void panzer::ScatterResidual_Epetra<panzer::Traits::Hessian, TRAITS,LO,GO>::
evaluateFields(typename TRAITS::EvalData workset)
{
std::vector<int> cLIDs, rLIDs;
std::vector<double> jacRow;
bool useColumnIndexer = colGlobalIndexer_!=Teuchos::null;
// for convenience pull out some objects from workset
std::string blockId = this->wda(workset).block_id;
const std::vector<std::size_t> & localCellIds = this->wda(workset).cell_local_ids;
Teuchos::RCP<Epetra_Vector> r = epetraContainer_->get_f();
Teuchos::RCP<Epetra_CrsMatrix> Jac = epetraContainer_->get_A();
const Teuchos::RCP<const panzer::UniqueGlobalIndexer<LO,GO> >&
colGlobalIndexer = useColumnIndexer ? colGlobalIndexer_ : globalIndexer_;
// NOTE: A reordering of these loops will likely improve performance
// The "getGIDFieldOffsets" may be expensive. However the
// "getElementGIDs" can be cheaper. However the lookup for LIDs
// may be more expensive!
// scatter operation for each cell in workset
for(std::size_t worksetCellIndex=0;worksetCellIndex<localCellIds.size();++worksetCellIndex) {
std::size_t cellLocalId = localCellIds[worksetCellIndex];
rLIDs = globalIndexer_->getElementLIDs(cellLocalId);
cLIDs = colGlobalIndexer->getElementLIDs(cellLocalId);
if (Teuchos::nonnull(workset.other)) {
const std::size_t other_cellLocalId = workset.other->cell_local_ids[worksetCellIndex];
const std::vector<int> other_cLIDs = colGlobalIndexer->getElementLIDs(other_cellLocalId);
cLIDs.insert(cLIDs.end(), other_cLIDs.begin(), other_cLIDs.end());
}
// loop over each field to be scattered
for(std::size_t fieldIndex = 0; fieldIndex < scatterFields_.size(); fieldIndex++) {
int fieldNum = fieldIds_[fieldIndex];
const std::vector<int> & elmtOffset = globalIndexer_->getGIDFieldOffsets(blockId,fieldNum);
// loop over the basis functions (currently they are nodes)
for(std::size_t rowBasisNum = 0; rowBasisNum < elmtOffset.size(); rowBasisNum++) {
const ScalarT scatterField = (scatterFields_[fieldIndex])(worksetCellIndex,rowBasisNum);
int rowOffset = elmtOffset[rowBasisNum];
int row = rLIDs[rowOffset];
// loop over the sensitivity indices: all DOFs on a cell
jacRow.resize(scatterField.size());
for(int sensIndex=0;sensIndex<scatterField.size();++sensIndex)
jacRow[sensIndex] = scatterField.fastAccessDx(sensIndex).fastAccessDx(0);
{
int err = Jac->SumIntoMyValues(
row,
std::min(cLIDs.size(), static_cast<size_t>(scatterField.size())),
panzer::ptrFromStlVector(jacRow),
panzer::ptrFromStlVector(cLIDs));
TEUCHOS_ASSERT_EQUALITY(err,0);
}
} // end rowBasisNum
} // end fieldIndex
}
}
void panzer::ScatterDirichletResidual_Tpetra<panzer::Traits::Jacobian, TRAITS,LO,GO,NodeT>::
evaluateFields(typename TRAITS::EvalData workset)
{
std::vector<GO> GIDs;
// for convenience pull out some objects from workset
std::string blockId = workset.block_id;
const std::vector<std::size_t> & localCellIds = workset.cell_local_ids;
Teuchos::RCP<typename LOC::VectorType> r = tpetraContainer_->get_f();
Teuchos::RCP<typename LOC::CrsMatrixType> Jac = tpetraContainer_->get_A();
Teuchos::ArrayRCP<double> r_array = r->get1dViewNonConst();
Teuchos::ArrayRCP<double> dc_array = dirichletCounter_->get1dViewNonConst();
// NOTE: A reordering of these loops will likely improve performance
// The "getGIDFieldOffsets may be expensive. However the
// "getElementGIDs" can be cheaper. However the lookup for LIDs
// may be more expensive!
// scatter operation for each cell in workset
for(std::size_t worksetCellIndex=0;worksetCellIndex<localCellIds.size();++worksetCellIndex) {
std::size_t cellLocalId = localCellIds[worksetCellIndex];
globalIndexer_->getElementGIDs(cellLocalId,GIDs);
const std::vector<LO> & LIDs = globalIndexer_->getElementLIDs(cellLocalId);
// loop over each field to be scattered
for(std::size_t fieldIndex = 0; fieldIndex < scatterFields_.size(); fieldIndex++) {
int fieldNum = fieldIds_[fieldIndex];
// this call "should" get the right ordering according to the Intrepid basis
const std::pair<std::vector<int>,std::vector<int> > & indicePair
= globalIndexer_->getGIDFieldOffsets_closure(blockId,fieldNum, side_subcell_dim_, local_side_id_);
const std::vector<int> & elmtOffset = indicePair.first;
const std::vector<int> & basisIdMap = indicePair.second;
// loop over basis functions
for(std::size_t basis=0;basis<elmtOffset.size();basis++) {
int offset = elmtOffset[basis];
LO lid = LIDs[offset];
if(lid<0) // not on this processor
continue;
int basisId = basisIdMap[basis];
if (checkApplyBC_)
if (!applyBC_[fieldIndex](worksetCellIndex,basisId))
continue;
// zero out matrix row
{
std::size_t sz = Jac->getNumEntriesInLocalRow(lid);
std::size_t numEntries = 0;
Teuchos::Array<LO> rowIndices(sz);
Teuchos::Array<double> rowValues(sz);
// Jac->getLocalRowView(lid,numEntries,rowValues,rowIndices);
Jac->getLocalRowCopy(lid,rowIndices,rowValues,numEntries);
for(std::size_t i=0;i<numEntries;i++)
rowValues[i] = 0.0;
Jac->replaceLocalValues(lid,rowIndices,rowValues);
}
GO gid = GIDs[offset];
const ScalarT scatterField = (scatterFields_[fieldIndex])(worksetCellIndex,basisId);
r_array[lid] = scatterField.val();
dc_array[lid] = 1.0; // mark row as dirichlet
// loop over the sensitivity indices: all DOFs on a cell
std::vector<double> jacRow(scatterField.size(),0.0);
for(int sensIndex=0;sensIndex<scatterField.size();++sensIndex)
jacRow[sensIndex] = scatterField.fastAccessDx(sensIndex);
TEUCHOS_ASSERT(jacRow.size()==GIDs.size());
Jac->replaceGlobalValues(gid, GIDs, jacRow);
}
}
}
}
void panzer::ScatterResidual_BlockedEpetra<panzer::Traits::Jacobian, TRAITS,LO,GO>::
evaluateFields(typename TRAITS::EvalData workset)
{
using Teuchos::RCP;
using Teuchos::ArrayRCP;
using Teuchos::ptrFromRef;
using Teuchos::rcp_dynamic_cast;
using Thyra::VectorBase;
using Thyra::SpmdVectorBase;
using Thyra::ProductVectorBase;
using Thyra::BlockedLinearOpBase;
typedef BlockedEpetraLinearObjContainer BLOC;
std::vector<std::pair<int,GO> > GIDs;
std::vector<LO> LIDs;
std::vector<double> jacRow;
// for convenience pull out some objects from workset
std::string blockId = this->wda(workset).block_id;
const std::vector<std::size_t> & localCellIds = this->wda(workset).cell_local_ids;
RCP<const BLOC> blockedContainer = blockedContainer_;
RCP<ProductVectorBase<double> > r = rcp_dynamic_cast<ProductVectorBase<double> >(blockedContainer->get_f());
Teuchos::RCP<BlockedLinearOpBase<double> > Jac = rcp_dynamic_cast<BlockedLinearOpBase<double> >(blockedContainer->get_A());
int numFieldBlocks = globalIndexer_->getNumFieldBlocks();
std::vector<int> blockOffsets(numFieldBlocks+1); // number of fields, plus a sentinnel
for(int blk=0;blk<numFieldBlocks;blk++) {
int blockOffset = globalIndexer_->getBlockGIDOffset(blockId,blk);
blockOffsets[blk] = blockOffset;
}
std::unordered_map<std::pair<int,int>,Teuchos::RCP<Epetra_CrsMatrix>,panzer::pair_hash> jacEpetraBlocks;
// NOTE: A reordering of these loops will likely improve performance
// The "getGIDFieldOffsets" may be expensive. However the
// "getElementGIDs" can be cheaper. However the lookup for LIDs
// may be more expensive!
// scatter operation for each cell in workset
for(std::size_t worksetCellIndex=0;worksetCellIndex<localCellIds.size();++worksetCellIndex) {
std::size_t cellLocalId = localCellIds[worksetCellIndex];
globalIndexer_->getElementGIDs(cellLocalId,GIDs,blockId);
// caculate the local IDs for this element
LIDs.resize(GIDs.size());
for(std::size_t i=0;i<GIDs.size();i++) {
// used for doing local ID lookups
RCP<const Epetra_Map> r_map = blockedContainer->getMapForBlock(GIDs[i].first);
LIDs[i] = r_map->LID(GIDs[i].second);
}
// loop over each field to be scattered
Teuchos::ArrayRCP<double> local_r;
for(std::size_t fieldIndex = 0; fieldIndex < scatterFields_.size(); fieldIndex++) {
int fieldNum = fieldIds_[fieldIndex];
int blockRowIndex = globalIndexer_->getFieldBlock(fieldNum);
// grab local data for inputing
if(r!=Teuchos::null) {
RCP<SpmdVectorBase<double> > block_r = rcp_dynamic_cast<SpmdVectorBase<double> >(r->getNonconstVectorBlock(blockRowIndex));
block_r->getNonconstLocalData(ptrFromRef(local_r));
}
const std::vector<int> & elmtOffset = globalIndexer_->getGIDFieldOffsets(blockId,fieldNum);
// loop over the basis functions (currently they are nodes)
for(std::size_t rowBasisNum = 0; rowBasisNum < elmtOffset.size(); rowBasisNum++) {
const ScalarT scatterField = (scatterFields_[fieldIndex])(worksetCellIndex,rowBasisNum);
int rowOffset = elmtOffset[rowBasisNum];
int r_lid = LIDs[rowOffset];
// Sum residual
if(local_r!=Teuchos::null)
local_r[r_lid] += (scatterField.val());
blockOffsets[numFieldBlocks] = scatterField.size(); // add the sentinel
// loop over the sensitivity indices: all DOFs on a cell
jacRow.resize(scatterField.size());
// For Neumann conditions with no dependence on degrees of freedom, there should be no Jacobian contribution
if(scatterField.size() == 0)
continue;
for(int sensIndex=0;sensIndex<scatterField.size();++sensIndex) {
jacRow[sensIndex] = scatterField.fastAccessDx(sensIndex);
}
for(int blockColIndex=0;blockColIndex<numFieldBlocks;blockColIndex++) {
int start = blockOffsets[blockColIndex];
int end = blockOffsets[blockColIndex+1];
if(end-start<=0)
continue;
// check hash table for jacobian sub block
std::pair<int,int> blockIndex = std::make_pair(blockRowIndex,blockColIndex);
Teuchos::RCP<Epetra_CrsMatrix> subJac = jacEpetraBlocks[blockIndex];
// if you didn't find one before, add it to the hash table
if(subJac==Teuchos::null) {
Teuchos::RCP<Thyra::LinearOpBase<double> > tOp = Jac->getNonconstBlock(blockIndex.first,blockIndex.second);
// block operator is null, don't do anything (it is excluded)
if(Teuchos::is_null(tOp))
continue;
Teuchos::RCP<Epetra_Operator> eOp = Thyra::get_Epetra_Operator(*tOp);
subJac = rcp_dynamic_cast<Epetra_CrsMatrix>(eOp,true);
jacEpetraBlocks[blockIndex] = subJac;
}
// Sum Jacobian
int err = subJac->SumIntoMyValues(r_lid, end-start, &jacRow[start],&LIDs[start]);
if(err!=0) {
RCP<const Epetra_Map> rr = blockedContainer->getMapForBlock(GIDs[start].first);
bool sameColMap = subJac->ColMap().SameAs(*rr);
std::stringstream ss;
ss << "Failed inserting row: " << GIDs[rowOffset].second << " (" << r_lid << "): ";
for(int i=start;i<end;i++)
ss << GIDs[i].second << " (" << LIDs[i] << ") ";
ss << std::endl;
ss << "Into block " << blockRowIndex << ", " << blockColIndex << std::endl;
ss << "scatter field = ";
scatterFields_[fieldIndex].print(ss);
ss << std::endl;
ss << "Same map = " << (sameColMap ? "true" : "false") << std::endl;
TEUCHOS_TEST_FOR_EXCEPTION(err!=0,std::runtime_error,ss.str());
}
}
} // end rowBasisNum
} // end fieldIndex
}
}
void panzer::ScatterDirichletResidual_BlockedTpetra<panzer::Traits::Jacobian, TRAITS,LO,GO,NodeT>::
evaluateFields(typename TRAITS::EvalData workset)
{
using Teuchos::RCP;
using Teuchos::ArrayRCP;
using Teuchos::ptrFromRef;
using Teuchos::rcp_dynamic_cast;
using Thyra::VectorBase;
using Thyra::SpmdVectorBase;
using Thyra::ProductVectorBase;
using Thyra::BlockedLinearOpBase;
std::vector<std::pair<int,GO> > GIDs;
std::vector<LO> LIDs;
// for convenience pull out some objects from workset
std::string blockId = this->wda(workset).block_id;
const std::vector<std::size_t> & localCellIds = this->wda(workset).cell_local_ids;
RCP<ProductVectorBase<double> > r = rcp_dynamic_cast<ProductVectorBase<double> >(blockedContainer_->get_f());
Teuchos::RCP<BlockedLinearOpBase<double> > Jac = rcp_dynamic_cast<BlockedLinearOpBase<double> >(blockedContainer_->get_A());
int numFieldBlocks = globalIndexer_->getNumFieldBlocks();
std::vector<int> blockOffsets(numFieldBlocks+1); // number of fields, plus a sentinnel
for(int blk=0; blk<numFieldBlocks; blk++) {
int blockOffset = globalIndexer_->getBlockGIDOffset(blockId,blk);
blockOffsets[blk] = blockOffset;
}
std::unordered_map<std::pair<int,int>,Teuchos::RCP<CrsMatrixType>,panzer::pair_hash> jacTpetraBlocks;
// NOTE: A reordering of these loops will likely improve performance
// The "getGIDFieldOffsets may be expensive. However the
// "getElementGIDs" can be cheaper. However the lookup for LIDs
// may be more expensive!
// scatter operation for each cell in workset
for(std::size_t worksetCellIndex=0; worksetCellIndex<localCellIds.size(); ++worksetCellIndex) {
std::size_t cellLocalId = localCellIds[worksetCellIndex];
globalIndexer_->getElementGIDs(cellLocalId,GIDs);
blockOffsets[numFieldBlocks] = GIDs.size();
// caculate the local IDs for this element
LIDs.resize(GIDs.size());
for(std::size_t i=0; i<GIDs.size(); i++) {
// used for doing local ID lookups
RCP<const MapType> r_map = blockedContainer_->getMapForBlock(GIDs[i].first);
LIDs[i] = r_map->getLocalElement(GIDs[i].second);
}
// loop over each field to be scattered
Teuchos::ArrayRCP<double> local_r, local_dc;
for(std::size_t fieldIndex = 0; fieldIndex < scatterFields_.size(); fieldIndex++) {
int fieldNum = fieldIds_[fieldIndex];
int blockRowIndex = globalIndexer_->getFieldBlock(fieldNum);
RCP<SpmdVectorBase<double> > dc = rcp_dynamic_cast<SpmdVectorBase<double> >(dirichletCounter_->getNonconstVectorBlock(blockRowIndex));
dc->getNonconstLocalData(ptrFromRef(local_dc));
// grab local data for inputing
RCP<SpmdVectorBase<double> > block_r = rcp_dynamic_cast<SpmdVectorBase<double> >(r->getNonconstVectorBlock(blockRowIndex));
block_r->getNonconstLocalData(ptrFromRef(local_r));
// this call "should" get the right ordering according to the Intrepid basis
const std::pair<std::vector<int>,std::vector<int> > & indicePair
= globalIndexer_->getGIDFieldOffsets_closure(blockId,fieldNum, side_subcell_dim_, local_side_id_);
const std::vector<int> & elmtOffset = indicePair.first;
const std::vector<int> & basisIdMap = indicePair.second;
// loop over basis functions
for(std::size_t basis=0; basis<elmtOffset.size(); basis++) {
int offset = elmtOffset[basis];
int lid = LIDs[offset];
if(lid<0) // not on this processor
continue;
int basisId = basisIdMap[basis];
if (checkApplyBC_)
if (!applyBC_[fieldIndex](worksetCellIndex,basisId))
continue;
// zero out matrix row
for(int blockColIndex=0; blockColIndex<numFieldBlocks; blockColIndex++) {
int start = blockOffsets[blockColIndex];
int end = blockOffsets[blockColIndex+1];
if(end-start<=0)
continue;
// check hash table for jacobian sub block
std::pair<int,int> blockIndex = std::make_pair(blockRowIndex,blockColIndex);
Teuchos::RCP<CrsMatrixType> subJac = jacTpetraBlocks[blockIndex];
// if you didn't find one before, add it to the hash table
if(subJac==Teuchos::null) {
Teuchos::RCP<Thyra::LinearOpBase<double> > tOp = Jac->getNonconstBlock(blockIndex.first,blockIndex.second);
// block operator is null, don't do anything (it is excluded)
if(Teuchos::is_null(tOp))
continue;
Teuchos::RCP<OperatorType> tpetra_Op = rcp_dynamic_cast<ThyraLinearOp>(tOp)->getTpetraOperator();
subJac = rcp_dynamic_cast<CrsMatrixType>(tpetra_Op,true);
jacTpetraBlocks[blockIndex] = subJac;
}
std::size_t sz = subJac->getNumEntriesInLocalRow(lid);
std::size_t numEntries = 0;
Teuchos::Array<LO> rowIndices(sz);
Teuchos::Array<double> rowValues(sz);
subJac->getLocalRowCopy(lid,rowIndices,rowValues,numEntries);
for(std::size_t i=0; i<numEntries; i++)
rowValues[i] = 0.0;
subJac->replaceLocalValues(lid,rowIndices,rowValues);
}
const ScalarT scatterField = (scatterFields_[fieldIndex])(worksetCellIndex,basisId);
local_r[lid] = scatterField.val();
local_dc[lid] = 1.0; // mark row as dirichlet
// loop over the sensitivity indices: all DOFs on a cell
std::vector<double> jacRow(scatterField.size(),0.0);
for(int sensIndex=0; sensIndex<scatterField.size(); ++sensIndex)
jacRow[sensIndex] = scatterField.fastAccessDx(sensIndex);
TEUCHOS_ASSERT(jacRow.size()==GIDs.size());
for(int blockColIndex=0; blockColIndex<numFieldBlocks; blockColIndex++) {
int start = blockOffsets[blockColIndex];
int end = blockOffsets[blockColIndex+1];
if(end-start<=0)
continue;
// check hash table for jacobian sub block
std::pair<int,int> blockIndex = std::make_pair(blockRowIndex,blockColIndex);
Teuchos::RCP<CrsMatrixType> subJac = jacTpetraBlocks[blockIndex];
// if you didn't find one before, add it to the hash table
if(subJac==Teuchos::null) {
Teuchos::RCP<Thyra::LinearOpBase<double> > tOp = Jac->getNonconstBlock(blockIndex.first,blockIndex.second);
// block operator is null, don't do anything (it is excluded)
if(Teuchos::is_null(tOp))
continue;
Teuchos::RCP<OperatorType> tpetra_Op = rcp_dynamic_cast<ThyraLinearOp>(tOp)->getTpetraOperator();
subJac = rcp_dynamic_cast<CrsMatrixType>(tpetra_Op,true);
jacTpetraBlocks[blockIndex] = subJac;
}
// Sum Jacobian
subJac->replaceLocalValues(lid, Teuchos::arrayViewFromVector(LIDs).view(start,end-start),
Teuchos::arrayViewFromVector(jacRow).view(start,end-start));
}
}
}
}
}
void panzer::ScatterDirichletResidual_Tpetra<panzer::Traits::Tangent, TRAITS,LO,GO,NodeT>::
evaluateFields(typename TRAITS::EvalData workset)
{
std::vector<GO> GIDs;
std::vector<LO> LIDs;
// for convenience pull out some objects from workset
std::string blockId = this->wda(workset).block_id;
const std::vector<std::size_t> & localCellIds = this->wda(workset).cell_local_ids;
Teuchos::RCP<typename LOC::VectorType> r = (!scatterIC_) ?
tpetraContainer_->get_f() :
tpetraContainer_->get_x();
Teuchos::ArrayRCP<double> r_array = r->get1dViewNonConst();
Teuchos::ArrayRCP<double> dc_array = dirichletCounter_->get1dViewNonConst();
// NOTE: A reordering of these loops will likely improve performance
// The "getGIDFieldOffsets may be expensive. However the
// "getElementGIDs" can be cheaper. However the lookup for LIDs
// may be more expensive!
// scatter operation for each cell in workset
for(std::size_t worksetCellIndex=0;worksetCellIndex<localCellIds.size();++worksetCellIndex) {
std::size_t cellLocalId = localCellIds[worksetCellIndex];
globalIndexer_->getElementGIDs(cellLocalId,GIDs);
// caculate the local IDs for this element
LIDs.resize(GIDs.size());
for(std::size_t i=0;i<GIDs.size();i++)
LIDs[i] = r->getMap()->getLocalElement(GIDs[i]);
// loop over each field to be scattered
for(std::size_t fieldIndex = 0; fieldIndex < scatterFields_.size(); fieldIndex++) {
int fieldNum = fieldIds_[fieldIndex];
if (!scatterIC_) {
// this call "should" get the right ordering according to the Intrepid2 basis
const std::pair<std::vector<int>,std::vector<int> > & indicePair
= globalIndexer_->getGIDFieldOffsets_closure(blockId,fieldNum, side_subcell_dim_, local_side_id_);
const std::vector<int> & elmtOffset = indicePair.first;
const std::vector<int> & basisIdMap = indicePair.second;
// loop over basis functions
for(std::size_t basis=0;basis<elmtOffset.size();basis++) {
int offset = elmtOffset[basis];
LO lid = LIDs[offset];
if(lid<0) // not on this processor!
continue;
int basisId = basisIdMap[basis];
if (checkApplyBC_)
if (!applyBC_[fieldIndex](worksetCellIndex,basisId))
continue;
ScalarT value = (scatterFields_[fieldIndex])(worksetCellIndex,basisId);
//r_array[lid] = (scatterFields_[fieldIndex])(worksetCellIndex,basisId).val();
// then scatter the sensitivity vectors
if(value.size()==0)
for(std::size_t d=0;d<dfdp_vectors_.size();d++)
dfdp_vectors_[d][lid] = 0.0;
else
for(int d=0;d<value.size();d++) {
dfdp_vectors_[d][lid] = value.fastAccessDx(d);
}
// record that you set a dirichlet condition
dc_array[lid] = 1.0;
}
} else {
// this call "should" get the right ordering according to the Intrepid2 basis
const std::vector<int> & elmtOffset = globalIndexer_->getGIDFieldOffsets(blockId,fieldNum);
// loop over basis functions
for(std::size_t basis=0;basis<elmtOffset.size();basis++) {
int offset = elmtOffset[basis];
LO lid = LIDs[offset];
if(lid<0) // not on this processor!
continue;
ScalarT value = (scatterFields_[fieldIndex])(worksetCellIndex,basis);
//r_array[lid] = (scatterFields_[fieldIndex])(worksetCellIndex,basis).val();
// then scatter the sensitivity vectors
if(value.size()==0)
for(std::size_t d=0;d<dfdp_vectors_.size();d++)
dfdp_vectors_[d][lid] = 0.0;
else
for(int d=0;d<value.size();d++) {
dfdp_vectors_[d][lid] = value.fastAccessDx(d);
}
// record that you set a dirichlet condition
dc_array[lid] = 1.0;
}
}
}
}
}
