void RAPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::CheckMainDiagonal(RCP<Matrix> & Ac) const {
// plausibility check: no zeros on diagonal
RCP<Vector> diagVec = VectorFactory::Build(Ac->getRowMap());
Ac->getLocalDiagCopy(*diagVec);
SC zero = Teuchos::ScalarTraits<SC>::zero(), one = Teuchos::ScalarTraits<SC>::one();
LO lZeroDiags = 0;
Teuchos::ArrayRCP< Scalar > diagVal = diagVec->getDataNonConst(0);
for (size_t r = 0; r < Ac->getRowMap()->getNodeNumElements(); r++) {
if (diagVal[r] == zero) {
lZeroDiags++;
if (repairZeroDiagonals_) {
GO grid = Ac->getRowMap()->getGlobalElement(r);
LO lcid = Ac->getColMap()->getLocalElement(grid);
Teuchos::ArrayRCP<LO> indout(1, lcid);
Teuchos::ArrayRCP<SC> valout(1, one);
Ac->insertLocalValues(r, indout.view(0, indout.size()), valout.view(0, valout.size()));
}
}
}
if (IsPrint(Warnings0)) {
const RCP<const Teuchos::Comm<int> > & comm = Ac->getRowMap()->getComm();
GO lZeroDiagsGO = Teuchos::as<GO>(lZeroDiags); /* LO->GO conversion */
GO gZeroDiags = 0;
sumAll(comm, lZeroDiagsGO, gZeroDiags);
if (repairZeroDiagonals_) GetOStream(Warnings0,0) << "RAPFactory (WARNING): repaired " << gZeroDiags << " zeros on main diagonal of Ac." << std::endl;
else GetOStream(Warnings0,0) << "RAPFactory (WARNING): found " << gZeroDiags << " zeros on main diagonal of Ac." << std::endl;
}
}
void DropNegativeEntriesFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Build(Level& currentLevel) const {
FactoryMonitor m(*this, "Matrix filtering (springs)", currentLevel);
RCP<Matrix> Ain = Get< RCP<Matrix> >(currentLevel, "A");
LocalOrdinal nDofsPerNode = Ain->GetFixedBlockSize();
// create new empty Operator
Teuchos::RCP<Matrix> Aout = MatrixFactory::Build(Ain->getRowMap(), Ain->getGlobalMaxNumRowEntries(), Xpetra::StaticProfile);
size_t numLocalRows = Ain->getNodeNumRows();
for(size_t row=0; row<numLocalRows; row++) {
GlobalOrdinal grid = Ain->getRowMap()->getGlobalElement(row);
int rDofID = Teuchos::as<int>(grid % nDofsPerNode);
// extract row information from input matrix
Teuchos::ArrayView<const LocalOrdinal> indices;
Teuchos::ArrayView<const Scalar> vals;
Ain->getLocalRowView(row, indices, vals);
// just copy all values in output
Teuchos::ArrayRCP<GlobalOrdinal> indout(indices.size(),Teuchos::ScalarTraits<GlobalOrdinal>::zero());
Teuchos::ArrayRCP<Scalar> valout(indices.size(),Teuchos::ScalarTraits<Scalar>::zero());
size_t nNonzeros = 0;
for(size_t i=0; i<(size_t)indices.size(); i++) {
GlobalOrdinal gcid = Ain->getColMap()->getGlobalElement(indices[i]); // global column id
int cDofID = Teuchos::as<int>(gcid % nDofsPerNode);
if(rDofID == cDofID && Teuchos::ScalarTraits<Scalar>::magnitude(vals[i]) >= Teuchos::ScalarTraits<Scalar>::magnitude(Teuchos::ScalarTraits<Scalar>::zero())) {
indout [nNonzeros] = gcid;
valout [nNonzeros] = vals[i];
nNonzeros++;
}
}
indout.resize(nNonzeros);
valout.resize(nNonzeros);
Aout->insertGlobalValues(Ain->getRowMap()->getGlobalElement(row), indout.view(0,indout.size()), valout.view(0,valout.size()));
}
Aout->fillComplete(Ain->getDomainMap(), Ain->getRangeMap());
// copy block size information
Aout->SetFixedBlockSize(nDofsPerNode);
GetOStream(Statistics0, 0) << "Nonzeros in A (input): " << Ain->getGlobalNumEntries() << ", Nonzeros after filtering A: " << Aout->getGlobalNumEntries() << std::endl;
Set(currentLevel, "A", Aout);
}
void BlockedRAPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::CheckMainDiagonal(RCP<BlockedCrsMatrix> & bAc, bool repairZeroDiagonals) {
RCP<CrsMatrix> c00 = bAc->getMatrix(0, 0);
RCP<CrsMatrix> Aout = CrsMatrixFactory::Build(c00->getRowMap(), c00->getGlobalMaxNumRowEntries(), Xpetra::StaticProfile);
RCP<Vector> diagVec = VectorFactory::Build(c00->getRowMap());
c00->getLocalDiagCopy(*diagVec);
ArrayRCP<SC> diagVal = diagVec->getDataNonConst(0);
// loop over local rows
for(size_t row=0; row<c00->getNodeNumRows(); row++) {
// get global row id
GO grid = c00->getRowMap()->getGlobalElement(row); // global row id
ArrayView<const LO> indices;
ArrayView<const SC> vals;
c00->getLocalRowView(row, indices, vals);
// just copy all values in output
ArrayRCP<GO> indout(indices.size(), Teuchos::OrdinalTraits<GO>::zero());
ArrayRCP<SC> valout(indices.size(), Teuchos::ScalarTraits<SC>::zero());
// just copy values
for (size_t i = 0; i < as<size_t>(indices.size()); i++) {
GO gcid = c00->getColMap()->getGlobalElement(indices[i]); // LID -> GID (column)
indout [i] = gcid;
valout [i] = vals[i];
}
Aout->insertGlobalValues(grid, indout.view(0, indout.size()), valout.view(0, valout.size()));
if (diagVal[row] == 0.0 && repairZeroDiagonals) {
// always overwrite diagonal entry
Aout->insertGlobalValues(grid, Teuchos::tuple<GO>(grid), Teuchos::tuple<SC>(1.0));
}
}
Aout->fillComplete(c00->getDomainMap(), c00->getRangeMap());
bAc->setMatrix(0, 0, Aout);
}
void AlgebraicPermutationStrategy<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::BuildPermutation(const Teuchos::RCP<Matrix> & A, const Teuchos::RCP<const Map> permRowMap, Level & currentLevel, const FactoryBase* genFactory) const {
#ifndef HAVE_MUELU_INST_COMPLEX_INT_INT
const Teuchos::RCP< const Teuchos::Comm< int > > comm = A->getRowMap()->getComm();
int numProcs = comm->getSize();
int myRank = comm->getRank();
/*if( permRowMap == Teuchos::null ) {
permRowMap = A->getRowMap(); // use full row map of A
}*/
size_t nDofsPerNode = 1;
if (A->IsView("stridedMaps")) {
Teuchos::RCP<const Map> permRowMapStrided = A->getRowMap("stridedMaps");
nDofsPerNode = Teuchos::rcp_dynamic_cast<const StridedMap>(permRowMapStrided)->getFixedBlockSize();
}
//GetOStream(Runtime0, 0) << "Perform generation of permutation operators on " << mapName_ << " map with " << permRowMap->getGlobalNumElements() << " elements" << std::endl;
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > permutedDiagCandidates;
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > keepDiagonalEntries;
std::vector<Scalar> Weights;
// loop over all local rows in matrix A and keep diagonal entries if corresponding
// matrix rows are not contained in permRowMap
for (size_t row = 0; row < A->getRowMap()->getNodeNumElements(); row++) {
GlobalOrdinal grow = A->getRowMap()->getGlobalElement(row);
if(permRowMap->isNodeGlobalElement(grow) == true) continue;
size_t nnz = A->getNumEntriesInLocalRow(row);
// extract local row information from matrix
Teuchos::ArrayView<const LocalOrdinal> indices;
Teuchos::ArrayView<const Scalar> vals;
A->getLocalRowView(row, indices, vals);
TEUCHOS_TEST_FOR_EXCEPTION(Teuchos::as<size_t>(indices.size()) != nnz, Exceptions::RuntimeError, "MueLu::PermutationFactory::Build: number of nonzeros not equal to number of indices? Error.");
// find column entry with max absolute value
GlobalOrdinal gMaxValIdx = 0;
Scalar norm1 = 0.0;
Scalar maxVal = 0.0;
for (size_t j = 0; j < Teuchos::as<size_t>(indices.size()); j++) {
norm1 += std::abs(vals[j]);
if(std::abs(vals[j]) > maxVal) {
maxVal = std::abs(vals[j]);
gMaxValIdx = A->getColMap()->getGlobalElement(indices[j]);
}
}
if(grow == gMaxValIdx) // only keep row/col pair if it's diagonal dominant!!!
keepDiagonalEntries.push_back(std::make_pair(grow,grow));
}
//////////
// handle rows that are marked to be relevant for permutations
for (size_t row = 0; row < permRowMap->getNodeNumElements(); row++) {
GlobalOrdinal grow = permRowMap->getGlobalElement(row);
LocalOrdinal lArow = A->getRowMap()->getLocalElement(grow);
size_t nnz = A->getNumEntriesInLocalRow(lArow);
// extract local row information from matrix
Teuchos::ArrayView<const LocalOrdinal> indices;
Teuchos::ArrayView<const Scalar> vals;
A->getLocalRowView(lArow, indices, vals);
TEUCHOS_TEST_FOR_EXCEPTION(Teuchos::as<size_t>(indices.size()) != nnz, Exceptions::RuntimeError, "MueLu::PermutationFactory::Build: number of nonzeros not equal to number of indices? Error.");
// find column entry with max absolute value
GlobalOrdinal gMaxValIdx = 0;
Scalar norm1 = 0.0;
Scalar maxVal = 0.0;
for (size_t j = 0; j < Teuchos::as<size_t>(indices.size()); j++) {
norm1 += std::abs(vals[j]);
if(std::abs(vals[j]) > maxVal) {
maxVal = std::abs(vals[j]);
gMaxValIdx = A->getColMap()->getGlobalElement(indices[j]);
}
}
if(std::abs(maxVal) > 0.0) { // keep only max Entries \neq 0.0
permutedDiagCandidates.push_back(std::make_pair(grow,gMaxValIdx));
Weights.push_back(maxVal/(norm1*Teuchos::as<Scalar>(nnz)));
} else {
std::cout << "ATTENTION: row " << grow << " has only zero entries -> singular matrix!" << std::endl;
}
}
// sort Weights in descending order
std::vector<int> permutation;
sortingPermutation(Weights,permutation);
// create new vector with exactly one possible entry for each column
// each processor which requests the global column id gcid adds 1 to gColVec
// gColVec will be summed up over all processors and communicated to gDomVec
// which is based on the non-overlapping domain map of A.
Teuchos::RCP<Vector> gColVec = VectorFactory::Build(A->getColMap());
Teuchos::RCP<Vector> gDomVec = VectorFactory::Build(A->getDomainMap());
gColVec->putScalar(0.0);
gDomVec->putScalar(0.0);
// put in all keep diagonal entries
for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = keepDiagonalEntries.begin(); p != keepDiagonalEntries.end(); ++p) {
gColVec->sumIntoGlobalValue((*p).second,1.0);
}
Teuchos::RCP<Export> exporter = ExportFactory::Build(gColVec->getMap(), gDomVec->getMap());
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD); // communicate blocked gcolids to all procs
gColVec->doImport(*gDomVec,*exporter,Xpetra::INSERT);
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > permutedDiagCandidatesFiltered; // TODO reserve memory
std::map<GlobalOrdinal, Scalar> gColId2Weight;
Teuchos::ArrayRCP< Scalar > ddata = gColVec->getDataNonConst(0);
for(size_t i = 0; i < permutedDiagCandidates.size(); ++i) {
// loop over all candidates
std::pair<GlobalOrdinal, GlobalOrdinal> pp = permutedDiagCandidates[permutation[i]];
GlobalOrdinal grow = pp.first;
GlobalOrdinal gcol = pp.second;
LocalOrdinal lcol = A->getColMap()->getLocalElement(gcol);
//Teuchos::ArrayRCP< Scalar > ddata = gColVec->getDataNonConst(0);
if(ddata[lcol] > 0.0){
continue; // skip lcol: column already handled by another row
}
// mark column as already taken
ddata[lcol]++;
permutedDiagCandidatesFiltered.push_back(std::make_pair(grow,gcol));
gColId2Weight[gcol] = Weights[permutation[i]];
}
// communicate how often each column index is requested by the different procs
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD);
gColVec->doImport(*gDomVec,*exporter,Xpetra::INSERT); // probably not needed // TODO check me
//*****************************************************************************************
// first communicate ALL global ids of column indices which are requested by more
// than one proc to all other procs
// detect which global col indices are requested by more than one proc
// and store them in the multipleColRequests vector
std::vector<GlobalOrdinal> multipleColRequests; // store all global column indices from current processor that are also
// requested by another processor. This is possible, since they are stored
// in gDomVec which is based on the nonoverlapping domain map. That is, each
// global col id is handled by exactly one proc.
std::queue<GlobalOrdinal> unusedColIdx; // unused column indices on current processor
for(size_t sz = 0; sz<gDomVec->getLocalLength(); ++sz) {
Teuchos::ArrayRCP< const Scalar > arrDomVec = gDomVec->getData(0);
if(arrDomVec[sz] > 1.0) {
multipleColRequests.push_back(gDomVec->getMap()->getGlobalElement(sz));
} else if(arrDomVec[sz] == 0.0) {
unusedColIdx.push(gDomVec->getMap()->getGlobalElement(sz));
}
}
// communicate the global number of column indices which are requested by more than one proc
LocalOrdinal localMultColRequests = Teuchos::as<LocalOrdinal>(multipleColRequests.size());
LocalOrdinal globalMultColRequests = 0;
// sum up all entries in multipleColRequests over all processors
sumAll(gDomVec->getMap()->getComm(), (LocalOrdinal)localMultColRequests, globalMultColRequests);
if(globalMultColRequests > 0) {
// special handling: two processors request the same global column id.
// decide which processor gets it
// distribute number of multipleColRequests to all processors
// each processor stores how many column ids for exchange are handled by the cur proc
std::vector<GlobalOrdinal> numMyMultColRequests(numProcs,0);
std::vector<GlobalOrdinal> numGlobalMultColRequests(numProcs,0);
numMyMultColRequests[myRank] = localMultColRequests;
Teuchos::reduceAll(*comm,Teuchos::REDUCE_MAX,numProcs,&numMyMultColRequests[0],&numGlobalMultColRequests[0]);
// communicate multipleColRequests entries to all processors
int nMyOffset = 0;
for (int i=0; i<myRank-1; i++)
nMyOffset += numGlobalMultColRequests[i]; // calculate offset to store the weights on the corresponding place in procOverlappingWeights
GlobalOrdinal zero=0;
std::vector<GlobalOrdinal> procMultRequestedColIds(globalMultColRequests,zero);
std::vector<GlobalOrdinal> global_procMultRequestedColIds(globalMultColRequests,zero);
// loop over all local column GIDs that are also requested by other procs
for(size_t i = 0; i < multipleColRequests.size(); i++) {
procMultRequestedColIds[nMyOffset + i] = multipleColRequests[i]; // all weights are > 0 ?
}
// template ordinal, package (double)
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, Teuchos::as<int>(globalMultColRequests), &procMultRequestedColIds[0], &global_procMultRequestedColIds[0]);
// loop over global_procOverlappingWeights and eliminate wrong entries...
for (size_t k = 0; k<global_procMultRequestedColIds.size(); k++) {
GlobalOrdinal globColId = global_procMultRequestedColIds[k];
std::vector<Scalar> MyWeightForColId(numProcs,0);
std::vector<Scalar> GlobalWeightForColId(numProcs,0);
if(gColVec->getMap()->isNodeGlobalElement(globColId)) {
MyWeightForColId[myRank] = gColId2Weight[globColId];
} else {
MyWeightForColId[myRank] = 0.0;
}
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &MyWeightForColId[0], &GlobalWeightForColId[0]);
if(gColVec->getMap()->isNodeGlobalElement(globColId)) {
// note: 2 procs could have the same weight for a column index.
// pick the first one.
Scalar winnerValue = 0.0;
int winnerProcRank = 0;
for (int proc = 0; proc < numProcs; proc++) {
if(GlobalWeightForColId[proc] > winnerValue) {
winnerValue = GlobalWeightForColId[proc];
winnerProcRank = proc;
}
}
// winnerProcRank is the winner for handling globColId.
// winnerProcRank is unique (even if two procs have the same weight for a column index)
if(myRank != winnerProcRank) {
// remove corresponding entry from permutedDiagCandidatesFiltered
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator p = permutedDiagCandidatesFiltered.begin();
while(p != permutedDiagCandidatesFiltered.end() )
{
if((*p).second == globColId)
p = permutedDiagCandidatesFiltered.erase(p);
else
p++;
}
}
} // end if isNodeGlobalElement
} // end loop over global_procOverlappingWeights and eliminate wrong entries...
} // end if globalMultColRequests > 0
// put together all pairs:
//size_t sizeRowColPairs = keepDiagonalEntries.size() + permutedDiagCandidatesFiltered.size();
std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> > RowColPairs;
RowColPairs.insert( RowColPairs.end(), keepDiagonalEntries.begin(), keepDiagonalEntries.end());
RowColPairs.insert( RowColPairs.end(), permutedDiagCandidatesFiltered.begin(), permutedDiagCandidatesFiltered.end());
#ifdef DEBUG_OUTPUT
//&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&
// plausibility check
gColVec->putScalar(0.0);
gDomVec->putScalar(0.0);
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator pl = RowColPairs.begin();
while(pl != RowColPairs.end() )
{
//GlobalOrdinal ik = (*pl).first;
GlobalOrdinal jk = (*pl).second;
gColVec->sumIntoGlobalValue(jk,1.0);
pl++;
}
gDomVec->doExport(*gColVec,*exporter,Xpetra::ADD);
for(size_t sz = 0; sz<gDomVec->getLocalLength(); ++sz) {
Teuchos::ArrayRCP< const Scalar > arrDomVec = gDomVec->getData(0);
if(arrDomVec[sz] > 1.0) {
GetOStream(Runtime0,0) << "RowColPairs has multiple column [" << sz << "]=" << arrDomVec[sz] << std::endl;
} else if(arrDomVec[sz] == 0.0) {
GetOStream(Runtime0,0) << "RowColPairs has empty column [" << sz << "]=" << arrDomVec[sz] << std::endl;
}
}
//&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&
#endif
//////////////////////////////////////////////////
// assumption: on each processor RowColPairs now contains
// a valid set of (row,column) pairs, where the row entries
// are a subset of the processor's rows and the column entries
// are unique throughout all processors.
// Note: the RowColPairs are only defined for a subset of all rows,
// so there might be rows without an entry in RowColPairs.
// It can be, that some rows seem to be missing in RowColPairs, since
// the entry in that row with maximum absolute value has been reserved
// by another row already (e.g. as already diagonal dominant row outside
// of perRowMap).
// In fact, the RowColPairs vector only defines the (row,column) pairs
// that will be definitely moved to the diagonal after permutation.
#ifdef DEBUG_OUTPUT
// for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = RowColPairs.begin(); p != RowColPairs.end(); ++p) {
// std::cout << "proc: " << myRank << " r/c: " << (*p).first << "/" << (*p).second << std::endl;
// }
// for (typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::const_iterator p = RowColPairs.begin(); p != RowColPairs.end(); ++p)
// {
//// if((*p).first != (*p).second) std::cout << "difference: " << (*p).first << " " << (*p).second << std::endl;
// std::cout << (*p).first +1 << " " << (*p).second+1 << std::endl;
// }
// std::cout << "\n";
#endif
// vectors to store permutation information
Teuchos::RCP<Vector> Pperm = VectorFactory::Build(A->getRowMap());
Teuchos::RCP<Vector> Qperm = VectorFactory::Build(A->getDomainMap()); // global variant (based on domain map)
Teuchos::RCP<Vector> lQperm = VectorFactory::Build(A->getColMap()); // local variant (based on column map)
Teuchos::ArrayRCP< Scalar > PpermData = Pperm->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > QpermData = Qperm->getDataNonConst(0);
Pperm->putScalar(0.0);
Qperm->putScalar(0.0);
lQperm->putScalar(0.0);
// setup exporter for Qperm
Teuchos::RCP<Export> QpermExporter = ExportFactory::Build(lQperm->getMap(), Qperm->getMap());
Teuchos::RCP<Vector> RowIdStatus = VectorFactory::Build(A->getRowMap());
Teuchos::RCP<Vector> ColIdStatus = VectorFactory::Build(A->getDomainMap()); // global variant (based on domain map)
Teuchos::RCP<Vector> lColIdStatus = VectorFactory::Build(A->getColMap()); // local variant (based on column map)
Teuchos::RCP<Vector> ColIdUsed = VectorFactory::Build(A->getDomainMap()); // mark column ids to be already in use
Teuchos::ArrayRCP< Scalar > RowIdStatusArray = RowIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > ColIdStatusArray = ColIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > lColIdStatusArray = lColIdStatus->getDataNonConst(0);
Teuchos::ArrayRCP< Scalar > ColIdUsedArray = ColIdUsed->getDataNonConst(0); // not sure about this
RowIdStatus->putScalar(0.0);
ColIdStatus->putScalar(0.0);
lColIdStatus->putScalar(0.0);
ColIdUsed->putScalar(0.0); // no column ids are used
// count wide-range permutations
// a wide-range permutation is defined as a permutation of rows/columns which do not
// belong to the same node
LocalOrdinal lWideRangeRowPermutations = 0;
GlobalOrdinal gWideRangeRowPermutations = 0;
LocalOrdinal lWideRangeColPermutations = 0;
GlobalOrdinal gWideRangeColPermutations = 0;
// run 1: mark all "identity" permutations
typename std::vector<std::pair<GlobalOrdinal, GlobalOrdinal> >::iterator p = RowColPairs.begin();
while(p != RowColPairs.end() )
{
GlobalOrdinal ik = (*p).first;
GlobalOrdinal jk = (*p).second;
LocalOrdinal lik = A->getRowMap()->getLocalElement(ik);
LocalOrdinal ljk = A->getColMap()->getLocalElement(jk);
if(RowIdStatusArray[lik] == 0.0) {
RowIdStatusArray[lik] = 1.0; // use this row id
lColIdStatusArray[ljk] = 1.0; // use this column id
Pperm->replaceLocalValue(lik, ik);
lQperm->replaceLocalValue(ljk, ik); // use column map
ColIdUsed->replaceGlobalValue(ik,1.0); // ik is now used
p = RowColPairs.erase(p);
// detect wide range permutations
if(floor(ik/nDofsPerNode) != floor(jk/nDofsPerNode)) {
lWideRangeColPermutations++;
}
}
else
p++;
}
// communicate column map -> domain map
Qperm->doExport(*lQperm,*QpermExporter,Xpetra::ABSMAX);
ColIdStatus->doExport(*lColIdStatus,*QpermExporter,Xpetra::ABSMAX);
// plausibility check
if(RowColPairs.size()>0) GetOStream(Warnings0,0) << "MueLu::PermutationFactory: There are Row/Col pairs left!!!" << std::endl; // TODO fix me
// close Pperm
// count, how many row permutations are missing on current proc
size_t cntFreeRowIdx = 0;
std::queue<GlobalOrdinal> qFreeGRowIdx; // store global row ids of "free" rows
for (size_t lik = 0; lik < RowIdStatus->getLocalLength(); ++lik) {
if(RowIdStatusArray[lik] == 0.0) {
cntFreeRowIdx++;
qFreeGRowIdx.push(RowIdStatus->getMap()->getGlobalElement(lik));
}
}
// fix Pperm
for (size_t lik = 0; lik < RowIdStatus->getLocalLength(); ++lik) {
if(RowIdStatusArray[lik] == 0.0) {
RowIdStatusArray[lik] = 1.0; // use this row id
Pperm->replaceLocalValue(lik, qFreeGRowIdx.front());
// detect wide range permutations
if(floor(qFreeGRowIdx.front()/nDofsPerNode) != floor(RowIdStatus->getMap()->getGlobalElement(lik)/nDofsPerNode)) {
lWideRangeRowPermutations++;
}
qFreeGRowIdx.pop();
}
}
// close Qperm (free permutation entries in Qperm)
size_t cntFreeColIdx = 0;
std::queue<GlobalOrdinal> qFreeGColIdx; // store global column ids of "free" available columns
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
if(ColIdStatusArray[ljk] == 0.0) {
cntFreeColIdx++;
qFreeGColIdx.push(ColIdStatus->getMap()->getGlobalElement(ljk));
}
}
size_t cntUnusedColIdx = 0;
std::queue<GlobalOrdinal> qUnusedGColIdx; // store global column ids of "free" available columns
for (size_t ljk = 0; ljk < ColIdUsed->getLocalLength(); ++ljk) {
if(ColIdUsedArray[ljk] == 0.0) {
cntUnusedColIdx++;
qUnusedGColIdx.push(ColIdUsed->getMap()->getGlobalElement(ljk));
}
}
// fix Qperm with local entries
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
// stop if no (local) unused column idx are left
if(cntUnusedColIdx == 0) break;
if(ColIdStatusArray[ljk] == 0.0) {
ColIdStatusArray[ljk] = 1.0; // use this row id
Qperm->replaceLocalValue(ljk, qUnusedGColIdx.front()); // loop over ColIdStatus (lives on domain map)
ColIdUsed->replaceGlobalValue(qUnusedGColIdx.front(),1.0); // ljk is now used, too
// detect wide range permutations
if(floor(qUnusedGColIdx.front()/nDofsPerNode) != floor(ColIdStatus->getMap()->getGlobalElement(ljk)/nDofsPerNode)) {
lWideRangeColPermutations++;
}
qUnusedGColIdx.pop();
cntUnusedColIdx--;
cntFreeColIdx--;
}
}
//Qperm->doExport(*lQperm,*QpermExporter,Xpetra::ABSMAX); // no export necessary, since changes only locally
//ColIdStatus->doExport(*lColIdStatus,*QpermExporter,Xpetra::ABSMAX);
// count, how many unused column idx are needed on current processor
// to complete Qperm
cntFreeColIdx = 0;
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) { // TODO avoid this loop
if(ColIdStatusArray[ljk] == 0.0) {
cntFreeColIdx++;
}
}
GlobalOrdinal global_cntFreeColIdx = 0;
LocalOrdinal local_cntFreeColIdx = cntFreeColIdx;
sumAll(comm, Teuchos::as<GlobalOrdinal>(local_cntFreeColIdx), global_cntFreeColIdx);
#ifdef DEBUG_OUTPUT
std::cout << "global # of empty column idx entries in Qperm: " << global_cntFreeColIdx << std::endl;
#endif
// avoid global communication if possible
if(global_cntFreeColIdx > 0) {
// 1) count how many unused column ids are left
GlobalOrdinal global_cntUnusedColIdx = 0;
LocalOrdinal local_cntUnusedColIdx = cntUnusedColIdx;
sumAll(comm, Teuchos::as<GlobalOrdinal>(local_cntUnusedColIdx), global_cntUnusedColIdx);
#ifdef DEBUG_OUTPUT
std::cout << "global # of unused column idx: " << global_cntUnusedColIdx << std::endl;
#endif
// 2) communicate how many unused column ids are available on procs
std::vector<LocalOrdinal> local_UnusedColIdxOnProc (numProcs);
std::vector<LocalOrdinal> global_UnusedColIdxOnProc(numProcs);
local_UnusedColIdxOnProc[myRank] = local_cntUnusedColIdx;
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &local_UnusedColIdxOnProc[0], &global_UnusedColIdxOnProc[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global num unused indices per proc: ";
for (size_t ljk = 0; ljk < global_UnusedColIdxOnProc.size(); ++ljk) {
std::cout << " " << global_UnusedColIdxOnProc[ljk];
}
std::cout << std::endl;
#endif
// 3) build array of length global_cntUnusedColIdx to globally replicate unused column idx
std::vector<GlobalOrdinal> local_UnusedColIdxVector(Teuchos::as<size_t>(global_cntUnusedColIdx));
std::vector<GlobalOrdinal> global_UnusedColIdxVector(Teuchos::as<size_t>(global_cntUnusedColIdx));
GlobalOrdinal global_cntUnusedColIdxStartIter = 0;
for(int proc=0; proc<myRank; proc++) {
global_cntUnusedColIdxStartIter += global_UnusedColIdxOnProc[proc];
}
for(GlobalOrdinal k = global_cntUnusedColIdxStartIter; k < global_cntUnusedColIdxStartIter+local_cntUnusedColIdx; k++) {
local_UnusedColIdxVector[k] = qUnusedGColIdx.front();
qUnusedGColIdx.pop();
}
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, Teuchos::as<int>(global_cntUnusedColIdx), &local_UnusedColIdxVector[0], &global_UnusedColIdxVector[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global UnusedGColIdx: ";
for (size_t ljk = 0; ljk < global_UnusedColIdxVector.size(); ++ljk) {
std::cout << " " << global_UnusedColIdxVector[ljk];
}
std::cout << std::endl;
#endif
// 4) communicate, how many column idx are needed on each processor
// to complete Qperm
std::vector<LocalOrdinal> local_EmptyColIdxOnProc (numProcs);
std::vector<LocalOrdinal> global_EmptyColIdxOnProc(numProcs);
local_EmptyColIdxOnProc[myRank] = local_cntFreeColIdx;
Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numProcs, &local_EmptyColIdxOnProc[0], &global_EmptyColIdxOnProc[0]);
#ifdef DEBUG_OUTPUT
std::cout << "PROC " << myRank << " global num of needed column indices: ";
for (size_t ljk = 0; ljk < global_EmptyColIdxOnProc.size(); ++ljk) {
std::cout << " " << global_EmptyColIdxOnProc[ljk];
}
std::cout << std::endl;
#endif
// 5) determine first index in global_UnusedColIdxVector for unused column indices,
// that are marked to be used by this processor
GlobalOrdinal global_UnusedColStartIdx = 0;
for(int proc=0; proc<myRank; proc++) {
global_UnusedColStartIdx += global_EmptyColIdxOnProc[proc];
}
#ifdef DEBUG_OUTPUT
GetOStream(Statistics0,0) << "PROC " << myRank << " is allowd to use the following column gids: ";
for(GlobalOrdinal k = global_UnusedColStartIdx; k < global_UnusedColStartIdx + Teuchos::as<GlobalOrdinal>(cntFreeColIdx); k++) {
GetOStream(Statistics0,0) << global_UnusedColIdxVector[k] << " ";
}
GetOStream(Statistics0,0) << std::endl;
#endif
// 6.) fix Qperm with global entries
GlobalOrdinal array_iter = 0;
for (size_t ljk = 0; ljk < ColIdStatus->getLocalLength(); ++ljk) {
if(ColIdStatusArray[ljk] == 0.0) {
ColIdStatusArray[ljk] = 1.0; // use this row id
Qperm->replaceLocalValue(ljk, global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter]);
ColIdUsed->replaceGlobalValue(global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter],1.0);
// detect wide range permutations
if(floor(global_UnusedColIdxVector[global_UnusedColStartIdx + array_iter]/nDofsPerNode) != floor(ColIdStatus->getMap()->getGlobalElement(ljk)/nDofsPerNode)) {
lWideRangeColPermutations++;
}
array_iter++;
//cntUnusedColIdx--; // check me
}
}
} // end if global_cntFreeColIdx > 0
/////////////////// Qperm should be fine now...
// create new empty Matrix
Teuchos::RCP<CrsMatrixWrap> permPTmatrix = Teuchos::rcp(new CrsMatrixWrap(A->getRowMap(),1,Xpetra::StaticProfile));
Teuchos::RCP<CrsMatrixWrap> permQTmatrix = Teuchos::rcp(new CrsMatrixWrap(A->getRowMap(),1,Xpetra::StaticProfile));
for(size_t row=0; row<A->getNodeNumRows(); row++) {
Teuchos::ArrayRCP<GlobalOrdinal> indoutP(1,Teuchos::as<GO>(PpermData[row])); // column idx for Perm^T
Teuchos::ArrayRCP<GlobalOrdinal> indoutQ(1,Teuchos::as<GO>(QpermData[row])); // column idx for Qperm
Teuchos::ArrayRCP<Scalar> valout(1,1.0);
permPTmatrix->insertGlobalValues(A->getRowMap()->getGlobalElement(row), indoutP.view(0,indoutP.size()), valout.view(0,valout.size()));
permQTmatrix->insertGlobalValues (A->getRowMap()->getGlobalElement(row), indoutQ.view(0,indoutQ.size()), valout.view(0,valout.size()));
}
permPTmatrix->fillComplete();
permQTmatrix->fillComplete();
Teuchos::RCP<Matrix> permPmatrix = Utils2::Transpose(permPTmatrix,true);
for(size_t row=0; row<permPTmatrix->getNodeNumRows(); row++) {
if(permPTmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permPTmatrix is " << permPTmatrix->getNumEntriesInLocalRow(row) << std::endl;
if(permPmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permPmatrix is " << permPmatrix->getNumEntriesInLocalRow(row) << std::endl;
if(permQTmatrix->getNumEntriesInLocalRow(row) != 1)
GetOStream(Warnings0,0) <<"#entries in row " << row << " of permQmatrix is " << permQTmatrix->getNumEntriesInLocalRow(row) << std::endl;
}
// build permP * A * permQT
Teuchos::RCP<Matrix> ApermQt = Utils::Multiply(*A, false, *permQTmatrix, false);
Teuchos::RCP<Matrix> permPApermQt = Utils::Multiply(*permPmatrix, false, *ApermQt, false);
/*
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("A.mat", *A);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permP.mat", *permPmatrix);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permQt.mat", *permQTmatrix);
MueLu::Utils<Scalar, LocalOrdinal, GlobalOrdinal, Node, LocalMatOps>::Write("permPApermQt.mat", *permPApermQt);
*/
// build scaling matrix
Teuchos::RCP<Vector> diagVec = VectorFactory::Build(permPApermQt->getRowMap(),true);
Teuchos::RCP<Vector> invDiagVec = VectorFactory::Build(permPApermQt->getRowMap(),true);
Teuchos::ArrayRCP< const Scalar > diagVecData = diagVec->getData(0);
Teuchos::ArrayRCP< Scalar > invDiagVecData = invDiagVec->getDataNonConst(0);
permPApermQt->getLocalDiagCopy(*diagVec);
for(size_t i = 0; i<diagVec->getMap()->getNodeNumElements(); ++i) {
if(diagVecData[i] != 0.0)
invDiagVecData[i] = 1/diagVecData[i];
else {
invDiagVecData[i] = 1.0;
GetOStream(Statistics0,0) << "MueLu::PermutationFactory: found zero on diagonal in row " << i << std::endl;
}
}
Teuchos::RCP<CrsMatrixWrap> diagScalingOp = Teuchos::rcp(new CrsMatrixWrap(permPApermQt->getRowMap(),1,Xpetra::StaticProfile));
for(size_t row=0; row<A->getNodeNumRows(); row++) {
Teuchos::ArrayRCP<GlobalOrdinal> indout(1,permPApermQt->getRowMap()->getGlobalElement(row)); // column idx for Perm^T
Teuchos::ArrayRCP<Scalar> valout(1,invDiagVecData[row]);
diagScalingOp->insertGlobalValues(A->getRowMap()->getGlobalElement(row), indout.view(0,indout.size()), valout.view(0,valout.size()));
}
diagScalingOp->fillComplete();
Teuchos::RCP<Matrix> scaledA = Utils::Multiply(*diagScalingOp, false, *permPApermQt, false);
currentLevel.Set("A", Teuchos::rcp_dynamic_cast<Matrix>(scaledA), genFactory/*this*/);
currentLevel.Set("permA", Teuchos::rcp_dynamic_cast<Matrix>(permPApermQt), genFactory/*this*/); // TODO careful with this!!!
currentLevel.Set("permP", Teuchos::rcp_dynamic_cast<Matrix>(permPmatrix), genFactory/*this*/);
currentLevel.Set("permQT", Teuchos::rcp_dynamic_cast<Matrix>(permQTmatrix), genFactory/*this*/);
currentLevel.Set("permScaling", Teuchos::rcp_dynamic_cast<Matrix>(diagScalingOp), genFactory/*this*/);
//// count row permutations
// count zeros on diagonal in P -> number of row permutations
Teuchos::RCP<Vector> diagPVec = VectorFactory::Build(permPmatrix->getRowMap(),true);
permPmatrix->getLocalDiagCopy(*diagPVec);
Teuchos::ArrayRCP< const Scalar > diagPVecData = diagPVec->getData(0);
LocalOrdinal lNumRowPermutations = 0;
GlobalOrdinal gNumRowPermutations = 0;
for(size_t i = 0; i<diagPVec->getMap()->getNodeNumElements(); ++i) {
if(diagPVecData[i] == 0.0) {
lNumRowPermutations++;
}
}
// sum up all entries in multipleColRequests over all processors
sumAll(diagPVec->getMap()->getComm(), Teuchos::as<GlobalOrdinal>(lNumRowPermutations), gNumRowPermutations);
//// count column permutations
// count zeros on diagonal in Q^T -> number of column permutations
Teuchos::RCP<Vector> diagQTVec = VectorFactory::Build(permQTmatrix->getRowMap(),true);
permQTmatrix->getLocalDiagCopy(*diagQTVec);
Teuchos::ArrayRCP< const Scalar > diagQTVecData = diagQTVec->getData(0);
LocalOrdinal lNumColPermutations = 0;
GlobalOrdinal gNumColPermutations = 0;
for(size_t i = 0; i<diagQTVec->getMap()->getNodeNumElements(); ++i) {
if(diagQTVecData[i] == 0.0) {
lNumColPermutations++;
}
}
// sum up all entries in multipleColRequests over all processors
sumAll(diagQTVec->getMap()->getComm(), Teuchos::as<GlobalOrdinal>(lNumColPermutations), gNumColPermutations);
currentLevel.Set("#RowPermutations", gNumRowPermutations, genFactory/*this*/);
currentLevel.Set("#ColPermutations", gNumColPermutations, genFactory/*this*/);
currentLevel.Set("#WideRangeRowPermutations", gWideRangeRowPermutations, genFactory/*this*/);
currentLevel.Set("#WideRangeColPermutations", gWideRangeColPermutations, genFactory/*this*/);
GetOStream(Statistics0, 0) << "#Row permutations/max possible permutations: " << gNumRowPermutations << "/" << diagPVec->getMap()->getGlobalNumElements() << std::endl;
GetOStream(Statistics0, 0) << "#Column permutations/max possible permutations: " << gNumColPermutations << "/" << diagQTVec->getMap()->getGlobalNumElements() << std::endl;
GetOStream(Runtime1, 0) << "#wide range row permutations: " << gWideRangeRowPermutations << " #wide range column permutations: " << gWideRangeColPermutations << std::endl;
#else
#warning PermutationFactory not compiling/working for Scalar==complex.
#endif // #ifndef HAVE_MUELU_INST_COMPLEX_INT_INT
}
void RAPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::CheckRepairMainDiagonal(RCP<Matrix>& Ac) const {
const Teuchos::ParameterList& pL = GetParameterList();
bool repairZeroDiagonals = pL.get<bool>("RepairMainDiagonal");
bool checkAc = pL.get<bool>("CheckMainDiagonal");
if (!checkAc && !repairZeroDiagonals)
return;
SC zero = Teuchos::ScalarTraits<SC>::zero(), one = Teuchos::ScalarTraits<SC>::one();
Teuchos::RCP<Teuchos::ParameterList> p = Teuchos::rcp(new Teuchos::ParameterList());
p->set("DoOptimizeStorage", true);
RCP<const Map> rowMap = Ac->getRowMap();
RCP<Vector> diagVec = VectorFactory::Build(rowMap);
Ac->getLocalDiagCopy(*diagVec);
LO lZeroDiags = 0;
Teuchos::ArrayRCP< Scalar > diagVal = diagVec->getDataNonConst(0);
for (size_t i = 0; i < rowMap->getNodeNumElements(); i++) {
if (diagVal[i] == zero) {
lZeroDiags++;
}
}
GO gZeroDiags;
MueLu_sumAll(rowMap->getComm(), Teuchos::as<GO>(lZeroDiags), gZeroDiags);
if (repairZeroDiagonals && gZeroDiags > 0) {
// TAW: If Ac has empty rows, put a 1 on the diagonal of Ac. Be aware that Ac might have empty rows AND columns.
// The columns might not exist in the column map at all.
//
// It would be nice to add the entries to the original matrix Ac. But then we would have to use
// insertLocalValues. However we cannot add new entries for local column indices that do not exist in the column map
// of Ac (at least Epetra is not able to do this).
//
// Here we build a diagonal matrix with zeros on the diagonal and ones on the diagonal for the rows where Ac has empty rows
// We have to build a new matrix to be able to use insertGlobalValues. Then we add the original matrix Ac to our new block
// diagonal matrix and use the result as new (non-singular) matrix Ac.
// This is very inefficient.
//
// If you know something better, please let me know.
RCP<Matrix> fixDiagMatrix = Teuchos::null;
fixDiagMatrix = MatrixFactory::Build(rowMap, 1);
for (size_t r = 0; r < rowMap->getNodeNumElements(); r++) {
if (diagVal[r] == zero) {
GO grid = rowMap->getGlobalElement(r);
Teuchos::ArrayRCP<GO> indout(1,grid);
Teuchos::ArrayRCP<SC> valout(1, one);
fixDiagMatrix->insertGlobalValues(grid,indout.view(0, 1), valout.view(0, 1));
}
}
{
Teuchos::TimeMonitor m1(*Teuchos::TimeMonitor::getNewTimer("CheckRepairMainDiagonal: fillComplete1"));
Ac->fillComplete(p);
}
MueLu::Utils2<Scalar, LocalOrdinal, GlobalOrdinal, Node>::TwoMatrixAdd(*Ac, false, 1.0, *fixDiagMatrix, 1.0);
if (Ac->IsView("stridedMaps"))
fixDiagMatrix->CreateView("stridedMaps", Ac);
Ac = Teuchos::null; // free singular coarse level matrix
Ac = fixDiagMatrix; // set fixed non-singular coarse level matrix
}
// call fillComplete with optimized storage option set to true
// This is necessary for new faster Epetra MM kernels.
{
Teuchos::TimeMonitor m1(*Teuchos::TimeMonitor::getNewTimer("CheckRepairMainDiagonal: fillComplete2"));
Ac->fillComplete(p);
}
// print some output
if (IsPrint(Warnings0))
GetOStream(Warnings0) << "RAPFactory (WARNING): " << (repairZeroDiagonals ? "repaired " : "found ")
<< gZeroDiags << " zeros on main diagonal of Ac." << std::endl;
#ifdef HAVE_MUELU_DEBUG // only for debugging
// check whether Ac has been repaired...
Ac->getLocalDiagCopy(*diagVec);
Teuchos::ArrayRCP< Scalar > diagVal2 = diagVec->getDataNonConst(0);
for (size_t r = 0; r < Ac->getRowMap()->getNodeNumElements(); r++) {
if (diagVal2[r] == zero) {
GetOStream(Errors,-1) << "Error: there are zeros left on diagonal after repair..." << std::endl;
break;
}
}
#endif
}
