TridiagonalCrsMatrix(
const Epetra_Map & Map,
double a,
double diag,
double c
) :
Epetra_CrsMatrix(Copy,Map,3)
{
// global number of rows
int NumGlobalElements = Map.NumGlobalElements();
// local number of rows
int NumMyElements = Map.NumMyElements();
// get update list
int * MyGlobalElements = new int [NumMyElements];
Map.MyGlobalElements( MyGlobalElements );
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
double *Values = new double[2];
Values[0] = a;
Values[1] = c;
int *Indices = new int[2];
int NumEntries;
for( int i=0 ; i<NumMyElements; ++i ) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
NumEntries = 1;
} else if (MyGlobalElements[i] == NumGlobalElements-1) {
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
} else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
NumEntries = 2;
}
InsertGlobalValues(MyGlobalElements[i], NumEntries, Values, Indices);
// Put in the diagonal entry
InsertGlobalValues(MyGlobalElements[i], 1, &diag, MyGlobalElements+i);
}
// Finish up
FillComplete();
delete [] MyGlobalElements;
delete [] Values;
delete [] Indices;
}
void build_test_map(const Epetra_Map & oldMap, Epetra_Map *& newMap) {
int NumProc = oldMap.Comm().NumProc();
int MyPID = oldMap.Comm().MyPID();
int num_global = oldMap.NumGlobalElements();
if(NumProc<3) {
// Dump everything onto -proc 0
int num_local = MyPID==0 ? num_global : 0;
newMap = new Epetra_Map(num_global,num_local,0,oldMap.Comm());
}
else {
// Split everything between procs 0 and 2 (leave proc 1 empty)
int num_local=0;
if(MyPID==0) num_local = num_global/2;
else if(MyPID==2) num_local = num_global - ((int)num_global/2);
newMap = new Epetra_Map(num_global,num_local,0,oldMap.Comm());
}
}
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;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
cout << Comm << endl;
int MyPID = Comm.MyPID();
bool verbose = false;
if (MyPID==0) verbose = true;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename [level_fill [level_overlap [absolute_threshold [ relative_threshold]]]]" << endl
<< "where:" << endl
<< "HB_filename - filename and path of a Harwell-Boeing data set" << endl
<< "level_fill - The amount of fill to use for ILU(k) preconditioner (default 0)" << endl
<< "level_overlap - The amount of overlap used for overlapping Schwarz subdomains (default 0)" << endl
<< "absolute_threshold - The minimum value to place on the diagonal prior to factorization (default 0.0)" << endl
<< "relative_threshold - The relative amount to perturb the diagonal prior to factorization (default 1.0)" << endl << endl
<< "To specify a non-default value for one of these parameters, you must specify all" << endl
<< " preceding values but not any subsequent parameters. Example:" << endl
<< "ifpackHbSerialMsr.exe mymatrix.hb 1 - loads mymatrix.hb, uses level fill of one, all other values are defaults" << endl
<< endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
// Construct ILU preconditioner
double elapsed_time, total_flops, MFLOPs;
Epetra_Time timer(Comm);
int LevelFill = 0;
if (argc > 2) LevelFill = atoi(argv[2]);
if (verbose) cout << "Using Level Fill = " << LevelFill << endl;
int Overlap = 0;
if (argc > 3) Overlap = atoi(argv[3]);
if (verbose) cout << "Using Level Overlap = " << Overlap << endl;
double Athresh = 0.0;
if (argc > 4) Athresh = atof(argv[4]);
if (verbose) cout << "Using Absolute Threshold Value of = " << Athresh << endl;
double Rthresh = 1.0;
if (argc > 5) Rthresh = atof(argv[5]);
if (verbose) cout << "Using Relative Threshold Value of = " << Rthresh << endl;
Ifpack_IlukGraph * IlukGraph = 0;
Ifpack_CrsRiluk * ILUK = 0;
if (LevelFill>-1) {
elapsed_time = timer.ElapsedTime();
IlukGraph = new Ifpack_IlukGraph(A.Graph(), LevelFill, Overlap);
assert(IlukGraph->ConstructFilledGraph()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
if (verbose) cout << "Time to construct ILUK graph = " << elapsed_time << endl;
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
ILUK = new Ifpack_CrsRiluk(*IlukGraph);
ILUK->SetFlopCounter(fact_counter);
ILUK->SetAbsoluteThreshold(Athresh);
ILUK->SetRelativeThreshold(Rthresh);
//assert(ILUK->InitValues()==0);
int initerr = ILUK->InitValues(A);
if (initerr!=0) cout << Comm << "InitValues error = " << initerr;
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
//cout << *ILUK << endl;
}
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
int Maxiter = 500;
double Tolerance = 1.0E-14;
Epetra_Vector xcomp(map);
Epetra_Vector resid(map);
Epetra_Flops counter;
A.SetFlopCounter(counter);
xcomp.SetFlopCounter(A);
b.SetFlopCounter(A);
resid.SetFlopCounter(A);
ILUK->SetFlopCounter(A);
elapsed_time = timer.ElapsedTime();
BiCGSTAB(A, xcomp, b, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
resid.Update(1.0, xcomp, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
if (verbose) cout << "Norm of the difference between exact and computed solutions = " << residual << endl;
if (ILUK!=0) delete ILUK;
if (IlukGraph!=0) delete IlukGraph;
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
/* Apply an identity matrix to the Schur complement operator. Drop the entries
entries using a relative threshold. Assemble the result in a Crs Matrix
which will be our approximate Schur complement.
*/
Teuchos::RCP<Epetra_CrsMatrix> computeApproxSchur(shylu_config *config,
shylu_symbolic *sym,
Epetra_CrsMatrix *G, Epetra_CrsMatrix *R,
Epetra_LinearProblem *LP, Amesos_BaseSolver *solver,
Ifpack_Preconditioner *ifSolver, Epetra_CrsMatrix *C,
Epetra_Map *localDRowMap)
{
double relative_thres = config->relative_threshold;
int nvectors = 16;
ShyLU_Probing_Operator probeop(config, sym, G, R, LP, solver, ifSolver, C,
localDRowMap, nvectors);
// Get row map
Epetra_Map rMap = G->RowMap();
int *rows = rMap.MyGlobalElements();
int totalElems = rMap.NumGlobalElements();
int localElems = rMap.NumMyElements();
//cout << " totalElems in Schur Complement" << totalElems << endl;
//cout << myPID << " localElems" << localElems << endl;
// **************** Two collectives here *********************
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
ftime.start();
#endif
int prefixSum;
G->Comm().ScanSum(&localElems, &prefixSum, 1);
//cout << " prefixSum" << prefixSum << endl;
// Start the index in prefixSum-localElems
int *mySGID = new int[totalElems]; // vector of size Schur complement !
int *allSGID = new int[totalElems]; // vector of size Schur complement !
int i, j;
for (i = 0, j = 0; i < totalElems ; i++)
{
if (i >= prefixSum - localElems && i < prefixSum)
{
mySGID[i] = rows[j];
j++;
}
else
{
mySGID[i] = 0;
}
allSGID[i] = 0;
}
C->Comm().SumAll(mySGID, allSGID, totalElems);
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time to Compute RowIDS" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
// Now everyone knows the GIDs in the Schur complement
//cout << rMap << endl;
j = 0;
Teuchos::RCP<Epetra_CrsMatrix> Sbar = Teuchos::rcp(new Epetra_CrsMatrix(
Copy, rMap, localElems));
int nentries;
double *values = new double[localElems]; // Need to adjust this for more
int *indices = new int[localElems]; // than one vector
double *vecvalues;
int dropped = 0;
double *maxvalue = new double[nvectors];
#ifdef TIMING_OUTPUT
ftime.start();
#endif
#ifdef TIMING_OUTPUT
Teuchos::Time app_time("Apply time");
#endif
int findex = totalElems / nvectors ;
for (i = 0 ; i < findex*nvectors ; i+=nvectors)
{
Epetra_MultiVector probevec(rMap, nvectors);
Epetra_MultiVector Scol(rMap, nvectors);
probevec.PutScalar(0.0);
int cindex;
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
if (cindex >= prefixSum - localElems && cindex < prefixSum)
{
probevec.ReplaceGlobalValue(allSGID[cindex], k, 1.0);
}
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
Scol.MaxValue(maxvalue);
for (int k = 0; k < nvectors; k++) //TODO:Need to switch these loops
{
cindex = k+i;
vecvalues = Scol[k];
//cout << "MAX" << maxvalue << endl;
for (j = 0 ; j < localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (allSGID[cindex] == rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[cindex];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else if (abs(vecvalues[j]/maxvalue[k]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[cindex];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else
{
if (vecvalues[j] != 0.0) dropped++;
}
}
}
}
probeop.ResetTempVectors(1);
for ( ; i < totalElems ; i++)
{
Epetra_MultiVector probevec(rMap, 1); // TODO: Try doing more than one
Epetra_MultiVector Scol(rMap, 1); // vector at a time
probevec.PutScalar(0.0);
if (i >= prefixSum - localElems && i < prefixSum)
{
probevec.ReplaceGlobalValue(allSGID[i], 0, 1.0);
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
vecvalues = Scol[0];
Scol.MaxValue(maxvalue);
//cout << "MAX" << maxvalue << endl;
for (j = 0 ; j < localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (allSGID[i] == rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[i];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else if (abs(vecvalues[j]/maxvalue[0]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[i];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else
{
if (vecvalues[j] != 0.0) dropped++;
}
}
}
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time in finding and dropping entries" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
#ifdef TIMING_OUTPUT
cout << "Time in Apply of probing" << app_time.totalElapsedTime() << endl;
#endif
Sbar->FillComplete();
cout << "#dropped entries" << dropped << endl;
delete[] allSGID;
delete[] mySGID;
delete[] values;
delete[] indices;
delete[] maxvalue;
return Sbar;
}
int CreateTridi(Epetra_CrsMatrix& A)
{
Epetra_Map Map = A.RowMap();
int NumMyElements = Map.NumMyElements();
int NumGlobalElements = Map.NumGlobalElements();
int * MyGlobalElements = new int[NumMyElements];
Map.MyGlobalElements(MyGlobalElements);
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
double *Values = new double[3];
int *Indices = new int[3];
int NumEntries;
for (int i=0; i<NumMyElements; i++)
{
if (MyGlobalElements[i]==0)
{
Indices[0] = 0;
Indices[1] = 1;
Values[0] = 2.0;
Values[1] = -1.0;
NumEntries = 2;
}
else if (MyGlobalElements[i] == NumGlobalElements-1)
{
Indices[0] = NumGlobalElements-1;
Indices[1] = NumGlobalElements-2;
Values[0] = 2.0;
Values[1] = -1.0;
NumEntries = 2;
}
else
{
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i];
Indices[2] = MyGlobalElements[i]+1;
Values[0] = -1.0;
Values[1] = 2.0;
Values[2] = -1.0;
NumEntries = 3;
}
assert(A.InsertGlobalValues(MyGlobalElements[i], NumEntries, Values, Indices)==0);
// Put in the diagonal entry
// assert(A.InsertGlobalValues(MyGlobalElements[i], 1, &two, &MyGlobalElements[i])==0);
}
// Finish up
assert(A.FillComplete()==0);
delete[] MyGlobalElements;
delete[] Values;
delete[] Indices;
return 0;
}
// *************************************************************
// main program - This benchmark code reads a Harwell-Boeing data
// set and finds the minimal eigenvalue of the matrix
// using inverse iteration.
// *************************************************************
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
cout << Comm << endl;
int MyPID = Comm.MyPID();
bool verbose = false;
if (MyPID==0) verbose = true; // Print out detailed results (turn off for best performance)
if(argc != 2) {
if (verbose) cerr << "Usage: " << argv[0] << " HB_data_file" << endl;
exit(1); // Error
}
// Define pointers that will be set by HB read function
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call function to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Not interested in x, b or xexact for an eigenvalue problem
delete readx;
delete readb;
delete readxexact;
#ifdef EPETRA_MPI // If running in parallel, we need to distribute matrix across all PEs.
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
A.Export(*readA, exporter, Add);
assert(A.FillComplete()==0);
delete readA;
delete readMap;
#else // If not running in parallel, we do not need to distribute the matrix
Epetra_CrsMatrix & A = *readA;
#endif
// Create flop counter to collect all FLOPS
Epetra_Flops counter;
A.SetFlopCounter(counter);
double lambda = 0; // Minimal eigenvalue returned here
// Call inverse iteration solver
Epetra_Time timer(Comm);
invIteration(A, lambda, verbose);
double elapsedTime = timer.ElapsedTime();
double totalFlops = counter.Flops();
double MFLOPS = totalFlops/elapsedTime/1000000.0;
cout << endl
<< "*************************************************" << endl
<< " Approximate smallest eigenvalue = " << lambda << endl
<< " Total Time = " << elapsedTime << endl
<< " Total FLOPS = " << totalFlops << endl
<< " Total MFLOPS = " << MFLOPS << endl
<< "*************************************************" << endl;
// All done
#ifdef EPETRA_MPI
MPI_Finalize();
#else
delete readA;
delete readMap;
#endif
return (0);
}
void
LOCA::Epetra::AugmentedOp::buildExtendedMap(const Epetra_BlockMap& uMap,
Epetra_Map*& eMapPtr,
bool buildImporter,
bool haveParam)
{
Epetra_BlockMap& nonconstUnderlyingMap = const_cast<Epetra_BlockMap&>(uMap);
// Convert underlying map to point map if necessary
Epetra_Map* uPointMapPtr =
dynamic_cast<Epetra_Map*>(&nonconstUnderlyingMap);
bool allocatedPointMap = false;
if (uPointMapPtr == NULL) {
allocatedPointMap = true;
blockMap2PointMap(uMap, uPointMapPtr);
}
int max_gid = uPointMapPtr->MaxAllGID();
int num_global_elements = uPointMapPtr->NumGlobalElements();
int num_my_elements = uPointMapPtr->NumMyElements();
int *global_elements = uPointMapPtr->MyGlobalElements();
const Epetra_Comm& comm = uPointMapPtr->Comm();
int index_base = uPointMapPtr->IndexBase();
int ext_num_global_elements;
int ext_num_my_elements;
int *ext_global_elements;
// Compute number of extended global elements
if (buildImporter)
ext_num_global_elements =
num_global_elements + numConstraints*comm.NumProc();
else
ext_num_global_elements = num_global_elements + numConstraints;
// Compute number of extended local elements
if (buildImporter || haveParam)
ext_num_my_elements = num_my_elements + numConstraints;
else
ext_num_my_elements = num_my_elements;
// Allocate extended global elements array
ext_global_elements = new int[ext_num_my_elements];
// Set extended global elements
for (int i=0; i<num_my_elements; i++) {
ext_global_elements[i] = global_elements[i];
}
if (buildImporter || haveParam)
for (int i=0; i<numConstraints; i++)
ext_global_elements[num_my_elements+i] = max_gid + 1 + i;
// Create extended point map
eMapPtr = new Epetra_Map(ext_num_global_elements, ext_num_my_elements,
ext_global_elements, index_base, comm);
// Free global elements array
delete [] ext_global_elements;
if (allocatedPointMap)
delete uPointMapPtr;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
bool verbose = true;
if (MyPID==0) verbose = true;
if (verbose)
cout << EpetraExt::EpetraExt_Version() << endl << endl;
cout << Comm << endl;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename" << endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
EpetraExt::RowMatrixToMatrixMarketFile("test.mm", A, "test matrix", "This is a test matrix");
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
//
// Amesos_TestMultiSolver.cpp reads in a matrix in Harwell-Boeing format,
// calls one of the sparse direct solvers, using blocked right hand sides
// and computes the error and residual.
//
// TestSolver ignores the Harwell-Boeing right hand sides, creating
// random right hand sides instead.
//
// Amesos_TestMultiSolver can test either A x = b or A^T x = b.
// This can be a bit confusing because sparse direct solvers
// use compressed column storage - the transpose of Trilinos'
// sparse row storage.
//
// Matrices:
// readA - Serial. As read from the file.
// transposeA - Serial. The transpose of readA.
// serialA - if (transpose) then transposeA else readA
// distributedA - readA distributed to all processes
// passA - if ( distributed ) then distributedA else serialA
//
//
int Amesos_TestMultiSolver( Epetra_Comm &Comm, char *matrix_file, int numsolves,
SparseSolverType SparseSolver, bool transpose,
int special, AMESOS_MatrixType matrix_type ) {
int iam = Comm.MyPID() ;
// int hatever;
// if ( iam == 0 ) std::cin >> hatever ;
Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
std::string FileName = matrix_file ;
int FN_Size = FileName.size() ;
std::string LastFiveBytes = FileName.substr( EPETRA_MAX(0,FN_Size-5), FN_Size );
std::string LastFourBytes = FileName.substr( EPETRA_MAX(0,FN_Size-4), FN_Size );
bool NonContiguousMap = false;
if ( LastFiveBytes == ".triU" ) {
NonContiguousMap = true;
// Call routine to read in unsymmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, false, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFiveBytes == ".triS" ) {
NonContiguousMap = true;
// Call routine to read in symmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, true, Comm,
readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFourBytes == ".mtx" ) {
EPETRA_CHK_ERR( Trilinos_Util_ReadMatrixMarket2Epetra( matrix_file, Comm, readMap,
readA, readx, readb, readxexact) );
} else {
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra( matrix_file, Comm, readMap, readA, readx,
readb, readxexact) ;
}
}
}
Epetra_CrsMatrix transposeA(Copy, *readMap, 0);
Epetra_CrsMatrix *serialA ;
if ( transpose ) {
assert( CrsMatrixTranspose( readA, &transposeA ) == 0 );
serialA = &transposeA ;
} else {
serialA = readA ;
}
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
Epetra_Map* map_;
if( NonContiguousMap ) {
//
// map gives us NumMyElements and MyFirstElement;
//
int NumGlobalElements = readMap->NumGlobalElements();
int NumMyElements = map.NumMyElements();
int MyFirstElement = map.MinMyGID();
std::vector<int> MapMap_( NumGlobalElements );
readMap->MyGlobalElements( &MapMap_[0] ) ;
Comm.Broadcast( &MapMap_[0], NumGlobalElements, 0 ) ;
map_ = new Epetra_Map( NumGlobalElements, NumMyElements, &MapMap_[MyFirstElement], 0, Comm);
} else {
map_ = new Epetra_Map( map ) ;
}
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, *map_);
Epetra_CrsMatrix A(Copy, *map_, 0);
Epetra_RowMatrix * passA = 0;
Epetra_MultiVector * passx = 0;
Epetra_MultiVector * passb = 0;
Epetra_MultiVector * passxexact = 0;
Epetra_MultiVector * passresid = 0;
Epetra_MultiVector * passtmp = 0;
Epetra_MultiVector x(*map_,numsolves);
Epetra_MultiVector b(*map_,numsolves);
Epetra_MultiVector xexact(*map_,numsolves);
Epetra_MultiVector resid(*map_,numsolves);
Epetra_MultiVector tmp(*map_,numsolves);
Epetra_MultiVector serialx(*readMap,numsolves);
Epetra_MultiVector serialb(*readMap,numsolves);
Epetra_MultiVector serialxexact(*readMap,numsolves);
Epetra_MultiVector serialresid(*readMap,numsolves);
Epetra_MultiVector serialtmp(*readMap,numsolves);
bool distribute_matrix = ( matrix_type == AMESOS_Distributed ) ;
if ( distribute_matrix ) {
//
// Initialize x, b and xexact to the values read in from the file
//
A.Export(*serialA, exporter, Add);
Comm.Barrier();
assert(A.FillComplete()==0);
Comm.Barrier();
passA = &A;
passx = &x;
passb = &b;
passxexact = &xexact;
passresid = &resid;
passtmp = &tmp;
} else {
passA = serialA;
passx = &serialx;
passb = &serialb;
passxexact = &serialxexact;
passresid = &serialresid;
passtmp = &serialtmp;
}
passxexact->SetSeed(131) ;
passxexact->Random();
passx->SetSeed(11231) ;
passx->Random();
passb->PutScalar( 0.0 );
passA->Multiply( transpose, *passxexact, *passb ) ;
Epetra_MultiVector CopyB( *passb ) ;
double Anorm = passA->NormInf() ;
SparseDirectTimingVars::SS_Result.Set_Anorm(Anorm) ;
Epetra_LinearProblem Problem( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb );
double max_resid = 0.0;
for ( int j = 0 ; j < special+1 ; j++ ) {
Epetra_Time TotalTime( Comm ) ;
if ( false ) {
#ifdef TEST_UMFPACK
unused code
} else if ( SparseSolver == UMFPACK ) {
UmfpackOO umfpack( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
umfpack.SetTrans( transpose ) ;
umfpack.Solve() ;
#endif
#ifdef TEST_SUPERLU
} else if ( SparseSolver == SuperLU ) {
SuperluserialOO superluserial( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
superluserial.SetPermc( SuperLU_permc ) ;
superluserial.SetTrans( transpose ) ;
superluserial.SetUseDGSSV( special == 0 ) ;
superluserial.Solve() ;
#endif
#ifdef HAVE_AMESOS_SLUD
} else if ( SparseSolver == SuperLUdist ) {
SuperludistOO superludist( Problem ) ;
superludist.SetTrans( transpose ) ;
EPETRA_CHK_ERR( superludist.Solve( true ) ) ;
#endif
#ifdef HAVE_AMESOS_SLUD2
} else if ( SparseSolver == SuperLUdist2 ) {
Superludist2_OO superludist2( Problem ) ;
superludist2.SetTrans( transpose ) ;
EPETRA_CHK_ERR( superludist2.Solve( true ) ) ;
#endif
#ifdef TEST_SPOOLES
} else if ( SparseSolver == SPOOLES ) {
SpoolesOO spooles( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spooles.SetTrans( transpose ) ;
spooles.Solve() ;
#endif
#ifdef HAVE_AMESOS_DSCPACK
} else if ( SparseSolver == DSCPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Dscpack dscpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( dscpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( dscpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_UMFPACK
} else if ( SparseSolver == UMFPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Umfpack umfpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( umfpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( umfpack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( umfpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_KLU
} else if ( SparseSolver == KLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Klu klu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( klu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( klu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( klu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( klu.NumericFactorization( ) );
EPETRA_CHK_ERR( klu.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARAKLETE
} else if ( SparseSolver == PARAKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Paraklete paraklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( paraklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( paraklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( paraklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( paraklete.NumericFactorization( ) );
EPETRA_CHK_ERR( paraklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SLUS
} else if ( SparseSolver == SuperLU ) {
Epetra_SLU superluserial( &Problem ) ;
EPETRA_CHK_ERR( superluserial.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superluserial.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superluserial.NumericFactorization( ) );
EPETRA_CHK_ERR( superluserial.Solve( ) );
#endif
#ifdef HAVE_AMESOS_LAPACK
} else if ( SparseSolver == LAPACK ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Lapack lapack( Problem ) ;
EPETRA_CHK_ERR( lapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( lapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( lapack.NumericFactorization( ) );
EPETRA_CHK_ERR( lapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_TAUCS
} else if ( SparseSolver == TAUCS ) {
Teuchos::ParameterList ParamList ;
Amesos_Taucs taucs( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( taucs.SetParameters( ParamList ) );
EPETRA_CHK_ERR( taucs.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( taucs.SymbolicFactorization( ) );
EPETRA_CHK_ERR( taucs.NumericFactorization( ) );
EPETRA_CHK_ERR( taucs.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARDISO
} else if ( SparseSolver == PARDISO ) {
Teuchos::ParameterList ParamList ;
Amesos_Pardiso pardiso( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( pardiso.SetParameters( ParamList ) );
EPETRA_CHK_ERR( pardiso.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( pardiso.SymbolicFactorization( ) );
EPETRA_CHK_ERR( pardiso.NumericFactorization( ) );
EPETRA_CHK_ERR( pardiso.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARKLETE
} else if ( SparseSolver == PARKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Parklete parklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( parklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( parklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( parklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( parklete.NumericFactorization( ) );
EPETRA_CHK_ERR( parklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_MUMPS
} else if ( SparseSolver == MUMPS ) {
Teuchos::ParameterList ParamList ;
Amesos_Mumps mumps( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( mumps.SetParameters( ParamList ) );
EPETRA_CHK_ERR( mumps.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( mumps.SymbolicFactorization( ) );
EPETRA_CHK_ERR( mumps.NumericFactorization( ) );
EPETRA_CHK_ERR( mumps.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SCALAPACK
} else if ( SparseSolver == SCALAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Scalapack scalapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( scalapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( scalapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( scalapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( scalapack.NumericFactorization( ) );
EPETRA_CHK_ERR( scalapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLUDIST
} else if ( SparseSolver == SUPERLUDIST ) {
Teuchos::ParameterList ParamList ;
Amesos_Superludist superludist( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( superludist.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superludist.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superludist.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superludist.NumericFactorization( ) );
EPETRA_CHK_ERR( superludist.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLU
} else if ( SparseSolver == SUPERLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Superlu superlu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( superlu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superlu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superlu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superlu.NumericFactorization( ) );
EPETRA_CHK_ERR( superlu.Solve( ) );
#endif
#ifdef TEST_SPOOLESSERIAL
} else if ( SparseSolver == SPOOLESSERIAL ) {
SpoolesserialOO spoolesserial( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spoolesserial.Solve() ;
#endif
} else {
SparseDirectTimingVars::log_file << "Solver not implemented yet" << std::endl ;
std::cerr << "\n\n#################### Requested solver not available (Or not tested with blocked RHS) on this platform #####################\n" << std::endl ;
}
SparseDirectTimingVars::SS_Result.Set_Total_Time( TotalTime.ElapsedTime() );
// SparseDirectTimingVars::SS_Result.Set_First_Time( 0.0 );
// SparseDirectTimingVars::SS_Result.Set_Middle_Time( 0.0 );
// SparseDirectTimingVars::SS_Result.Set_Last_Time( 0.0 );
//
// Compute the error = norm(xcomp - xexact )
//
std::vector <double> error(numsolves) ;
double max_error = 0.0;
passresid->Update(1.0, *passx, -1.0, *passxexact, 0.0);
passresid->Norm2(&error[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( error[i] > max_error ) max_error = error[i] ;
SparseDirectTimingVars::SS_Result.Set_Error(max_error) ;
// passxexact->Norm2(&error[0] ) ;
// passx->Norm2(&error ) ;
//
// Compute the residual = norm(Ax - b)
//
std::vector <double> residual(numsolves) ;
passtmp->PutScalar(0.0);
passA->Multiply( transpose, *passx, *passtmp);
passresid->Update(1.0, *passtmp, -1.0, *passb, 0.0);
// passresid->Update(1.0, *passtmp, -1.0, CopyB, 0.0);
passresid->Norm2(&residual[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( residual[i] > max_resid ) max_resid = residual[i] ;
SparseDirectTimingVars::SS_Result.Set_Residual(max_resid) ;
std::vector <double> bnorm(numsolves);
passb->Norm2( &bnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Bnorm(bnorm[0]) ;
std::vector <double> xnorm(numsolves);
passx->Norm2( &xnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Xnorm(xnorm[0]) ;
if ( false && iam == 0 ) {
std::cout << " Amesos_TestMutliSolver.cpp " << std::endl ;
for ( int i = 0 ; i< numsolves && i < 10 ; i++ ) {
std::cout << "i=" << i
<< " error = " << error[i]
<< " xnorm = " << xnorm[i]
<< " residual = " << residual[i]
<< " bnorm = " << bnorm[i]
<< std::endl ;
}
std::cout << std::endl << " max_resid = " << max_resid ;
std::cout << " max_error = " << max_error << std::endl ;
std::cout << " Get_residual() again = " << SparseDirectTimingVars::SS_Result.Get_Residual() << std::endl ;
}
}
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
delete map_;
Comm.Barrier();
return 0 ;
}
// ============================================================================
void EpetraExt::XMLWriter::
Write(const std::string& Label, const Epetra_Map& Map)
{
TEUCHOS_TEST_FOR_EXCEPTION(IsOpen_ == false, std::logic_error,
"No file has been opened");
int NumGlobalElements = Map.NumGlobalElements();
int* MyGlobalElements = Map.MyGlobalElements();
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "<Map Label=\"" << Label
<< "\" NumElements=\"" << NumGlobalElements << '"'
<< " IndexBase=\"" << Map.IndexBase() << '"'
<< " NumProc=\"" << Comm_.NumProc() << '"';
of.close();
}
for (int iproc = 0; iproc < Comm_.NumProc(); ++iproc)
{
if (iproc == Comm_.MyPID())
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << " ElementsOnProc" << iproc << "=\"" << Map.NumMyElements() << '"';
of.close();
}
Comm_.Barrier();
}
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << '>' << std::endl;
of.close();
}
for (int iproc = 0; iproc < Comm_.NumProc(); iproc++)
{
if (iproc == Comm_.MyPID())
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "<Proc ID=\"" << Comm_.MyPID() << "\">" << std::endl;
for (int i = 0; i < Map.NumMyElements(); ++i)
{
of << MyGlobalElements[i] << std::endl;
}
of << "</Proc>" << std::endl;
of.close();
}
Comm_.Barrier();
}
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "</Map>" << std::endl;
of.close();
}
}
//
// Diagonal: 0=no change, 1=eliminate entry
// from the map for the largest row element in process 0
// 2=add diagonal entries to the matrix, with a zero value
// (assume row map contains all diagonal entries).
//
// ReindexRowMap:
// 0=no change, 1= add 2 (still contiguous), 2=non-contiguous
//
// ReindexColMap
// 0=same as RowMap, 1=add 4 - Different From RowMap, but contiguous)
//
// RangeMap:
// 0=no change, 1=serial map, 2=bizarre distribution, 3=replicated map
//
// DomainMap:
// 0=no change, 1=serial map, 2=bizarre distribution, 3=replicated map
//
RCP<Epetra_CrsMatrix> NewMatNewMap(Epetra_CrsMatrix& In,
int Diagonal,
int ReindexRowMap,
int ReindexColMap,
int RangeMapType,
int DomainMapType
)
{
//
// If we are making no change, return the original matrix (which has a linear map)
//
#if 0
std::cout << __FILE__ << "::" << __LINE__ << " "
<< Diagonal << " "
<< ReindexRowMap << " "
<< ReindexColMap << " "
<< RangeMapType << " "
<< DomainMapType << " " << std::endl ;
#endif
if ( Diagonal + ReindexRowMap + ReindexColMap + RangeMapType + DomainMapType == 0 ) {
RCP<Epetra_CrsMatrix> ReturnOrig = rcp( &In, false );
return ReturnOrig ;
}
//
// Diagonal==2 is used for a different purpose -
// Making sure that the diagonal of the matrix is non-empty.
// Note: The diagonal must exist in In.RowMap().
//
if ( Diagonal == 2 ) {
assert( ReindexRowMap==0 && ReindexColMap == 0 ) ;
}
int (*RowPermute)(int in) = 0;
int (*ColPermute)(int in) = 0;
assert( Diagonal >= 0 && Diagonal <= 2 );
assert( ReindexRowMap>=0 && ReindexRowMap<=2 );
assert( ReindexColMap>=0 && ReindexColMap<=1 );
assert( RangeMapType>=0 && RangeMapType<=3 );
assert( DomainMapType>=0 && DomainMapType<=3 );
Epetra_Map DomainMap = In.DomainMap();
Epetra_Map RangeMap = In.RangeMap();
Epetra_Map ColMap = In.ColMap();
Epetra_Map RowMap = In.RowMap();
int NumMyRowElements = RowMap.NumMyElements();
int NumMyColElements = ColMap.NumMyElements();
int NumMyRangeElements = RangeMap.NumMyElements();
int NumMyDomainElements = DomainMap.NumMyElements();
int NumGlobalRowElements = RowMap.NumGlobalElements();
int NumGlobalColElements = ColMap.NumGlobalElements();
int NumGlobalRangeElements = RangeMap.NumGlobalElements();
int NumGlobalDomainElements = DomainMap.NumGlobalElements();
assert( NumGlobalRangeElements == NumGlobalDomainElements ) ;
std::vector<int> MyGlobalRowElements( NumMyRowElements ) ;
std::vector<int> NumEntriesPerRow( NumMyRowElements ) ;
std::vector<int> MyPermutedGlobalRowElements( NumMyRowElements ) ;
std::vector<int> MyGlobalColElements( NumMyColElements ) ;
std::vector<int> MyPermutedGlobalColElements( NumMyColElements ) ; // Used to create the column map
std::vector<int> MyPermutedGlobalColElementTable( NumMyColElements ) ; // To convert local indices to global
std::vector<int> MyGlobalRangeElements( NumMyRangeElements ) ;
std::vector<int> MyPermutedGlobalRangeElements( NumMyRangeElements ) ;
std::vector<int> MyGlobalDomainElements( NumMyDomainElements ) ;
std::vector<int> MyPermutedGlobalDomainElements( NumMyDomainElements ) ;
RowMap.MyGlobalElements(&MyGlobalRowElements[0]);
ColMap.MyGlobalElements(&MyGlobalColElements[0]);
RangeMap.MyGlobalElements(&MyGlobalRangeElements[0]);
DomainMap.MyGlobalElements(&MyGlobalDomainElements[0]);
switch( ReindexRowMap ) {
case 0:
RowPermute = &NoPermute ;
break;
case 1:
RowPermute = &SmallRowPermute ;
break;
case 2:
RowPermute = BigRowPermute ;
break;
}
switch( ReindexColMap ) {
case 0:
ColPermute = RowPermute ;
break;
case 1:
ColPermute = &SmallColPermute ;
break;
}
//
// Create Serial Range and Domain Maps based on the permuted indexing
//
int nlocal = 0;
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements;
std::vector<int> AllIDs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDs[i] = (*RowPermute)( i ) ;
Epetra_Map SerialRangeMap( -1, nlocal, &AllIDs[0], 0, In.Comm());
std::vector<int> AllIDBs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDBs[i] = (*ColPermute)( i ) ;
Epetra_Map SerialDomainMap( -1, nlocal, &AllIDBs[0], 0, In.Comm());
//
// Create Bizarre Range and Domain Maps based on the permuted indexing
// These are nearly serial, having all but one element on process 0
// The goal here is to make sure that we can use Domain and Range maps
// that are neither serial, nor distributed in the normal manner.
//
std::vector<int> AllIDCs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDCs[i] = (*ColPermute)( i ) ;
if ( In.Comm().NumProc() > 1 ) {
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements-1;
if (In.Comm().MyPID()==1) {
nlocal = 1;
AllIDCs[0] = (*ColPermute)( NumGlobalRangeElements - 1 );
}
}
int iam = In.Comm().MyPID();
Epetra_Map BizarreDomainMap( -1, nlocal, &AllIDCs[0], 0, In.Comm());
std::vector<int> AllIDDs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDDs[i] = (*RowPermute)( i ) ;
if ( In.Comm().NumProc() > 1 ) {
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements-1;
if (In.Comm().MyPID()==1) {
nlocal = 1;
AllIDDs[0] = (*RowPermute)( NumGlobalRangeElements -1 ) ;
}
}
Epetra_Map BizarreRangeMap( -1, nlocal, &AllIDDs[0], 0, In.Comm());
//
// Compute the column map
//
// If Diagonal==1, remove the column corresponding to the last row owned
// by process 0. Removing this column from a tridiagonal matrix, leaves
// a disconnected, but non-singular matrix.
//
int NumMyColElementsOut = 0 ;
int NumGlobalColElementsOut ;
if ( Diagonal == 1 )
NumGlobalColElementsOut = NumGlobalColElements-1;
else
NumGlobalColElementsOut = NumGlobalColElements;
if ( Diagonal == 1 && iam==0 ) {
for ( int i=0; i < NumMyColElements ; i++ ) {
if ( MyGlobalColElements[i] != MyGlobalRowElements[NumMyRowElements-1] ) {
MyPermutedGlobalColElements[NumMyColElementsOut++] =
(*ColPermute)( MyGlobalColElements[i] ) ;
}
}
assert( NumMyColElementsOut == NumMyColElements-1 );
} else {
for ( int i=0; i < NumMyColElements ; i++ )
MyPermutedGlobalColElements[i] =
(*ColPermute)( MyGlobalColElements[i] ) ;
NumMyColElementsOut = NumMyColElements ;
if ( Diagonal == 2 ) {
// For each row, make sure that the column map has this row in it,
// if it doesn't, add it to the column map.
// Note: MyPermutedGlobalColElements == MyGlobalColElements when
// Diagonal==2 because ( Diagonal == 2 ) implies:
// ReindexRowMap==0 && ReindexColMap == 0 - see assert above
for ( int i=0; i < NumMyRowElements ; i++ ) {
bool MissingDiagonal = true;
for ( int j=0; j < NumMyColElements; j++ ) {
if ( MyGlobalRowElements[i] == MyGlobalColElements[j] ) {
MissingDiagonal = false;
}
}
if ( MissingDiagonal ) {
MyPermutedGlobalColElements.resize(NumMyColElements+1);
MyPermutedGlobalColElements[NumMyColElementsOut] = MyGlobalRowElements[i];
NumMyColElementsOut++;
}
}
In.Comm().SumAll(&NumMyColElementsOut,&NumGlobalColElementsOut,1);
}
}
//
// These tables are used both as the permutation tables and to create the maps.
//
for ( int i=0; i < NumMyColElements ; i++ )
MyPermutedGlobalColElementTable[i] =
(*ColPermute)( MyGlobalColElements[i] ) ;
for ( int i=0; i < NumMyRowElements ; i++ )
MyPermutedGlobalRowElements[i] =
(*RowPermute)( MyGlobalRowElements[i] ) ;
for ( int i=0; i < NumMyRangeElements ; i++ )
MyPermutedGlobalRangeElements[i] =
(*RowPermute)( MyGlobalRangeElements[i] ) ;
for ( int i=0; i < NumMyDomainElements ; i++ )
MyPermutedGlobalDomainElements[i] =
(*ColPermute)( MyGlobalDomainElements[i] ) ;
RCP<Epetra_Map> PermutedRowMap =
rcp( new Epetra_Map( NumGlobalRowElements, NumMyRowElements,
&MyPermutedGlobalRowElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedColMap =
rcp( new Epetra_Map( NumGlobalColElementsOut, NumMyColElementsOut,
&MyPermutedGlobalColElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedRangeMap =
rcp( new Epetra_Map( NumGlobalRangeElements, NumMyRangeElements,
&MyPermutedGlobalRangeElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedDomainMap =
rcp( new Epetra_Map( NumGlobalDomainElements, NumMyDomainElements,
&MyPermutedGlobalDomainElements[0], 0, In.Comm() ) );
//
// These vectors are filled and then passed to InsertGlobalValues
//
std::vector<int> ThisRowIndices( In.MaxNumEntries() );
std::vector<double> ThisRowValues( In.MaxNumEntries() );
std::vector<int> PermutedGlobalColIndices( In.MaxNumEntries() );
//std::cout << __FILE__ << "::" <<__LINE__ << std::endl ;
RCP<Epetra_CrsMatrix> Out =
rcp( new Epetra_CrsMatrix( Copy, *PermutedRowMap, *PermutedColMap, 0 ) );
for (int i=0; i<NumMyRowElements; i++)
{
int NumIndicesThisRow = 0;
assert( In.ExtractMyRowCopy( i,
In.MaxNumEntries(),
NumIndicesThisRow,
&ThisRowValues[0],
&ThisRowIndices[0] ) == 0 ) ;
for (int j = 0 ; j < NumIndicesThisRow ; j++ )
{
PermutedGlobalColIndices[j] = MyPermutedGlobalColElementTable[ ThisRowIndices[j] ] ;
}
bool MissingDiagonal = false;
if ( Diagonal==2 ) {
//
assert( MyGlobalRowElements[i] == MyPermutedGlobalRowElements[i] );
MissingDiagonal = true;
for( int j =0 ; j < NumIndicesThisRow ; j++ ) {
if ( PermutedGlobalColIndices[j] == MyPermutedGlobalRowElements[i] ) {
MissingDiagonal = false ;
}
}
#if 0
std::cout << __FILE__ << "::" << __LINE__
<< " i = " << i
<< " MyPermutedGlobalRowElements[i] = " << MyPermutedGlobalRowElements[i]
<< " MissingDiagonal = " << MissingDiagonal << std::endl ;
#endif
}
if ( MissingDiagonal ) {
ThisRowValues.resize(NumIndicesThisRow+1) ;
ThisRowValues[NumIndicesThisRow] = 0.0;
PermutedGlobalColIndices.resize(NumIndicesThisRow+1);
PermutedGlobalColIndices[NumIndicesThisRow] = MyPermutedGlobalRowElements[i] ;
#if 0
std::cout << __FILE__ << "::" << __LINE__
<< " i = " << i
<< "NumIndicesThisRow = " << NumIndicesThisRow
<< "ThisRowValues[NumIndicesThisRow = " << ThisRowValues[NumIndicesThisRow]
<< " PermutedGlobalColIndices[NumIndcesThisRow] = " << PermutedGlobalColIndices[NumIndicesThisRow]
<< std::endl ;
#endif
NumIndicesThisRow++ ;
}
assert( Out->InsertGlobalValues( MyPermutedGlobalRowElements[i],
NumIndicesThisRow,
&ThisRowValues[0],
&PermutedGlobalColIndices[0] ) >= 0 );
}
//
Epetra_LocalMap ReplicatedMap( NumGlobalRangeElements, 0, In.Comm() );
RCP<Epetra_Map> OutRangeMap ;
RCP<Epetra_Map> OutDomainMap ;
switch( RangeMapType ) {
case 0:
OutRangeMap = PermutedRangeMap ;
break;
case 1:
OutRangeMap = rcp(&SerialRangeMap, false);
break;
case 2:
OutRangeMap = rcp(&BizarreRangeMap, false);
break;
case 3:
OutRangeMap = rcp(&ReplicatedMap, false);
break;
}
// switch( DomainMapType ) {
switch( DomainMapType ) {
case 0:
OutDomainMap = PermutedDomainMap ;
break;
case 1:
OutDomainMap = rcp(&SerialDomainMap, false);
break;
case 2:
OutDomainMap = rcp(&BizarreDomainMap, false);
break;
case 3:
OutDomainMap = rcp(&ReplicatedMap, false);
break;
}
#if 0
assert(Out->FillComplete( *PermutedDomainMap, *PermutedRangeMap )==0);
#else
assert(Out->FillComplete( *OutDomainMap, *OutRangeMap )==0);
#endif
#if 0
std::cout << __FILE__ << "::" << __LINE__ << std::endl ;
Out->Print( std::cout ) ;
#endif
return Out;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
cout << Comm << endl;
int MyPID = Comm.MyPID();
bool verbose = false;
bool verbose1 = true;
if (MyPID==0) verbose = true;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename [level_fill [level_overlap [absolute_threshold [ relative_threshold]]]]" << endl
<< "where:" << endl
<< "HB_filename - filename and path of a Harwell-Boeing data set" << endl
<< "level_fill - The amount of fill to use for ILU(k) preconditioner (default 0)" << endl
<< "level_overlap - The amount of overlap used for overlapping Schwarz subdomains (default 0)" << endl
<< "absolute_threshold - The minimum value to place on the diagonal prior to factorization (default 0.0)" << endl
<< "relative_threshold - The relative amount to perturb the diagonal prior to factorization (default 1.0)" << endl << endl
<< "To specify a non-default value for one of these parameters, you must specify all" << endl
<< " preceding values but not any subsequent parameters. Example:" << endl
<< "ifpackHpcSerialMsr.exe mymatrix.hpc 1 - loads mymatrix.hpc, uses level fill of one, all other values are defaults" << endl
<< endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
bool smallProblem = false;
if (A.RowMap().NumGlobalElements()<100) smallProblem = true;
if (smallProblem)
cout << "Original Matrix = " << endl << A << endl;
x.PutScalar(0.0);
Epetra_LinearProblem FullProblem(&A, &x, &b);
double normb, norma;
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
Epetra_Time ReductionTimer(Comm);
Epetra_CrsSingletonFilter SingletonFilter;
Comm.Barrier();
double reduceInitTime = ReductionTimer.ElapsedTime();
SingletonFilter.Analyze(&A);
Comm.Barrier();
double reduceAnalyzeTime = ReductionTimer.ElapsedTime() - reduceInitTime;
if (SingletonFilter.SingletonsDetected())
cout << "Singletons found" << endl;
else {
cout << "Singletons not found" << endl;
exit(1);
}
SingletonFilter.ConstructReducedProblem(&FullProblem);
Comm.Barrier();
double reduceConstructTime = ReductionTimer.ElapsedTime() - reduceInitTime;
double totalReduceTime = ReductionTimer.ElapsedTime();
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Reduction init time (sec) = " << reduceInitTime<< endl
<< " Reduction Analyze time (sec) = " << reduceAnalyzeTime << endl
<< " Construct Reduced Problem time (sec) = " << reduceConstructTime << endl
<< " Reduction Total time (sec) = " << totalReduceTime << endl<< endl;
Statistics(SingletonFilter);
Epetra_LinearProblem * ReducedProblem = SingletonFilter.ReducedProblem();
Epetra_CrsMatrix * Ap = dynamic_cast<Epetra_CrsMatrix *>(ReducedProblem->GetMatrix());
Epetra_Vector * bp = (*ReducedProblem->GetRHS())(0);
Epetra_Vector * xp = (*ReducedProblem->GetLHS())(0);
if (smallProblem)
cout << " Reduced Matrix = " << endl << *Ap << endl
<< " LHS before sol = " << endl << *xp << endl
<< " RHS = " << endl << *bp << endl;
// Construct ILU preconditioner
double elapsed_time, total_flops, MFLOPs;
Epetra_Time timer(Comm);
int LevelFill = 0;
if (argc > 2) LevelFill = atoi(argv[2]);
if (verbose) cout << "Using Level Fill = " << LevelFill << endl;
int Overlap = 0;
if (argc > 3) Overlap = atoi(argv[3]);
if (verbose) cout << "Using Level Overlap = " << Overlap << endl;
double Athresh = 0.0;
if (argc > 4) Athresh = atof(argv[4]);
if (verbose) cout << "Using Absolute Threshold Value of = " << Athresh << endl;
double Rthresh = 1.0;
if (argc > 5) Rthresh = atof(argv[5]);
if (verbose) cout << "Using Relative Threshold Value of = " << Rthresh << endl;
Ifpack_IlukGraph * IlukGraph = 0;
Ifpack_CrsRiluk * ILUK = 0;
if (LevelFill>-1) {
elapsed_time = timer.ElapsedTime();
IlukGraph = new Ifpack_IlukGraph(Ap->Graph(), LevelFill, Overlap);
assert(IlukGraph->ConstructFilledGraph()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
if (verbose) cout << "Time to construct ILUK graph = " << elapsed_time << endl;
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
ILUK = new Ifpack_CrsRiluk(*IlukGraph);
ILUK->SetFlopCounter(fact_counter);
ILUK->SetAbsoluteThreshold(Athresh);
ILUK->SetRelativeThreshold(Rthresh);
//assert(ILUK->InitValues()==0);
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
//cout << *ILUK << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
int Maxiter = 100;
double Tolerance = 1.0E-8;
Epetra_Flops counter;
Ap->SetFlopCounter(counter);
xp->SetFlopCounter(*Ap);
bp->SetFlopCounter(*Ap);
if (ILUK!=0) ILUK->SetFlopCounter(*Ap);
elapsed_time = timer.ElapsedTime();
double normreducedb, normreduceda;
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
Epetra_Vector resid(x);
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
double normx, normxexact;
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (verbose)
cout << "********************************************************" << endl
<< " Solving again with 2*Ax=2*b" << endl
<< "********************************************************" << endl;
A.Scale(1.0); // A = 2*A
b.Scale(1.0); // b = 2*b
x.PutScalar(0.0);
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
double updateReducedProblemTime = ReductionTimer.ElapsedTime();
SingletonFilter.UpdateReducedProblem(&FullProblem);
Comm.Barrier();
updateReducedProblemTime = ReductionTimer.ElapsedTime() - updateReducedProblemTime;
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Update Reduced Problem time (sec) = " << updateReducedProblemTime<< endl
<< "****************************************************" << endl;
Statistics(SingletonFilter);
if (LevelFill>-1) {
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (ILUK!=0) delete ILUK;
if (IlukGraph!=0) delete IlukGraph;
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
int main(int argc, char *argv[])
{
int ierr = 0;
double elapsed_time;
double total_flops;
double MFLOPs;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm comm;
#endif
bool verbose = false;
bool summary = false;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='v') verbose = true;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='s') summary = true;
if(argc < 6) {
cerr << "Usage: " << argv[0]
<< " NumNodesX NumNodesY NumProcX NumProcY NumPoints [-v|-s]" << endl
<< "where:" << endl
<< "NumNodesX - Number of mesh nodes in X direction per processor" << endl
<< "NumNodesY - Number of mesh nodes in Y direction per processor" << endl
<< "NumProcX - Number of processors to use in X direction" << endl
<< "NumProcY - Number of processors to use in Y direction" << endl
<< "NumPoints - Number of points to use in stencil (5, 9 or 25 only)" << endl
<< "-v|-s - (Optional) Run in verbose mode if -v present or summary mode if -s present" << endl
<< " NOTES: NumProcX*NumProcY must equal the number of processors used to run the problem." << endl << endl
<< " Serial example:" << endl
<< argv[0] << " 16 12 1 1 25 -v" << endl
<< " Run this program in verbose mode on 1 processor using a 16 X 12 grid with a 25 point stencil."<< endl <<endl
<< " MPI example:" << endl
<< "mpirun -np 32 " << argv[0] << " 10 12 4 8 9 -v" << endl
<< " Run this program in verbose mode on 32 processors putting a 10 X 12 subgrid on each processor using 4 processors "<< endl
<< " in the X direction and 8 in the Y direction. Total grid size is 40 points in X and 96 in Y with a 9 point stencil."<< endl
<< endl;
return(1);
}
//char tmp;
//if (comm.MyPID()==0) cout << "Press any key to continue..."<< endl;
//if (comm.MyPID()==0) cin >> tmp;
//comm.Barrier();
comm.SetTracebackMode(0); // This should shut down any error traceback reporting
if (verbose && comm.MyPID()==0)
cout << Epetra_Version() << endl << endl;
if (summary && comm.MyPID()==0) {
if (comm.NumProc()==1)
cout << Epetra_Version() << endl << endl;
else
cout << endl << endl; // Print two blank line to keep output columns lined up
}
if (verbose) cout << comm <<endl;
// Redefine verbose to only print on PE 0
if (verbose && comm.MyPID()!=0) verbose = false;
if (summary && comm.MyPID()!=0) summary = false;
int numNodesX = atoi(argv[1]);
int numNodesY = atoi(argv[2]);
int numProcsX = atoi(argv[3]);
int numProcsY = atoi(argv[4]);
int numPoints = atoi(argv[5]);
if (verbose || (summary && comm.NumProc()==1)) {
cout << " Number of local nodes in X direction = " << numNodesX << endl
<< " Number of local nodes in Y direction = " << numNodesY << endl
<< " Number of global nodes in X direction = " << numNodesX*numProcsX << endl
<< " Number of global nodes in Y direction = " << numNodesY*numProcsY << endl
<< " Number of local nonzero entries = " << numNodesX*numNodesY*numPoints << endl
<< " Number of global nonzero entries = " << numNodesX*numNodesY*numPoints*numProcsX*numProcsY << endl
<< " Number of Processors in X direction = " << numProcsX << endl
<< " Number of Processors in Y direction = " << numProcsY << endl
<< " Number of Points in stencil = " << numPoints << endl << endl;
}
// Print blank line to keep output columns lined up
if (summary && comm.NumProc()>1)
cout << endl << endl << endl << endl << endl << endl << endl << endl<< endl << endl;
if (numProcsX*numProcsY!=comm.NumProc()) {
cerr << "Number of processors = " << comm.NumProc() << endl
<< " is not the product of " << numProcsX << " and " << numProcsY << endl << endl;
return(1);
}
if (numPoints!=5 && numPoints!=9 && numPoints!=25) {
cerr << "Number of points specified = " << numPoints << endl
<< " is not 5, 9, 25" << endl << endl;
return(1);
}
if (numNodesX*numNodesY<=0) {
cerr << "Product of number of nodes is <= zero" << endl << endl;
return(1);
}
Epetra_IntSerialDenseVector Xoff, XLoff, XUoff;
Epetra_IntSerialDenseVector Yoff, YLoff, YUoff;
if (numPoints==5) {
// Generate a 5-point 2D Finite Difference matrix
Xoff.Size(5);
Yoff.Size(5);
Xoff[0] = -1; Xoff[1] = 1; Xoff[2] = 0; Xoff[3] = 0; Xoff[4] = 0;
Yoff[0] = 0; Yoff[1] = 0; Yoff[2] = 0; Yoff[3] = -1; Yoff[4] = 1;
// Generate a 2-point 2D Lower triangular Finite Difference matrix
XLoff.Size(2);
YLoff.Size(2);
XLoff[0] = -1; XLoff[1] = 0;
YLoff[0] = 0; YLoff[1] = -1;
// Generate a 3-point 2D upper triangular Finite Difference matrix
XUoff.Size(3);
YUoff.Size(3);
XUoff[0] = 0; XUoff[1] = 1; XUoff[2] = 0;
YUoff[0] = 0; YUoff[1] = 0; YUoff[2] = 1;
}
else if (numPoints==9) {
// Generate a 9-point 2D Finite Difference matrix
Xoff.Size(9);
Yoff.Size(9);
Xoff[0] = -1; Xoff[1] = 0; Xoff[2] = 1;
Yoff[0] = -1; Yoff[1] = -1; Yoff[2] = -1;
Xoff[3] = -1; Xoff[4] = 0; Xoff[5] = 1;
Yoff[3] = 0; Yoff[4] = 0; Yoff[5] = 0;
Xoff[6] = -1; Xoff[7] = 0; Xoff[8] = 1;
Yoff[6] = 1; Yoff[7] = 1; Yoff[8] = 1;
// Generate a 5-point lower triangular 2D Finite Difference matrix
XLoff.Size(5);
YLoff.Size(5);
XLoff[0] = -1; XLoff[1] = 0; Xoff[2] = 1;
YLoff[0] = -1; YLoff[1] = -1; Yoff[2] = -1;
XLoff[3] = -1; XLoff[4] = 0;
YLoff[3] = 0; YLoff[4] = 0;
// Generate a 4-point upper triangular 2D Finite Difference matrix
XUoff.Size(4);
YUoff.Size(4);
XUoff[0] = 1;
YUoff[0] = 0;
XUoff[1] = -1; XUoff[2] = 0; XUoff[3] = 1;
YUoff[1] = 1; YUoff[2] = 1; YUoff[3] = 1;
}
else {
// Generate a 25-point 2D Finite Difference matrix
Xoff.Size(25);
Yoff.Size(25);
int xi = 0, yi = 0;
int xo = -2, yo = -2;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
// Generate a 13-point lower triangular 2D Finite Difference matrix
XLoff.Size(13);
YLoff.Size(13);
xi = 0, yi = 0;
xo = -2, yo = -2;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
// Generate a 13-point upper triangular 2D Finite Difference matrix
XUoff.Size(13);
YUoff.Size(13);
xi = 0, yi = 0;
xo = 0, yo = 0;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
}
Epetra_Map * map;
Epetra_Map * mapL;
Epetra_Map * mapU;
Epetra_CrsMatrix * A;
Epetra_CrsMatrix * L;
Epetra_CrsMatrix * U;
Epetra_MultiVector * b;
Epetra_MultiVector * bt;
Epetra_MultiVector * xexact;
Epetra_MultiVector * bL;
Epetra_MultiVector * btL;
Epetra_MultiVector * xexactL;
Epetra_MultiVector * bU;
Epetra_MultiVector * btU;
Epetra_MultiVector * xexactU;
Epetra_SerialDenseVector resvec(0);
//Timings
Epetra_Flops flopcounter;
Epetra_Time timer(comm);
#ifdef EPETRA_VERY_SHORT_PERFTEST
int jstop = 1;
#elif EPETRA_SHORT_PERFTEST
int jstop = 1;
#else
int jstop = 2;
#endif
for (int j=0; j<jstop; j++) {
for (int k=1; k<17; k++) {
#ifdef EPETRA_VERY_SHORT_PERFTEST
if (k<3 || (k%4==0 && k<9)) {
#elif EPETRA_SHORT_PERFTEST
if (k<6 || k%4==0) {
#else
if (k<7 || k%2==0) {
#endif
int nrhs=k;
if (verbose) cout << "\n*************** Results for " << nrhs << " RHS with ";
bool StaticProfile = (j!=0);
if (verbose)
if (StaticProfile) cout << " static profile\n";
else cout << " dynamic profile\n";
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
Xoff.Values(), Yoff.Values(), nrhs, comm, verbose, summary,
map, A, b, bt, xexact, StaticProfile, false);
#ifdef EPETRA_HAVE_JADMATRIX
timer.ResetStartTime();
Epetra_JadMatrix JA(*A);
elapsed_time = timer.ElapsedTime();
if (verbose) cout << "Time to create Jagged diagonal matrix = " << elapsed_time << endl;
//cout << "A = " << *A << endl;
//cout << "JA = " << JA << endl;
runJadMatrixTests(&JA, b, bt, xexact, StaticProfile, verbose, summary);
#endif
runMatrixTests(A, b, bt, xexact, StaticProfile, verbose, summary);
delete A;
delete b;
delete bt;
delete xexact;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XLoff.Length(),
XLoff.Values(), YLoff.Values(), nrhs, comm, verbose, summary,
mapL, L, bL, btL, xexactL, StaticProfile, true);
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XUoff.Length(),
XUoff.Values(), YUoff.Values(), nrhs, comm, verbose, summary,
mapU, U, bU, btU, xexactU, StaticProfile, true);
runLUMatrixTests(L, bL, btL, xexactL, U, bU, btU, xexactU, StaticProfile, verbose, summary);
delete L;
delete bL;
delete btL;
delete xexactL;
delete mapL;
delete U;
delete bU;
delete btU;
delete xexactU;
delete mapU;
Epetra_MultiVector q(*map, nrhs);
Epetra_MultiVector z(q);
Epetra_MultiVector r(q);
delete map;
q.SetFlopCounter(flopcounter);
z.SetFlopCounter(q);
r.SetFlopCounter(q);
resvec.Resize(nrhs);
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 norms
for( int i = 0; i < 10; ++i )
q.Norm2( resvec.Values() );
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "\nTotal MFLOPs for 10 Norm2's= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Norm2" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Dot(z, resvec.Values());
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Dot's = " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "DotProd" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Update(1.0, z, 1.0, r, 0.0);
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Updates= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Update" << '\t';
cout << MFLOPs << endl;
}
}
}
}
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return ierr ;
}
// Constructs a 2D PDE finite difference matrix using the list of x and y offsets.
//
// nx (In) - number of grid points in x direction
// ny (In) - number of grid points in y direction
// The total number of equations will be nx*ny ordered such that the x direction changes
// most rapidly:
// First equation is at point (0,0)
// Second at (1,0)
// ...
// nx equation at (nx-1,0)
// nx+1st equation at (0,1)
// numPoints (In) - number of points in finite difference stencil
// xoff (In) - stencil offsets in x direction (of length numPoints)
// yoff (In) - stencil offsets in y direction (of length numPoints)
// A standard 5-point finite difference stencil would be described as:
// numPoints = 5
// xoff = [-1, 1, 0, 0, 0]
// yoff = [ 0, 0, 0, -1, 1]
// nrhs - Number of rhs to generate. (First interface produces vectors, so nrhs is not needed
// comm (In) - an Epetra_Comm object describing the parallel machine (numProcs and my proc ID)
// map (Out) - Epetra_Map describing distribution of matrix and vectors/multivectors
// A (Out) - Epetra_CrsMatrix constructed for nx by ny grid using prescribed stencil
// Off-diagonal values are random between 0 and 1. If diagonal is part of stencil,
// diagonal will be slightly diag dominant.
// b (Out) - Generated RHS. Values satisfy b = A*xexact
// bt (Out) - Generated RHS. Values satisfy b = A'*xexact
// xexact (Out) - Generated exact solution to Ax = b and b' = A'xexact
// Note: Caller of this function is responsible for deleting all output objects.
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_Vector *& b,
Epetra_Vector *& bt,
Epetra_Vector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_MultiVector * b1, * bt1, * xexact1;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
xoff, yoff, 1, comm, verbose, summary,
map, A, b1, bt1, xexact1, StaticProfile, MakeLocalOnly);
b = dynamic_cast<Epetra_Vector *>(b1);
bt = dynamic_cast<Epetra_Vector *>(bt1);
xexact = dynamic_cast<Epetra_Vector *>(xexact1);
return;
}
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff, int nrhs,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_MultiVector *& b,
Epetra_MultiVector *& bt,
Epetra_MultiVector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_Time timer(comm);
// Determine my global IDs
int * myGlobalElements;
GenerateMyGlobalElements(numNodesX, numNodesY, numProcsX, numProcsY, comm.MyPID(), myGlobalElements);
int numMyEquations = numNodesX*numNodesY;
map = new Epetra_Map(-1, numMyEquations, myGlobalElements, 0, comm); // Create map with 2D block partitioning.
delete [] myGlobalElements;
int numGlobalEquations = map->NumGlobalElements();
int profile = 0; if (StaticProfile) profile = numPoints;
#ifdef EPETRA_HAVE_STATICPROFILE
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile, StaticProfile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile, StaticProfile); // Construct matrix
#else
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile); // Construct matrix
#endif
int * indices = new int[numPoints];
double * values = new double[numPoints];
double dnumPoints = (double) numPoints;
int nx = numNodesX*numProcsX;
for (int i=0; i<numMyEquations; i++) {
int rowID = map->GID(i);
int numIndices = 0;
for (int j=0; j<numPoints; j++) {
int colID = rowID + xoff[j] + nx*yoff[j]; // Compute column ID based on stencil offsets
if (colID>-1 && colID<numGlobalEquations) {
indices[numIndices] = colID;
double value = - ((double) rand())/ ((double) RAND_MAX);
if (colID==rowID)
values[numIndices++] = dnumPoints - value; // Make diagonal dominant
else
values[numIndices++] = value;
}
}
//cout << "Building row " << rowID << endl;
A->InsertGlobalValues(rowID, numIndices, values, indices);
}
delete [] indices;
delete [] values;
double insertTime = timer.ElapsedTime();
timer.ResetStartTime();
A->FillComplete(false);
double fillCompleteTime = timer.ElapsedTime();
if (verbose)
cout << "Time to insert matrix values = " << insertTime << endl
<< "Time to complete fill = " << fillCompleteTime << endl;
if (summary) {
if (comm.NumProc()==1) cout << "InsertTime" << '\t';
cout << insertTime << endl;
if (comm.NumProc()==1) cout << "FillCompleteTime" << '\t';
cout << fillCompleteTime << endl;
}
if (nrhs<=1) {
b = new Epetra_Vector(*map);
bt = new Epetra_Vector(*map);
xexact = new Epetra_Vector(*map);
}
else {
b = new Epetra_MultiVector(*map, nrhs);
bt = new Epetra_MultiVector(*map, nrhs);
xexact = new Epetra_MultiVector(*map, nrhs);
}
xexact->Random(); // Fill xexact with random values
A->Multiply(false, *xexact, *b);
A->Multiply(true, *xexact, *bt);
return;
}
// build maps to make other conversions
void buildSubMaps(const Epetra_Map & globalMap,const std::vector<int> & vars,const Epetra_Comm & comm,
std::vector<std::pair<int,Teuchos::RCP<Epetra_Map> > > & subMaps)
{
buildSubMaps(globalMap.NumGlobalElements(),globalMap.NumMyElements(),globalMap.MinMyGID(),
vars,comm,subMaps);
}
int MultiVectorTests(const Epetra_Map & Map, int NumVectors, bool verbose)
{
const Epetra_Comm & Comm = Map.Comm();
int ierr = 0, i, j;
/* get number of processors and the name of this processor */
int MyPID = Comm.MyPID();
// Construct FEVbrMatrix
if (verbose && MyPID==0) cout << "constructing Epetra_FEVbrMatrix" << endl;
//
//we'll set up a tri-diagonal matrix.
//
int numGlobalRows = Map.NumGlobalElements();
int minLocalRow = Map.MinMyGID();
int rowLengths = 3;
Epetra_FEVbrMatrix A(Copy, Map, rowLengths);
if (verbose && MyPID==0) {
cout << "calling A.InsertGlobalValues with 1-D data array"<<endl;
}
int numCols = 3;
int* ptIndices = new int[numCols];
for(int k=0; k<numCols; ++k) {
ptIndices[k] = minLocalRow+k;
}
double* values_1d = new double[numCols*numCols];
for(j=0; j<numCols*numCols; ++j) {
values_1d[j] = 3.0;
}
//For an extreme test, we'll have all processors sum into all rows.
int minGID = Map.MinAllGID();
//For now we're going to assume that there's just one point associated with
//each GID (element).
double* ptCoefs = new double[3];
{for(i=0; i<numGlobalRows; ++i) {
if (i>0 && i<numGlobalRows-1) {
ptIndices[0] = minGID+i-1;
ptIndices[1] = minGID+i;
ptIndices[2] = minGID+i+1;
ptCoefs[0] = -1.0;
ptCoefs[1] = 2.0;
ptCoefs[2] = -1.0;
numCols = 3;
}
else if (i == 0) {
ptIndices[0] = minGID+i;
ptIndices[1] = minGID+i+1;
ptIndices[2] = minGID+i+2;
ptCoefs[0] = 2.0;
ptCoefs[1] = -1.0;
ptCoefs[2] = -1.0;
numCols = 3;
}
else {
ptIndices[0] = minGID+i-2;
ptIndices[1] = minGID+i-1;
ptIndices[2] = minGID+i;
ptCoefs[0] = -1.0;
ptCoefs[1] = -1.0;
ptCoefs[2] = 2.0;
numCols = 3;
}
int row = minGID+i;
EPETRA_TEST_ERR( A.BeginInsertGlobalValues(row, rowLengths, ptIndices), ierr);
for(j=0; j<rowLengths; ++j) {
EPETRA_TEST_ERR( A.SubmitBlockEntry(&(ptCoefs[j]), 1, 1, 1), ierr);
}
EPETRA_TEST_ERR( A.EndSubmitEntries(), ierr);
}}
if (verbose&&MyPID==0) {
cout << "calling A.GlobalAssemble()" << endl;
}
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
if (verbose&&MyPID==0) {
cout << "after globalAssemble"<<endl;
}
if (verbose) {
A.Print(cout);
}
delete [] values_1d;
delete [] ptIndices;
delete [] ptCoefs;
return(ierr);
}
int main(int argc, char *argv[]) {
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
bool verbose = false;
if (MyPID==0) verbose = true;
// matrix downloaded from MatrixMarket
char FileName[] = "../HBMatrices/fidap005.rua";
Epetra_Map * readMap; // Pointers because of Trilinos_Util_ReadHb2Epetra
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(FileName, Comm, readMap, readA, readx,
readb, readxexact);
int NumGlobalElements = readMap->NumGlobalElements();
// Create uniform distributed map
Epetra_Map map(NumGlobalElements, 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
A.Export(*readA, exporter, Add);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
A.FillComplete();
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
// ============================ //
// Construct ILU preconditioner //
// ---------------------------- //
// modify those parameters
int LevelFill = 1;
double DropTol = 0.0;
double Condest;
Ifpack_CrsIct * ICT = NULL;
ICT = new Ifpack_CrsIct(A,DropTol,LevelFill);
// Init values from A
ICT->InitValues(A);
// compute the factors
ICT->Factor();
// and now estimate the condition number
ICT->Condest(false,Condest);
cout << Condest << endl;
if( Comm.MyPID() == 0 ) {
cout << "Condition number estimate (level-of-fill = "
<< LevelFill << ") = " << Condest << endl;
}
// Define label for printing out during the solve phase
string label = "Ifpack_CrsIct Preconditioner: LevelFill = " + toString(LevelFill) +
" Overlap = 0";
ICT->SetLabel(label.c_str());
// Here we create an AztecOO object
AztecOO solver;
solver.SetUserMatrix(&A);
solver.SetLHS(&x);
solver.SetRHS(&b);
solver.SetAztecOption(AZ_solver,AZ_cg);
// Here we set the IFPACK preconditioner and specify few parameters
solver.SetPrecOperator(ICT);
int Niters = 1200;
solver.SetAztecOption(AZ_kspace, Niters);
solver.SetAztecOption(AZ_output, 20);
solver.Iterate(Niters, 5.0e-5);
if (ICT!=0) delete ICT;
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
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 main(int argc, char *argv[]) {
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
// matrix downloaded from MatrixMarket
char FileName[] = "../HBMatrices/fidap005.rua";
Epetra_Map * readMap; // Pointers because of Trilinos_Util_ReadHb2Epetra
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(FileName, Comm, readMap, readA, readx,
readb, readxexact);
int NumGlobalElements = readMap->NumGlobalElements();
// Create uniform distributed map
Epetra_Map map(NumGlobalElements, 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
A.FillComplete();
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if( MyPID==0 ) {
cout << "Vector redistribute time (sec) = "
<< vectorRedistributeTime<< endl;
cout << "Matrix redistribute time (sec) = "
<< matrixRedistributeTime << endl;
cout << "Transform to Local time (sec) = "
<< fillCompleteTime << endl<< endl;
}
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return(EXIT_SUCCESS);
}
int Amesos_TestSolver( Epetra_Comm &Comm, char *matrix_file,
SparseSolverType SparseSolver,
bool transpose,
int special, AMESOS_MatrixType matrix_type ) {
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
std::string FileName = matrix_file ;
int FN_Size = FileName.size() ;
std::string LastFiveBytes = FileName.substr( EPETRA_MAX(0,FN_Size-5), FN_Size );
std::string LastFourBytes = FileName.substr( EPETRA_MAX(0,FN_Size-4), FN_Size );
bool NonContiguousMap = false;
if ( LastFiveBytes == ".triU" ) {
// Call routine to read in unsymmetric Triplet matrix
NonContiguousMap = true;
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, false, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFiveBytes == ".triS" ) {
NonContiguousMap = true;
// Call routine to read in symmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, true, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFourBytes == ".mtx" ) {
EPETRA_CHK_ERR( Trilinos_Util_ReadMatrixMarket2Epetra( matrix_file, Comm, readMap,
readA, readx, readb, readxexact) );
} else {
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra( matrix_file, Comm, readMap, readA, readx,
readb, readxexact) ;
}
}
}
Epetra_CrsMatrix transposeA(Copy, *readMap, 0);
Epetra_CrsMatrix *serialA ;
if ( transpose ) {
assert( CrsMatrixTranspose( readA, &transposeA ) == 0 );
serialA = &transposeA ;
} else {
serialA = readA ;
}
Epetra_RowMatrix * passA = 0;
Epetra_Vector * passx = 0;
Epetra_Vector * passb = 0;
Epetra_Vector * passxexact = 0;
Epetra_Vector * passresid = 0;
Epetra_Vector * passtmp = 0;
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
Epetra_Map* map_;
if( NonContiguousMap ) {
//
// map gives us NumMyElements and MyFirstElement;
//
int NumGlobalElements = readMap->NumGlobalElements();
int NumMyElements = map.NumMyElements();
int MyFirstElement = map.MinMyGID();
std::vector<int> MapMap_( NumGlobalElements );
readMap->MyGlobalElements( &MapMap_[0] ) ;
Comm.Broadcast( &MapMap_[0], NumGlobalElements, 0 ) ;
map_ = new Epetra_Map( NumGlobalElements, NumMyElements, &MapMap_[MyFirstElement], 0, Comm);
} else {
map_ = new Epetra_Map( map ) ;
}
Epetra_CrsMatrix A(Copy, *map_, 0);
const Epetra_Map &OriginalMap = serialA->RowMatrixRowMap() ;
assert( OriginalMap.SameAs(*readMap) );
Epetra_Export exporter(OriginalMap, *map_);
Epetra_Export exporter2(OriginalMap, *map_);
Epetra_Export MatrixExporter(OriginalMap, *map_);
Epetra_CrsMatrix AwithDiag(Copy, *map_, 0);
Epetra_Vector x(*map_);
Epetra_Vector b(*map_);
Epetra_Vector xexact(*map_);
Epetra_Vector resid(*map_);
Epetra_Vector readresid(*readMap);
Epetra_Vector tmp(*map_);
Epetra_Vector readtmp(*readMap);
// Epetra_Vector xcomp(*map_); // X as computed by the solver
bool distribute_matrix = ( matrix_type == AMESOS_Distributed ) ;
if ( distribute_matrix ) {
// Create Exporter to distribute read-in matrix and vectors
//
// Initialize x, b and xexact to the values read in from the file
//
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
A.Export(*serialA, exporter, Add);
assert(A.FillComplete()==0);
Comm.Barrier();
passA = &A;
passx = &x;
passb = &b;
passxexact = &xexact;
passresid = &resid;
passtmp = &tmp;
} else {
passA = serialA;
passx = readx;
passb = readb;
passxexact = readxexact;
passresid = &readresid;
passtmp = &readtmp;
}
Epetra_MultiVector CopyB( *passb ) ;
double Anorm = passA->NormInf() ;
SparseDirectTimingVars::SS_Result.Set_Anorm(Anorm) ;
Epetra_LinearProblem Problem( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb );
for ( int i = 0; i < 1+special ; i++ ) {
Epetra_Time TotalTime( Comm ) ;
if ( false ) {
// TEST_UMFPACK is never set by configure
#ifdef HAVE_AMESOS_SUPERLUDIST
} else if ( SparseSolver == SUPERLUDIST ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Superludist A_Superludist( Problem ) ;
//ParamList.set( "Redistribute", true );
//ParamList.set( "AddZeroToDiag", true );
Teuchos::ParameterList& SuperludistParams = ParamList.sublist("Superludist") ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_Superludist.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_Superludist.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_Superludist.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_Superludist.NumericFactorization( ) );
EPETRA_CHK_ERR( A_Superludist.Solve( ) );
#endif
#ifdef HAVE_AMESOS_DSCPACK
} else if ( SparseSolver == DSCPACK ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Dscpack A_dscpack( Problem ) ;
EPETRA_CHK_ERR( A_dscpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_dscpack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_dscpack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_dscpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SCALAPACK
} else if ( SparseSolver == SCALAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Scalapack A_scalapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_scalapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_scalapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_scalapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_scalapack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_scalapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_TAUCS
} else if ( SparseSolver == TAUCS ) {
Teuchos::ParameterList ParamList ;
Amesos_Taucs A_taucs( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_taucs.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_taucs.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_taucs.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_taucs.NumericFactorization( ) );
EPETRA_CHK_ERR( A_taucs.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARDISO
} else if ( SparseSolver == PARDISO ) {
Teuchos::ParameterList ParamList ;
Amesos_Pardiso A_pardiso( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_pardiso.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_pardiso.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_pardiso.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_pardiso.NumericFactorization( ) );
EPETRA_CHK_ERR( A_pardiso.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARAKLETE
} else if ( SparseSolver == PARAKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Paraklete A_paraklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_paraklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_paraklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_paraklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_paraklete.NumericFactorization( ) );
EPETRA_CHK_ERR( A_paraklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_MUMPS
} else if ( SparseSolver == MUMPS ) {
Teuchos::ParameterList ParamList ;
Amesos_Mumps A_mumps( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_mumps.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_mumps.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_mumps.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_mumps.NumericFactorization( ) );
EPETRA_CHK_ERR( A_mumps.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLU
} else if ( SparseSolver == SUPERLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Superlu A_superlu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_superlu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_superlu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_superlu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_superlu.NumericFactorization( ) );
EPETRA_CHK_ERR( A_superlu.Solve( ) );
#endif
#ifdef HAVE_AMESOS_LAPACK
} else if ( SparseSolver == LAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Lapack A_lapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_lapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_lapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_lapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_lapack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_lapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_UMFPACK
} else if ( SparseSolver == UMFPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Umfpack A_umfpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_umfpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_umfpack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_umfpack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_umfpack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_umfpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_KLU
} else if ( SparseSolver == KLU ) {
using namespace Teuchos;
Amesos_Time AT;
int setupTimePtr = -1, symTimePtr = -1, numTimePtr = -1, refacTimePtr = -1, solveTimePtr = -1;
AT.CreateTimer(Comm, 2);
AT.ResetTimer(0);
Teuchos::ParameterList ParamList ;
// ParamList.set("OutputLevel",2);
Amesos_Klu A_klu( Problem );
ParamList.set( "MaxProcs", -3 );
ParamList.set( "TrustMe", false );
// ParamList.set( "Refactorize", true );
EPETRA_CHK_ERR( A_klu.SetParameters( ParamList ) ) ;
EPETRA_CHK_ERR( A_klu.SetUseTranspose( transpose ) );
setupTimePtr = AT.AddTime("Setup", setupTimePtr, 0);
EPETRA_CHK_ERR( A_klu.SymbolicFactorization( ) );
symTimePtr = AT.AddTime("Symbolic", symTimePtr, 0);
EPETRA_CHK_ERR( A_klu.NumericFactorization( ) );
numTimePtr = AT.AddTime("Numeric", numTimePtr, 0);
EPETRA_CHK_ERR( A_klu.NumericFactorization( ) );
refacTimePtr = AT.AddTime("Refactor", refacTimePtr, 0);
// for ( int i=0; i<100000 ; i++ )
EPETRA_CHK_ERR( A_klu.Solve( ) );
solveTimePtr = AT.AddTime("Solve", solveTimePtr, 0);
double SetupTime = AT.GetTime(setupTimePtr);
double SymbolicTime = AT.GetTime(symTimePtr);
double NumericTime = AT.GetTime(numTimePtr);
double RefactorTime = AT.GetTime(refacTimePtr);
double SolveTime = AT.GetTime(solveTimePtr);
std::cout << __FILE__ << "::" << __LINE__ << " SetupTime = " << SetupTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " SymbolicTime = " << SymbolicTime - SetupTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " NumericTime = " << NumericTime - SymbolicTime<< std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " RefactorTime = " << RefactorTime - NumericTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " SolveTime = " << SolveTime - RefactorTime << std::endl ;
#endif
} else {
SparseDirectTimingVars::log_file << "Solver not implemented yet" << std::endl ;
std::cerr << "\n\n#################### Requested solver not available on this platform ##################### ATS\n" << std::endl ;
std::cout << " SparseSolver = " << SparseSolver << std::endl ;
std::cerr << " SparseSolver = " << SparseSolver << std::endl ;
}
SparseDirectTimingVars::SS_Result.Set_Total_Time( TotalTime.ElapsedTime() );
} // end for (int i=0; i<special; i++ )
//
// Compute the error = norm(xcomp - xexact )
//
double error;
passresid->Update(1.0, *passx, -1.0, *passxexact, 0.0);
passresid->Norm2(&error);
SparseDirectTimingVars::SS_Result.Set_Error(error) ;
// passxexact->Norm2(&error ) ;
// passx->Norm2(&error ) ;
//
// Compute the residual = norm(Ax - b)
//
double residual ;
passA->Multiply( transpose, *passx, *passtmp);
passresid->Update(1.0, *passtmp, -1.0, *passb, 0.0);
// passresid->Update(1.0, *passtmp, -1.0, CopyB, 0.0);
passresid->Norm2(&residual);
SparseDirectTimingVars::SS_Result.Set_Residual(residual) ;
double bnorm;
passb->Norm2( &bnorm ) ;
SparseDirectTimingVars::SS_Result.Set_Bnorm(bnorm) ;
double xnorm;
passx->Norm2( &xnorm ) ;
SparseDirectTimingVars::SS_Result.Set_Xnorm(xnorm) ;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
delete map_;
Comm.Barrier();
return 0;
}