyMax = d;
}
A.Comm().MinAll(&MyMin, &GlobalMin, 1);
A.Comm().MaxAll(&MyMax, &GlobalMax, 1);
A.Comm().SumAll(&MyAvg, &GlobalAvg, 1);
// does not really work fine if the number of global
// elements is not a multiple of NumPDEEqns
GlobalAvg /= (Diag.GlobalLength64() / NumPDEEqns);
if (verbose) {
char str[80];
sprintf(str, " A(k,k), eq %d", ie);
print<double>(str, GlobalMin, GlobalAvg, GlobalMax);
}
}
}
// ======== //
// row sums //
// ======== //
RowSum.MinValue(&GlobalMin);
RowSum.MaxValue(&GlobalMax);
RowSum.MeanValue(&GlobalAvg);
if (verbose) {
print();
print<double>(" sum_j A(k,j)", GlobalMin, GlobalAvg, GlobalMax);
}
// ===================================== //
// cycle over all equations for row sums //
// ===================================== //
if (NumPDEEqns > 1 ) {
if (verbose) print();
for (int ie = 0 ; ie < NumPDEEqns ; ie++) {
MyMin = DBL_MAX;
MyMax = -DBL_MAX;
MyAvg = 0.0;
for (int i = ie ; i < Diag.MyLength() ; i += NumPDEEqns) {
double d = RowSum[i];
MyAvg += d;
if (d < MyMin)
MyMin = d;
if (d > MyMax)
MyMax = d;
}
A.Comm().MinAll(&MyMin, &GlobalMin, 1);
A.Comm().MaxAll(&MyMax, &GlobalMax, 1);
A.Comm().SumAll(&MyAvg, &GlobalAvg, 1);
// does not really work fine if the number of global
// elements is not a multiple of NumPDEEqns
GlobalAvg /= (Diag.GlobalLength64() / NumPDEEqns);
if (verbose) {
char str[80];
sprintf(str, " sum_j A(k,j), eq %d", ie);
print<double>(str, GlobalMin, GlobalAvg, GlobalMax);
}
}
}
if (verbose)
Ifpack_PrintLine();
return(0);
}