static Epetra_MultiVector getComponent( spaceT const& Xh, Epetra_MultiVector const& sol )
{
Epetra_Map componentMap ( epetraMap( Xh->template functionSpace<index>()->map() ) );
Epetra_Map globalMap ( epetraMap( Xh->map() ) );
//DVLOG(2) << "Component map: " << componentMap << "\n";
Epetra_MultiVector component( componentMap, 1 );
int Length = component.MyLength();
int shift = Xh->nDofStart( index );
for ( int i=0; i < Length; i++ )
{
int compGlobalID = componentMap.GID( i );
if ( compGlobalID >= 0 )
{
int compLocalID = componentMap.LID( compGlobalID );
int localID = globalMap.LID( compGlobalID+shift );
// int globalID = globalMap.GID(localID);
DVLOG(2) << "[MyBackend] Copy entry sol[" << localID << "]=" << sol[0][localID]
<< " to component[" << compLocalID << "]\n";
component[0][compLocalID] = sol[0][localID];
DVLOG(2) << component[0][compLocalID] << "\n";
}
}
return component;
}
static void UpdateComponent( spaceT const& Xh, Epetra_MultiVector& sol, Epetra_MultiVector& comp )
{
Epetra_Map componentMap ( epetraMap( Xh->template functionSpace<index>()->map() ) );
Epetra_Map globalMap ( epetraMap( Xh->map() ) );
int shift = Xh->nDofStart( index );
int Length = comp.MyLength();
for ( int i=0; i < Length; i++ )
{
int compGlobalID = componentMap.GID( i );
if ( compGlobalID >= 0 )
{
int compLocalID = componentMap.LID( compGlobalID );
int localID = globalMap.LID( compGlobalID+shift );
// int globalID = globalMap.GID(localID);
DVLOG(2) << "Copy entry component[" << compLocalID << "] to sol[" << localID << "]="
<< sol[0][localID]
<< "]\n";
sol[0][localID] = comp[0][compLocalID] ;
DVLOG(2) << comp[0][compLocalID] << "\n";
}
}
}
size_t
AbstractConcreteMatrixAdapter<
Epetra_RowMatrix,
DerivedMat>::getGlobalRowNNZ_impl(global_ordinal_t row) const
{
// check whether row is local, then transform to local index
Epetra_Map rowmap = this->mat_->RowMatrixRowMap();
int gid = Teuchos::as<int>(row);
TEUCHOS_TEST_FOR_EXCEPTION( !rowmap.MyGID(gid),
std::invalid_argument,
"The specified global row id does not belong to me" );
int lid = rowmap.LID(gid);
int nnz = 0;
this->mat_->NumMyRowEntries(lid, nnz);
return nnz;
}
Teuchos::RCP<Epetra_CrsMatrix> Epetra_Operator_to_Epetra_Matrix::constructInverseMatrix(const Epetra_Operator &op, const Epetra_Map &map)
{
int numEntriesPerRow = 0;
Teuchos::RCP<Epetra_FECrsMatrix> matrix = Teuchos::rcp(new Epetra_FECrsMatrix(::Copy, map, numEntriesPerRow));
int numRows = map.NumGlobalElements();
Epetra_Vector X(map);
Epetra_Vector Y(map);
double tol = 1e-15; // values below this will be considered 0
for (int rowIndex=0; rowIndex<numRows; rowIndex++)
{
int lid = map.LID(rowIndex);
if (lid != -1)
{
X[lid] = 1.0;
}
op.ApplyInverse(X, Y);
if (lid != -1)
{
X[lid] = 0.0;
}
std::vector<double> values;
std::vector<int> indices;
for (int i=0; i<map.NumMyElements(); i++)
{
if (abs(Y[i]) > tol)
{
values.push_back(Y[i]);
indices.push_back(map.GID(i));
}
}
matrix->InsertGlobalValues(rowIndex, values.size(), &values[0], &indices[0]);
}
matrix->GlobalAssemble();
return matrix;
}
// Serial only. Very straightforward: one eval per column. Could speed up a lot
// using graph coloring.
double
testJacobian (panzer::AssemblyEngine_TemplateManager<panzer::Traits>& ae_tm,
const Teuchos::RCP<panzer::LinearObjFactory<panzer::Traits> >& linObjFactory,
const Teuchos::RCP<const Epetra_Vector> x0 = Teuchos::null)
{
using Teuchos::RCP;
using Teuchos::rcp;
const RCP<panzer::LinearObjContainer> ghost_con = linObjFactory->buildGhostedLinearObjContainer();
const RCP<panzer::LinearObjContainer> con = linObjFactory->buildLinearObjContainer();
linObjFactory->initializeGhostedContainer(panzer::LinearObjContainer::X |
panzer::LinearObjContainer::F |
panzer::LinearObjContainer::Mat, *ghost_con);
const RCP<panzer::EpetraLinearObjContainer>
ep_con = Teuchos::rcp_dynamic_cast<panzer::EpetraLinearObjContainer>(con);
SparseTriple t;
{ // Finite-difference Jacobian. Should use graph coloring to greatly reduce
// number of evaluations. But brute force for now.
RCP<Epetra_Vector> x, f0;
for (int i = -1; ; ++i) {
linObjFactory->initializeContainer(panzer::LinearObjContainer::X | panzer::LinearObjContainer::F |
panzer::LinearObjContainer::Mat, *con);
ghost_con->initialize();
con->initialize();
const Epetra_Map& row_map = ep_con->get_A()->RowMap(), col_map = ep_con->get_A()->ColMap();
if (i == -1) {
if (Teuchos::nonnull(x0))
x = rcp(new Epetra_Vector(*x0));
else
x = rcp(new Epetra_Vector(ep_con->get_x()->Map()));
t.m = t.n = x->GlobalLength();
}
// For a linear problem, could make delta 1 to remove cancellation
// error. But I want to look at a nonlinear Robin condition, so do a true
// finite difference.
const double delta = 1e-6;
const bool i_mine = row_map.MyGID(i);
double x_prev = 0;
int i_lid = 0;
if (i_mine && i >= 0) {
i_lid = col_map.LID(i);
x_prev = (*x)[i_lid];
(*x)[i_lid] += delta;
}
ep_con->set_x(x);
panzer::AssemblyEngineInArgs input(ghost_con, ep_con);
input.alpha = 0;
input.beta = 1;
ae_tm.getAsObject<panzer::Traits::Residual>()->evaluate(input);
if (i == -1)
f0 = ep_con->get_f();
else {
const Epetra_Vector& f = *ep_con->get_f();
if (i_mine) (*x)[i_lid] = x_prev;
for (int k = 0; k < f.MyLength(); ++k) {
const double d = f[k] - (*f0)[k];
if (d == 0) continue;
t.i.push_back(row_map.GID(k));
t.j.push_back(i);
t.v.push_back(d/delta);
}
}
if (i + 1 == t.m) break;
}
}
{ // AD Jacobian.
linObjFactory->initializeContainer(panzer::LinearObjContainer::X | panzer::LinearObjContainer::F |
panzer::LinearObjContainer::Mat, *con);
ghost_con->initialize();
con->initialize();
if (Teuchos::nonnull(x0))
ep_con->set_x(rcp(new Epetra_Vector(*x0)));
else
ep_con->get_x()->PutScalar(0);
panzer::AssemblyEngineInArgs input(ghost_con, ep_con);
input.alpha = 0;
input.beta = 1;
ae_tm.getAsObject<panzer::Traits::Jacobian>()->evaluate(input);
EpetraExt::RowMatrixToMatrixMarketFile("A_ad.mm", *ep_con->get_A());
}
RCP<Epetra_CrsMatrix> A_fd;
{
const Epetra_Map& row_map = ep_con->get_A()->RowMap(), col_map = ep_con->get_A()->ColMap();
A_fd = rcp(new Epetra_CrsMatrix(Copy, row_map, 40));
for (int i = 0; i < static_cast<int>(t.v.size()); ++i)
A_fd->InsertGlobalValues(t.i[i], 1, &t.v[i], &t.j[i]);
A_fd->FillComplete();
}
EpetraExt::RowMatrixToMatrixMarketFile("A_fd.mm", *A_fd);
{ // Check error.
const double A_fd_inf_norm = A_fd->NormInf();
RCP<Epetra_CrsMatrix> D = rcp(new Epetra_CrsMatrix(Copy, ep_con->get_A()->Graph()));
D->FillComplete();
Epetra_CrsMatrix* D_ptr = D.get();
EpetraExt::MatrixMatrix::Add(*ep_con->get_A(), false, -1, *A_fd, false, 1, D_ptr);
const double D_inf_norm = D->NormInf();
return D_inf_norm / A_fd_inf_norm;
}
}
int checkmap(Epetra_Map & Map, int NumGlobalElements, int NumMyElements,
int *MyGlobalElements, int IndexBase, Epetra_Comm& Comm,
bool DistributedGlobal)
{
int i, ierr=0, forierr = 0;
EPETRA_TEST_ERR(!Map.ConstantElementSize(),ierr);
EPETRA_TEST_ERR(DistributedGlobal!=Map.DistributedGlobal(),ierr);
EPETRA_TEST_ERR(Map.ElementSize()!=1,ierr);
int *MyElementSizeList = new int[NumMyElements];
EPETRA_TEST_ERR(Map.ElementSizeList(MyElementSizeList)!=0,ierr);
forierr = 0;
for (i=0; i<NumMyElements; i++) forierr += MyElementSizeList[i]!=1;
EPETRA_TEST_ERR(forierr,ierr);
delete [] MyElementSizeList;
const Epetra_Comm & Comm1 = Map.Comm();
EPETRA_TEST_ERR(Comm1.NumProc()!=Comm.NumProc(),ierr);
EPETRA_TEST_ERR(Comm1.MyPID()!=Comm.MyPID(),ierr);
EPETRA_TEST_ERR(Map.IndexBase()!=IndexBase,ierr);
EPETRA_TEST_ERR(!Map.LinearMap() && MyGlobalElements==0,ierr);
EPETRA_TEST_ERR(Map.LinearMap() && MyGlobalElements!=0,ierr);
EPETRA_TEST_ERR(Map.MaxAllGID()!=NumGlobalElements-1+IndexBase,ierr);
EPETRA_TEST_ERR(Map.MaxElementSize()!=1,ierr);
int MaxLID = Map.MaxLID();
EPETRA_TEST_ERR(MaxLID!=NumMyElements-1,ierr);
int MaxMyGID = (Comm.MyPID()+1)*NumMyElements-1+IndexBase;
if (Comm.MyPID()>2) MaxMyGID+=3;
if (!DistributedGlobal) MaxMyGID = NumMyElements-1+IndexBase;
EPETRA_TEST_ERR(Map.MaxMyGID()!=MaxMyGID,ierr);
EPETRA_TEST_ERR(Map.MinAllGID()!=IndexBase,ierr);
EPETRA_TEST_ERR(Map.MinElementSize()!=1,ierr);
EPETRA_TEST_ERR(Map.MinLID()!=0,ierr);
int MinMyGID = Comm.MyPID()*NumMyElements+IndexBase;
if (Comm.MyPID()>2) MinMyGID+=3;
if (!DistributedGlobal) MinMyGID = 0;
EPETRA_TEST_ERR(Map.MinMyGID()!=MinMyGID,ierr);
int * MyGlobalElements1 = new int[NumMyElements];
EPETRA_TEST_ERR(Map.MyGlobalElements(MyGlobalElements1)!=0,ierr);
forierr = 0;
if (MyGlobalElements==0)
{
for (i=0; i<NumMyElements; i++)
forierr += MyGlobalElements1[i]!=MinMyGID+i;
EPETRA_TEST_ERR(forierr,ierr);
}
else {
for (i=0; i<NumMyElements; i++)
forierr += MyGlobalElements[i]!=MyGlobalElements1[i];
EPETRA_TEST_ERR(forierr,ierr);
}
EPETRA_TEST_ERR(Map.NumGlobalElements()!=NumGlobalElements,ierr);
EPETRA_TEST_ERR(Map.NumGlobalPoints()!=NumGlobalElements,ierr);
EPETRA_TEST_ERR(Map.NumMyElements()!=NumMyElements,ierr);
EPETRA_TEST_ERR(Map.NumMyPoints()!=NumMyElements,ierr);
int MaxMyGID2 = Map.GID(Map.LID(MaxMyGID));
EPETRA_TEST_ERR(MaxMyGID2 != MaxMyGID,ierr);
int MaxLID2 = Map.LID(Map.GID(MaxLID));
EPETRA_TEST_ERR(MaxLID2 != MaxLID,ierr);
EPETRA_TEST_ERR(Map.GID(MaxLID+1) != IndexBase-1,ierr);// MaxLID+1 doesn't exist
EPETRA_TEST_ERR(Map.LID(MaxMyGID+1) != -1,ierr);// MaxMyGID+1 doesn't exist or is on a different processor
EPETRA_TEST_ERR(!Map.MyGID(MaxMyGID),ierr);
EPETRA_TEST_ERR(Map.MyGID(MaxMyGID+1),ierr);
EPETRA_TEST_ERR(!Map.MyLID(MaxLID),ierr);
EPETRA_TEST_ERR(Map.MyLID(MaxLID+1),ierr);
EPETRA_TEST_ERR(!Map.MyGID(Map.GID(MaxLID)),ierr);
EPETRA_TEST_ERR(Map.MyGID(Map.GID(MaxLID+1)),ierr);
EPETRA_TEST_ERR(!Map.MyLID(Map.LID(MaxMyGID)),ierr);
EPETRA_TEST_ERR(Map.MyLID(Map.LID(MaxMyGID+1)),ierr);
// Check RemoteIDList function
// Get some GIDs off of each processor to test
int TotalNumEle, NumElePerProc, NumProc = Comm.NumProc();
int MinNumEleOnProc;
int NumMyEle=Map.NumMyElements();
Comm.MinAll(&NumMyEle,&MinNumEleOnProc,1);
if (MinNumEleOnProc > 5) NumElePerProc = 6;
else NumElePerProc = MinNumEleOnProc;
if (NumElePerProc > 0) {
TotalNumEle = NumElePerProc*NumProc;
int * MyGIDlist = new int[NumElePerProc];
int * GIDlist = new int[TotalNumEle];
int * PIDlist = new int[TotalNumEle];
int * LIDlist = new int[TotalNumEle];
for (i=0; i<NumElePerProc; i++)
MyGIDlist[i] = MyGlobalElements1[i];
Comm.GatherAll(MyGIDlist,GIDlist,NumElePerProc);// Get a few values from each proc
Map.RemoteIDList(TotalNumEle, GIDlist, PIDlist, LIDlist);
int MyPID= Comm.MyPID();
forierr = 0;
for (i=0; i<TotalNumEle; i++) {
if (Map.MyGID(GIDlist[i])) {
forierr += PIDlist[i] != MyPID;
forierr += !Map.MyLID(Map.LID(GIDlist[i])) || Map.LID(GIDlist[i]) != LIDlist[i] || Map.GID(LIDlist[i]) != GIDlist[i];
}
else {
forierr += PIDlist[i] == MyPID; // If MyGID comes back false, the PID listed should be that of another proc
}
}
EPETRA_TEST_ERR(forierr,ierr);
delete [] MyGIDlist;
delete [] GIDlist;
delete [] PIDlist;
delete [] LIDlist;
}
delete [] MyGlobalElements1;
// Check RemoteIDList function (assumes all maps are linear, even if not stored that way)
if (Map.LinearMap()) {
int * GIDList = new int[3];
int * PIDList = new int[3];
int * LIDList = new int[3];
int MyPID = Map.Comm().MyPID();
int NumIDs = 0;
//GIDList[NumIDs++] = Map.MaxAllGID()+1; // Should return -1 for both PID and LID
if (Map.MinMyGID()-1>=Map.MinAllGID()) GIDList[NumIDs++] = Map.MinMyGID()-1;
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) GIDList[NumIDs++] = Map.MaxMyGID()+1;
Map.RemoteIDList(NumIDs, GIDList, PIDList, LIDList);
NumIDs = 0;
//EPETRA_TEST_ERR(!(PIDList[NumIDs]==-1),ierr);
//EPETRA_TEST_ERR(!(LIDList[NumIDs++]==-1),ierr);
if (Map.MinMyGID()-1>=Map.MinAllGID()) EPETRA_TEST_ERR(!(PIDList[NumIDs++]==MyPID-1),ierr);
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) EPETRA_TEST_ERR(!(PIDList[NumIDs]==MyPID+1),ierr);
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) EPETRA_TEST_ERR(!(LIDList[NumIDs++]==0),ierr);
delete [] GIDList;
delete [] PIDList;
delete [] LIDList;
}
return (ierr);
}
int TLowCommunicationMakeColMapAndReindex(int N, const int * rowptr, int * colind_LID, const int_type *colind_GID, const Epetra_Map& domainMap, const int * owningPIDs, bool SortGhostsAssociatedWithEachProcessor, std::vector<int>& RemotePIDs, MapType1 & NewColMap)
{
int i,j;
// Sanity checks
bool UseLL;
if(domainMap.GlobalIndicesLongLong()) UseLL=true;
else if(domainMap.GlobalIndicesInt()) UseLL=false;
else throw std::runtime_error("LowCommunicationMakeColMapAndReindex: cannot detect int type.");
// Scan all column indices and sort into two groups:
// Local: those whose GID matches a GID of the domain map on this processor and
// Remote: All others.
int numDomainElements = domainMap.NumMyElements();
bool * LocalGIDs = 0;
if (numDomainElements>0) LocalGIDs = new bool[numDomainElements];
for (i=0; i<numDomainElements; i++) LocalGIDs[i] = false; // Assume domain GIDs are not local
bool DoSizes = !domainMap.ConstantElementSize(); // If not constant element size, then error
if(DoSizes) throw std::runtime_error("LowCommunicationMakeColMapAndReindex: cannot handle non-constant sized domainMap.");
// In principle it is good to have RemoteGIDs and RemotGIDList be as long as the number of remote GIDs
// on this processor, but this would require two passes through the column IDs, so we make it the max of 100
// and the number of block rows.
const int numMyBlockRows = N;
int hashsize = numMyBlockRows; if (hashsize < 100) hashsize = 100;
Epetra_HashTable<int_type> RemoteGIDs(hashsize);
std::vector<int_type> RemoteGIDList; RemoteGIDList.reserve(hashsize);
std::vector<int> PIDList; PIDList.reserve(hashsize);
// Here we start using the *int* colind array. If int_type==int this clobbers the GIDs, if
// int_type==long long, then this is the first use of the colind array.
// For *local* GID's set colind with with their LID in the domainMap. For *remote* GIDs,
// we set colind with (numDomainElements+NumRemoteColGIDs) before the increment of
// the remote count. These numberings will be separate because no local LID is greater
// than numDomainElements.
int NumLocalColGIDs = 0;
int NumRemoteColGIDs = 0;
for(i = 0; i < numMyBlockRows; i++) {
for(j = rowptr[i]; j < rowptr[i+1]; j++) {
int_type GID = colind_GID[j];
// Check if GID matches a row GID
int LID = domainMap.LID(GID);
if(LID != -1) {
bool alreadyFound = LocalGIDs[LID];
if (!alreadyFound) {
LocalGIDs[LID] = true; // There is a column in the graph associated with this domain map GID
NumLocalColGIDs++;
}
colind_LID[j] = LID;
}
else {
int_type hash_value=RemoteGIDs.Get(GID);
if(hash_value == -1) { // This means its a new remote GID
int PID = owningPIDs[j];
if(PID==-1) throw std::runtime_error("LowCommunicationMakeColMapAndReindex: Cannot figure out if PID is owned.");
colind_LID[j] = numDomainElements + NumRemoteColGIDs;
RemoteGIDs.Add(GID, NumRemoteColGIDs);
RemoteGIDList.push_back(GID);
PIDList.push_back(PID);
NumRemoteColGIDs++;
}
else
colind_LID[j] = numDomainElements + hash_value;
}
}
}
// Possible short-circuit: If all domain map GIDs are present as column indices, then set ColMap=domainMap and quit
if (domainMap.Comm().NumProc()==1) {
if (NumRemoteColGIDs!=0) {
throw std::runtime_error("Some column IDs are not in domainMap. If matrix is rectangular, you must pass in a domainMap");
// Sanity test: When one processor,there can be no remoteGIDs
}
if (NumLocalColGIDs==numDomainElements) {
if (LocalGIDs!=0) delete [] LocalGIDs;
// In this case, we just use the domainMap's indices, which is, not coincidently, what we clobbered colind with up above anyway.
// No further reindexing is needed.
NewColMap = domainMap;
return 0;
}
}
// Now build integer array containing column GIDs
// Build back end, containing remote GIDs, first
int numMyBlockCols = NumLocalColGIDs + NumRemoteColGIDs;
std::vector<int_type> ColIndices;
int_type * RemoteColIndices=0;
if(numMyBlockCols > 0) {
ColIndices.resize(numMyBlockCols);
if(NumLocalColGIDs!=numMyBlockCols) RemoteColIndices = &ColIndices[NumLocalColGIDs]; // Points to back end of ColIndices
else RemoteColIndices=0;
}
for(i = 0; i < NumRemoteColGIDs; i++)
RemoteColIndices[i] = RemoteGIDList[i];
// Build permute array for *remote* reindexing.
std::vector<int> RemotePermuteIDs(NumRemoteColGIDs);
for(i=0; i<NumRemoteColGIDs; i++) RemotePermuteIDs[i]=i;
// Sort External column indices so that all columns coming from a given remote processor are contiguous
int NumListsInt=0;
int NumListsLL =0;
int * IntSortLists[2];
long long * LLSortLists[2];
int * RemotePermuteIDs_ptr = RemotePermuteIDs.size() ? &RemotePermuteIDs[0] : 0;
if(!UseLL) {
// int version
IntSortLists[0] = (int*) RemoteColIndices;
IntSortLists[1] = RemotePermuteIDs_ptr;
NumListsInt=2;
}
else {
//LL version
LLSortLists[0] = (long long*) RemoteColIndices;
IntSortLists[0] = RemotePermuteIDs_ptr;
NumListsInt = NumListsLL = 1;
}
int * PIDList_ptr = PIDList.size() ? &PIDList[0] : 0;
Epetra_Util::Sort(true, NumRemoteColGIDs, PIDList_ptr, 0, 0, NumListsInt, IntSortLists,NumListsLL,LLSortLists);
// Stash the RemotePIDs
PIDList.resize(NumRemoteColGIDs);
RemotePIDs = PIDList;
if (SortGhostsAssociatedWithEachProcessor) {
// Sort external column indices so that columns from a given remote processor are not only contiguous
// but also in ascending order. NOTE: I don't know if the number of externals associated
// with a given remote processor is known at this point ... so I count them here.
// NTS: Only sort the RemoteColIndices this time...
int StartCurrent, StartNext;
StartCurrent = 0; StartNext = 1;
while ( StartNext < NumRemoteColGIDs ) {
if (PIDList[StartNext]==PIDList[StartNext-1]) StartNext++;
else {
IntSortLists[0] = &RemotePermuteIDs[StartCurrent];
Epetra_Util::Sort(true,StartNext-StartCurrent, &(RemoteColIndices[StartCurrent]),0,0,1,IntSortLists,0,0);
StartCurrent = StartNext; StartNext++;
}
}
IntSortLists[0] = &RemotePermuteIDs[StartCurrent];
Epetra_Util::Sort(true, StartNext-StartCurrent, &(RemoteColIndices[StartCurrent]), 0, 0, 1,IntSortLists,0,0);
}
// Reverse the permutation to get the information we actually care about
std::vector<int> ReverseRemotePermuteIDs(NumRemoteColGIDs);
for(i=0; i<NumRemoteColGIDs; i++) ReverseRemotePermuteIDs[RemotePermuteIDs[i]]=i;
// Build permute array for *local* reindexing.
bool use_local_permute=false;
std::vector<int> LocalPermuteIDs(numDomainElements);
// Now fill front end. Two cases:
// (1) If the number of Local column GIDs is the same as the number of Local domain GIDs, we
// can simply read the domain GIDs into the front part of ColIndices, otherwise
// (2) We step through the GIDs of the domainMap, checking to see if each domain GID is a column GID.
// we want to do this to maintain a consistent ordering of GIDs between the columns and the domain.
if(NumLocalColGIDs == domainMap.NumMyElements()) {
if(NumLocalColGIDs > 0) {
domainMap.MyGlobalElements(&ColIndices[0]); // Load Global Indices into first numMyBlockCols elements column GID list
}
}
else {
int_type* MyGlobalElements = 0;
domainMap.MyGlobalElementsPtr(MyGlobalElements);
int* ElementSizeList = 0;
if(DoSizes)
ElementSizeList = domainMap.ElementSizeList();
int NumLocalAgain = 0;
use_local_permute = true;
for(i = 0; i < numDomainElements; i++) {
if(LocalGIDs[i]) {
LocalPermuteIDs[i] = NumLocalAgain;
ColIndices[NumLocalAgain++] = MyGlobalElements[i];
}
}
assert(NumLocalAgain==NumLocalColGIDs); // Sanity test
}
// Done with this array
if (LocalGIDs!=0) delete [] LocalGIDs;
// Make Column map with same element sizes as Domain map
int_type * ColIndices_ptr = ColIndices.size() ? &ColIndices[0] : 0;
MapType2 temp((int_type)(-1), numMyBlockCols, ColIndices_ptr, (int_type)domainMap.IndexBase64(), domainMap.Comm());
NewColMap = temp;
// Low-cost reindex of the matrix
for(i=0; i<numMyBlockRows; i++){
for(j=rowptr[i]; j<rowptr[i+1]; j++){
int ID=colind_LID[j];
if(ID < numDomainElements){
if(use_local_permute) colind_LID[j] = LocalPermuteIDs[colind_LID[j]];
// In the case where use_local_permute==false, we just copy the DomainMap's ordering, which it so happens
// is what we put in colind to begin with.
}
else
colind_LID[j] = NumLocalColGIDs + ReverseRemotePermuteIDs[colind_LID[j]-numDomainElements];
}
}
return 0;
}