StridedEpetraMap::StridedEpetraMap(global_size_t numGlobalElements, size_t numLocalElements, int indexBase,
std::vector<size_t>& stridingInfo, const Teuchos::RCP<const Teuchos::Comm<int> > &comm, LocalOrdinal stridedBlockId, GlobalOrdinal offset, const Teuchos::RCP<Node> &node)
: EpetraMap(numGlobalElements, numLocalElements, indexBase, comm, node), StridedMap<int, int>(numGlobalElements, numLocalElements, indexBase, stridingInfo, comm, stridedBlockId, offset)
{
// check input data and reorganize map
global_size_t numGlobalNodes = Teuchos::OrdinalTraits<global_size_t>::invalid();
if(numGlobalElements != Teuchos::OrdinalTraits<global_size_t>::invalid())
numGlobalNodes = numGlobalElements / getFixedBlockSize(); // number of nodes (over all processors)
size_t blockSize = getFixedBlockSize();
//if(stridedBlockId > -1) {
// blockSize = stridingInfo[stridedBlockId];
//}
size_t numLocalNodes = numLocalElements / blockSize; // number of nodes (on each processor)
// build an equally distributed node map
RCP<Epetra_Map> nodeMap = Teuchos::null;
IF_EPETRA_EXCEPTION_THEN_THROW_GLOBAL_INVALID_ARG((nodeMap = (rcp(new Epetra_Map(static_cast<int>(numGlobalNodes), numLocalNodes, indexBase, *toEpetra(comm))))));
// translate local node ids to local dofs
int nStridedOffset = 0;
int nDofsPerNode = Teuchos::as<int>(getFixedBlockSize()); // dofs per node for local striding block
if(stridedBlockId > -1) {
// determine nStridedOffset
for(int j=0; j<stridedBlockId; j++) {
nStridedOffset += stridingInfo[j];
}
nDofsPerNode = stridingInfo[stridedBlockId];
numGlobalElements = nodeMap->NumGlobalElements()*nDofsPerNode;
}
std::vector<int> dofgids;
for(int i = 0; i<nodeMap->NumMyElements(); i++) {
int gid = nodeMap->GID(i);
for(int dof = 0; dof < nDofsPerNode; ++dof) {
// dofs are calculated by
// global offset + node_GID * full size of strided map + striding offset of current striding block + dof id of current striding block
dofgids.push_back(offset_ + gid*getFixedBlockSize() + nStridedOffset + dof);
}
}
if (numGlobalElements == Teuchos::OrdinalTraits<global_size_t>::invalid()) {
IF_EPETRA_EXCEPTION_THEN_THROW_GLOBAL_INVALID_ARG((map_ = (rcp(new Epetra_BlockMap(-1, dofgids.size(), &dofgids[0], 1, indexBase, *toEpetra(comm))))));
} else {
IF_EPETRA_EXCEPTION_THEN_THROW_GLOBAL_INVALID_ARG((map_ = (rcp(new Epetra_BlockMap(numGlobalElements, dofgids.size(), &dofgids[0], 1, indexBase, *toEpetra(comm))))));
}
TEUCHOS_TEST_FOR_EXCEPTION(map_->NumMyPoints() % nDofsPerNode != 0, Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: wrong distribution of dofs among processors.");
if(stridedBlockId == -1) {
TEUCHOS_TEST_FOR_EXCEPTION(getNodeNumElements() != Teuchos::as<size_t>(nodeMap->NumMyElements()*nDofsPerNode), Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: wrong distribution of dofs among processors.");
TEUCHOS_TEST_FOR_EXCEPTION(getGlobalNumElements() != Teuchos::as<size_t>(nodeMap->NumGlobalElements()*nDofsPerNode), Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: wrong distribution of dofs among processors.");
} else {
TEUCHOS_TEST_FOR_EXCEPTION(stridingInfo.size() < Teuchos::as<size_t>(stridedBlockId), Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: stridedBlockId > stridingInfo.size()");
int nDofsInStridedBlock = stridingInfo[stridedBlockId];
TEUCHOS_TEST_FOR_EXCEPTION(getNodeNumElements() != Teuchos::as<size_t>(nodeMap->NumMyElements()*nDofsInStridedBlock), Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: wrong distribution of dofs among processors.");
TEUCHOS_TEST_FOR_EXCEPTION(getGlobalNumElements() != Teuchos::as<size_t>(nodeMap->NumGlobalElements()*nDofsInStridedBlock), Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: wrong distribution of dofs among processors.");
}
TEUCHOS_TEST_FOR_EXCEPTION(CheckConsistency() == false, Exceptions::RuntimeError, "StridedEpetraMap::StridedEpetraMap: CheckConsistency() == false");
}
int main(int argc, char *argv[]) {
//
int MyPID = 0;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
MyPID = Comm.MyPID();
#else
Epetra_SerialComm Comm;
#endif
//
typedef double ST;
typedef Teuchos::ScalarTraits<ST> SCT;
typedef SCT::magnitudeType MT;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Belos::MultiVecTraits<ST,MV> MVT;
typedef Belos::OperatorTraits<ST,MV,OP> OPT;
using Teuchos::ParameterList;
using Teuchos::RCP;
using Teuchos::rcp;
bool verbose = false, debug = false, proc_verbose = false;
int frequency = -1; // frequency of status test output.
int blocksize = 1; // blocksize
int numrhs = 1; // number of right-hand sides to solve for
int maxiters = -1; // maximum number of iterations allowed per linear system
std::string filenameMatrix("orsirr1_scaled.hb");
std::string filenameRHS; // blank mean unset
MT relResTol = 3.0e-4; // relative residual tolerance
// Like CG, LSQR is a short recurrence method that
// does not have the "n" step convergence property in finite precision arithmetic.
MT resGrowthFactor = 4.0; // In this example, warn if |resid| > resGrowthFactor * relResTol
// With no preconditioner, this is only the difference between the "implicit" and the "explict
// residual.
MT relMatTol = 1.e-4; // relative Matrix error, default value sqrt(eps)
MT maxCond = 1.e+8; // maximum condition number default value 1/eps
MT damp = 0.; // regularization (or damping) parameter
Teuchos::CommandLineProcessor cmdp(false,true); // e.g. ./a.out --tol=.1 --filename=foo.hb
cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
cmdp.setOption("debug","nondebug",&debug,"Print debugging information from solver.");
cmdp.setOption("frequency",&frequency,"Solvers frequency for printing residuals (#iters).");
cmdp.setOption("filename",&filenameMatrix,"Filename for test matrix. Acceptable file extensions: *.hb,*.mtx,*.triU,*.triS");
cmdp.setOption("rhsFilename",&filenameRHS,"Filename for right-hand side. Acceptable file extension: *.mtx");
cmdp.setOption("lambda",&damp,"Regularization parameter");
cmdp.setOption("tol",&relResTol,"Relative residual tolerance");
cmdp.setOption("matrixTol",&relMatTol,"Relative error in Matrix");
cmdp.setOption("max-cond",&maxCond,"Maximum condition number");
cmdp.setOption("num-rhs",&numrhs,"Number of right-hand sides to be solved for.");
cmdp.setOption("block-size",&blocksize,"Block size used by LSQR."); // must be one at this point
cmdp.setOption("max-iters",&maxiters,"Maximum number of iterations per linear system (-1 = adapted to problem/block size).");
if (cmdp.parse(argc,argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return -1;
}
if (!verbose)
frequency = -1; // reset frequency if test is not verbose
//
// Get the problem
//
RCP<Epetra_Map> Map;
RCP<Epetra_CrsMatrix> A;
RCP<Epetra_MultiVector> B, X;
RCP<Epetra_Vector> vecB, vecX;
EpetraExt::readEpetraLinearSystem(filenameMatrix, Comm, &A, &Map, &vecX, &vecB);
// Rectangular matrices are embedded in square matrices. vecX := 0, vecB = A*randVec
A->OptimizeStorage();
proc_verbose = verbose && (MyPID==0); /* Only print on the zero processor */
bool isRHS = false;
if ( filenameRHS != "" )
{
isRHS = true;
}
// Check to see if the number of right-hand sides is the same as requested.
if (numrhs>1) {
isRHS = false; // numrhs > 1 not yet supported
X = rcp( new Epetra_MultiVector( *Map, numrhs ) );
B = rcp( new Epetra_MultiVector( *Map, numrhs ) );
X->Seed();
X->Random();
OPT::Apply( *A, *X, *B ); // B := AX
X->PutScalar( 0.0 ); // annihilate X
}
else {
if ( isRHS )
{
Epetra_MultiVector * BmustDelete;
int mmRHSioflag = 0;
const char * charPtrRHSfn = filenameRHS.c_str();
mmRHSioflag = EpetraExt::MatrixMarketFileToMultiVector(charPtrRHSfn, *Map, BmustDelete);
//std::cout << "rhs from input file " << std::endl;
//BmustDelete->Print(std::cout);
if( mmRHSioflag )
{
if (proc_verbose)
std::cout << "Error " << mmRHSioflag << " occured while attempting to read file " << filenameRHS << std::endl;
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return -1;
}
X = rcp( new Epetra_MultiVector( *Map, numrhs ) );
X->Scale( 0.0 );
B = rcp( new MV(*BmustDelete));
delete BmustDelete;
}
else
{
int locNumCol = Map->MaxLID() + 1; // Create a known solution
int globNumCol = Map->MaxAllGID() + 1;
for( int li = 0; li < locNumCol; li++){ // assume consecutive lid
int gid = Map->GID(li);
double value = (double) ( globNumCol -1 - gid );
int numEntries = 1;
vecX->ReplaceGlobalValues( numEntries, &value, &gid );
}
bool Trans = false;
A->Multiply( Trans, *vecX, *vecB ); // Create a consistent linear system
// At this point, the initial guess is exact.
bool goodInitGuess = true; // perturb initial guess
bool zeroInitGuess = false; // annihilate initial guess
if( goodInitGuess )
{
double value = 1.e-2; // "Rel RHS Err" and "Rel Mat Err" apply to the residual equation,
int numEntries = 1; // norm( b - A x_k ) ?<? relResTol norm( b- Axo).
int index = 0; // norm(b) is inaccessible to LSQR.
vecX->SumIntoMyValues( numEntries, &value, &index);
}
if( zeroInitGuess )
{
vecX->PutScalar( 0.0 ); //
}
X = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecX);
B = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecB);
}
}
//
// ********Other information used by block solver***********
// *****************(can be user specified)******************
//
const int NumGlobalElements = B->GlobalLength();
if (maxiters == -1)
maxiters = NumGlobalElements/blocksize - 1; // maximum number of iterations to run
ParameterList belosList; // mechanism for configuring specific linear solver
belosList.set( "Block Size", blocksize ); // LSQR blocksize, must be one
belosList.set( "Lambda", damp ); // Regularization parameter
belosList.set( "Rel RHS Err", relResTol ); // Relative convergence tolerance requested
belosList.set( "Rel Mat Err", relMatTol ); // Maximum number of restarts allowed
belosList.set( "Condition Limit", maxCond); // upper bound for cond(A)
belosList.set( "Maximum Iterations", maxiters );// Maximum number of iterations allowed
int verbosity = Belos::Errors + Belos::Warnings;
if (verbose) {
verbosity += Belos::TimingDetails + Belos::StatusTestDetails;
if (frequency > 0)
belosList.set( "Output Frequency", frequency );
}
if (debug) {
verbosity += Belos::Debug;
}
belosList.set( "Verbosity", verbosity );
//
// Construct an unpreconditioned linear problem instance.
//
Belos::LinearProblem<double,MV,OP> problem( A, X, B );
bool set = problem.setProblem();
if (set == false) {
if (proc_verbose)
{
std::cout << std::endl << "ERROR: Belos::LinearProblem failed to set up correctly!" << std::endl;
}
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return -1;
}
// *******************************************************************
// ******************* Apply Single Vector LSQR **********************
// *******************************************************************
// Create an iterative solver manager.
RCP< Belos::LSQRSolMgr<double,MV,OP> > newSolver
= rcp( new Belos::LSQRSolMgr<double,MV,OP>(rcp(&problem,false), rcp(&belosList,false)));
if (proc_verbose) { // ******** Print a problem description *********
std::cout << std::endl << std::endl;
std::cout << "Dimension of matrix: " << NumGlobalElements << std::endl;
std::cout << "Number of right-hand sides: " << numrhs << std::endl;
std::cout << "Block size used by solver: " << blocksize << std::endl;
std::cout << "Max number of iterations for a linear system: " << maxiters << std::endl;
std::cout << "Relative residual tolerance: " << relResTol << std::endl;
std::cout << std::endl;
std::cout << "Solver's Description: " << std::endl;
std::cout << newSolver->description() << std::endl; // visually verify the parameter list
}
Belos::ReturnType ret = newSolver->solve(); // Perform solve
std::vector<double> solNorm( numrhs ); // get solution norm
MVT::MvNorm( *X, solNorm );
int numIters = newSolver->getNumIters(); // get number of solver iterations
MT condNum = newSolver->getMatCondNum();
MT matrixNorm= newSolver->getMatNorm();
MT resNorm = newSolver->getResNorm();
MT lsResNorm = newSolver->getMatResNorm();
if (proc_verbose)
std::cout << "Number of iterations performed for this solve: " << numIters << std::endl
<< "matrix condition number: " << condNum << std::endl
<< "matrix norm: " << matrixNorm << std::endl
<< "residual norm: " << resNorm << std::endl
<< "solution norm: " << solNorm[0] << std::endl
<< "least squares residual Norm: " << lsResNorm << std::endl;
bool badRes = false; // Compute the actual residuals.
std::vector<double> actual_resids( numrhs );
std::vector<double> rhs_norm( numrhs );
Epetra_MultiVector resid(*Map, numrhs);
OPT::Apply( *A, *X, resid );
MVT::MvAddMv( -1.0, resid, 1.0, *B, resid );
MVT::MvNorm( resid, actual_resids );
MVT::MvNorm( *B, rhs_norm );
if (proc_verbose) {
std::cout<< "---------- Actual Residuals (normalized) ----------"<<std::endl<<std::endl;
for ( int i=0; i<numrhs; i++) {
double actRes = actual_resids[i]/rhs_norm[i];
std::cout<<"Problem "<<i<<" : \t"<< actRes <<std::endl;
if (actRes > relResTol * resGrowthFactor)
{
badRes = true;
if (verbose) std::cout << "residual norm > " << relResTol * resGrowthFactor << std::endl;
}
}
}
if (ret!=Belos::Converged || badRes) {
if (proc_verbose)
std::cout << std::endl << "ERROR: Belos did not converge!" << std::endl;
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return -1;
}
if (proc_verbose)
std::cout << std::endl << "SUCCESS: Belos converged!" << std::endl;
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return 0; // Default return value
}
int main (int argc, char *argv[])
{
// Initialize MPI
Teuchos::GlobalMPISession (&argc, &argv, NULL);
// Create output stream. (Handy for multicore output.)
const RCP<Teuchos::FancyOStream> out =
Teuchos::VerboseObjectBase::getDefaultOStream();
// Create a communicator for Epetra objects
#ifdef HAVE_MPI
RCP<Epetra_MpiComm> eComm =
rcp<Epetra_MpiComm> (new Epetra_MpiComm (MPI_COMM_WORLD));
#else
RCP<Epetra_SerialComm> eComm =
rcp<Epetra_SerialComm> (new Epetra_SerialComm());
#endif
bool success = true;
try {
// Create map.
// Do strong scaling tests, so keep numGlobalElements independent of
// the number of processes.
int numGlobalElements = 5e7;
int indexBase = 0;
RCP<Epetra_Map> map =
rcp(new Epetra_Map (numGlobalElements, indexBase, *eComm));
//// Create map with overlay.
//int numMyOverlapNodes = 3;
//// Get an approximation of my nodes.
//int numMyElements = numGlobalElements / eComm->NumProc();
//int startIndex = eComm->MyPID() * numMyElements;
//// Calculate the resulting number of total nodes.
//int numTotalNodes = numMyElements * eComm->NumProc();
//// Add one node to the first numGlobalElements-numTotalNodes processes.
//if (eComm->MyPID() < numGlobalElements - numTotalNodes)
//{
// numMyElements++;
// startIndex += eComm->MyPID();
//}
//else
//{
// startIndex += numGlobalElements - numTotalNodes;
//}
//Teuchos::Array<int> indices(numMyElements);
//for (int k = 0; k<numMyElements; k++)
// indices[k] = startIndex + k;
//std::cout << numGlobalElements << std::endl;
//std::cout << numMyElements << std::endl;
//RCP<Epetra_Map> overlapMap =
// rcp(new Epetra_Map (numGlobalElements, numMyElements, indices.getRawPtr(), indexBase, *eComm));
//overlapMap->Print(std::cout);
//throw 1;
// tests on one vector
RCP<Epetra_Vector> u = rcp(new Epetra_Vector(*map));
u->Random();
RCP<Teuchos::Time> meanValueTime =
Teuchos::TimeMonitor::getNewTimer("Vector::MeanValue");
{
Teuchos::TimeMonitor tm(*meanValueTime);
double meanVal;
TEUCHOS_ASSERT_EQUALITY(0, u->MeanValue(&meanVal));
}
RCP<Teuchos::Time> maxValueTime =
Teuchos::TimeMonitor::getNewTimer("Vector::MaxValue");
{
Teuchos::TimeMonitor tm(*maxValueTime);
double maxValue;
TEUCHOS_ASSERT_EQUALITY(0, u->MaxValue(&maxValue));
}
RCP<Teuchos::Time> minValueTime =
Teuchos::TimeMonitor::getNewTimer("Vector::MinValue");
{
Teuchos::TimeMonitor tm(*minValueTime);
double minValue;
TEUCHOS_ASSERT_EQUALITY(0, u->MinValue(&minValue));
}
RCP<Teuchos::Time> norm1Time =
Teuchos::TimeMonitor::getNewTimer("Vector::Norm1");
{
Teuchos::TimeMonitor tm(*norm1Time);
double norm1;
TEUCHOS_ASSERT_EQUALITY(0, u->Norm1(&norm1));
}
RCP<Teuchos::Time> norm2Time =
Teuchos::TimeMonitor::getNewTimer("Vector::Norm2");
{
Teuchos::TimeMonitor tm(*norm2Time);
double norm2;
TEUCHOS_ASSERT_EQUALITY(0, u->Norm2(&norm2));
}
RCP<Teuchos::Time> normInfTime =
Teuchos::TimeMonitor::getNewTimer("Vector::NormInf");
{
Teuchos::TimeMonitor tm(*normInfTime);
double normInf;
TEUCHOS_ASSERT_EQUALITY(0, u->NormInf(&normInf));
}
RCP<Teuchos::Time> scaleTime =
Teuchos::TimeMonitor::getNewTimer("Vector::Scale");
{
Teuchos::TimeMonitor tm(*scaleTime);
double alpha = 0.5;
TEUCHOS_ASSERT_EQUALITY(0, u->Scale(0.5));
}
// tests involving two vectors
RCP<Epetra_Vector> v = rcp(new Epetra_Vector(*map));
v->Random();
RCP<Teuchos::Time> dotTime =
Teuchos::TimeMonitor::getNewTimer("Vector::Dot");
{
Teuchos::TimeMonitor tm(*dotTime);
double dot;
TEUCHOS_ASSERT_EQUALITY(0, u->Dot(*v, &dot));
}
RCP<Teuchos::Time> multiplyTime =
Teuchos::TimeMonitor::getNewTimer("Vector::Multiply");
{
Teuchos::TimeMonitor tm(*multiplyTime);
TEUCHOS_ASSERT_EQUALITY(0, u->Multiply(1.0, *u, *v, 1.0));
}
RCP<Teuchos::Time> updateTime =
Teuchos::TimeMonitor::getNewTimer("Vector::Update");
{
Teuchos::TimeMonitor tm(*updateTime);
TEUCHOS_ASSERT_EQUALITY(0, u->Update(1.0, *v, 1.0));
}
// matrix-vector tests
// diagonal test matrix
RCP<Epetra_CrsMatrix> D =
rcp(new Epetra_CrsMatrix(Copy, *map, 1));
for (int k = 0; k < map->NumMyElements(); k++) {
int col = map->GID(k);
double val = 1.0 / (col+1);
//TEUCHOS_ASSERT_EQUALITY(0, D->InsertMyValues(k, 1, &val, &col));
TEUCHOS_ASSERT_EQUALITY(0, D->InsertGlobalValues(col, 1, &val, &col));
}
TEUCHOS_ASSERT_EQUALITY(0, D->FillComplete());
// tridiagonal test matrix
RCP<Epetra_CrsMatrix> T =
rcp(new Epetra_CrsMatrix(Copy, *map, 3));
for (int k = 0; k < map->NumMyElements(); k++) {
int row = map->GID(k);
if (row > 0) {
int col = row-1;
double val = -1.0;
//TEUCHOS_ASSERT_EQUALITY(0, T->InsertMyValues(k, 1, &val, &col));
TEUCHOS_ASSERT_EQUALITY(0, T->InsertGlobalValues(row, 1, &val, &col));
}
{
int col = row;
double val = 2.0;
//TEUCHOS_ASSERT_EQUALITY(0, T->InsertMyValues(k, 1, &val, &col));
TEUCHOS_ASSERT_EQUALITY(0, T->InsertGlobalValues(row, 1, &val, &col));
}
if (row < numGlobalElements-1) {
int col = row+1;
double val = -1.0;
//TEUCHOS_ASSERT_EQUALITY(0, T->InsertMyValues(k, 1, &val, &col));
TEUCHOS_ASSERT_EQUALITY(0, T->InsertGlobalValues(row, 1, &val, &col));
}
}
TEUCHOS_ASSERT_EQUALITY(0, T->FillComplete());
// start timings
RCP<Teuchos::Time> mNorm1Time =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::Norm1");
{
Teuchos::TimeMonitor tm(*mNorm1Time);
double dNorm1 = D->NormOne();
double tNorm1 = T->NormOne();
}
RCP<Teuchos::Time> mNormInfTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::NormInf");
{
Teuchos::TimeMonitor tm(*mNormInfTime);
double dNormInf = D->NormInf();
double tNormInf = T->NormInf();
}
RCP<Teuchos::Time> mNormFrobTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::NormFrobenius");
{
Teuchos::TimeMonitor tm(*mNormFrobTime);
double dNormFrob = D->NormFrobenius();
double tNormFrob = T->NormFrobenius();
}
RCP<Teuchos::Time> mScaleTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::Scale");
{
Teuchos::TimeMonitor tm(*mScaleTime);
TEUCHOS_ASSERT_EQUALITY(0, D->Scale(2.0));
TEUCHOS_ASSERT_EQUALITY(0, T->Scale(2.0));
}
RCP<Teuchos::Time> leftScaleTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::LeftScale");
{
Teuchos::TimeMonitor tm(*leftScaleTime);
TEUCHOS_ASSERT_EQUALITY(0, D->LeftScale(*v));
TEUCHOS_ASSERT_EQUALITY(0, T->LeftScale(*v));
}
RCP<Teuchos::Time> rightScaleTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::RightScale");
{
Teuchos::TimeMonitor tm(*rightScaleTime);
TEUCHOS_ASSERT_EQUALITY(0, D->RightScale(*v));
TEUCHOS_ASSERT_EQUALITY(0, T->RightScale(*v));
}
RCP<Teuchos::Time> applyTime =
Teuchos::TimeMonitor::getNewTimer("CrsMatrix::Apply");
{
Teuchos::TimeMonitor tm(*applyTime);
TEUCHOS_ASSERT_EQUALITY(0, D->Apply(*u, *v));
TEUCHOS_ASSERT_EQUALITY(0, T->Apply(*u, *v));
}
// print timing data
Teuchos::TimeMonitor::summarize();
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(true, *out, success);
return success ? EXIT_SUCCESS : EXIT_FAILURE;
}
int main(int argc, char *argv[]) {
//
int MyPID = 0;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
MyPID = Comm.MyPID();
#else
Epetra_SerialComm Comm;
#endif
//
typedef double ST;
typedef Teuchos::ScalarTraits<ST> SCT;
typedef SCT::magnitudeType MT;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Belos::MultiVecTraits<ST,MV> MVT;
typedef Belos::OperatorTraits<ST,MV,OP> OPT;
using Teuchos::ParameterList;
using Teuchos::RCP;
using Teuchos::rcp;
bool verbose = false;
bool success = true;
try {
bool proc_verbose = false;
bool leftprec = true; // left preconditioning or right.
// LSQR applies the operator and the transposed operator.
// A preconditioner must support transpose multiply.
int frequency = -1; // frequency of status test output.
int blocksize = 1; // blocksize
// LSQR as currently implemented is a single vector algorithm.
// However some of the parameters that would be used by a block version
// have not been removed from this file.
int numrhs = 1; // number of right-hand sides to solve for
int maxiters = -1; // maximum number of iterations allowed per linear system
std::string filename("orsirr1_scaled.hb");
MT relResTol = 1.0e-5; // relative residual tolerance for the preconditioned linear system
MT resGrowthFactor = 1.0; // In this example, warn if |resid| > resGrowthFactor * relResTol
MT relMatTol = 1.e-10; // relative Matrix error, default value sqrt(eps)
MT maxCond = 1.e+5; // maximum condition number default value 1/eps
MT damp = 0.; // regularization (or damping) parameter
Teuchos::CommandLineProcessor cmdp(false,true);
cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
cmdp.setOption("left-prec","right-prec",&leftprec,"Left preconditioning or right.");
cmdp.setOption("frequency",&frequency,"Solvers frequency for printing residuals (#iters).");
cmdp.setOption("filename",&filename,"Filename for test matrix. Acceptable file extensions: *.hb,*.mtx,*.triU,*.triS");
cmdp.setOption("lambda",&damp,"Regularization parameter");
cmdp.setOption("tol",&relResTol,"Relative residual tolerance");
cmdp.setOption("matrixTol",&relMatTol,"Relative error in Matrix");
cmdp.setOption("max-cond",&maxCond,"Maximum condition number");
cmdp.setOption("num-rhs",&numrhs,"Number of right-hand sides to be solved for.");
cmdp.setOption("block-size",&blocksize,"Block size used by LSQR.");
cmdp.setOption("max-iters",&maxiters,"Maximum number of iterations per linear system (-1 = adapted to problem/block size).");
if (cmdp.parse(argc,argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
return -1;
}
if (!verbose)
frequency = -1; // reset frequency if test is not verbose
//
// *************Get the problem*********************
//
RCP<Epetra_Map> Map;
RCP<Epetra_CrsMatrix> A;
RCP<Epetra_MultiVector> B, X;
RCP<Epetra_Vector> vecB, vecX;
EpetraExt::readEpetraLinearSystem(filename, Comm, &A, &Map, &vecX, &vecB);
A->OptimizeStorage();
proc_verbose = verbose && (MyPID==0); /* Only print on the zero processor */
// Check to see if the number of right-hand sides is the same as requested.
if (numrhs>1) {
X = rcp( new Epetra_MultiVector( *Map, numrhs ) );
B = rcp( new Epetra_MultiVector( *Map, numrhs ) );
X->Random();
OPT::Apply( *A, *X, *B );
X->PutScalar( 0.0 );
}
else {
int locNumCol = Map->MaxLID() + 1; // Create a known solution
int globNumCol = Map->MaxAllGID() + 1;
for( int li = 0; li < locNumCol; li++){ // assume consecutive lid
int gid = Map->GID(li);
double value = (double) ( globNumCol -1 - gid );
int numEntries = 1;
vecX->ReplaceGlobalValues( numEntries, &value, &gid );
}
bool Trans = false;
A->Multiply( Trans, *vecX, *vecB ); // Create a consistent linear system
// At this point, the initial guess is exact.
bool zeroInitGuess = false; // annihilate initial guess
bool goodInitGuess = true; // initial guess near solution
if( zeroInitGuess )
{
vecX->PutScalar( 0.0 );
}
else
{
if( goodInitGuess )
{
double value = 1.e-2; // "Rel RHS Err" and "Rel Mat Err" apply to the residual equation,
int numEntries = 1; // norm( b - A x_k ) ?<? relResTol norm( b- Axo).
int index = 0; // norm(b) is inaccessible to LSQR.
vecX->SumIntoMyValues( numEntries, &value, &index);
}
}
X = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecX);
B = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecB);
}
//
// ************Construct preconditioner*************
//
ParameterList ifpackList;
// allocates an IFPACK factory. No data is associated
// to this object (only method Create()).
Ifpack Factory; // do support transpose multiply
// create the preconditioner. For valid PrecType values,
// please check the documentation
std::string PrecType = "ILU"; // incomplete LU
int OverlapLevel = 1; // nonnegative
RCP<Ifpack_Preconditioner> Prec = Teuchos::rcp( Factory.Create(PrecType, &*A, OverlapLevel) );
assert(Prec != Teuchos::null);
// specify parameters for ILU
ifpackList.set("fact: level-of-fill", 1);
// the combine mode is on the following:
// "Add", "Zero", "Insert", "InsertAdd", "Average", "AbsMax"
// Their meaning is as defined in file Epetra_CombineMode.h
ifpackList.set("schwarz: combine mode", "Add");
// sets the parameters
IFPACK_CHK_ERR(Prec->SetParameters(ifpackList));
// initialize the preconditioner. At this point the matrix must
// have been FillComplete()'d, but actual values are ignored.
IFPACK_CHK_ERR(Prec->Initialize());
// Builds the preconditioners, by looking for the values of
// the matrix.
IFPACK_CHK_ERR(Prec->Compute());
{
const int errcode = Prec->SetUseTranspose (true);
if (errcode != 0) {
throw std::logic_error ("Oh hai! Ifpack_Preconditioner doesn't know how to apply its transpose.");
} else {
(void) Prec->SetUseTranspose (false);
}
}
// Create the Belos preconditioned operator from the Ifpack preconditioner.
// NOTE: This is necessary because Belos expects an operator to apply the
// preconditioner with Apply() NOT ApplyInverse().
RCP<Belos::EpetraPrecOp> belosPrec = rcp( new Belos::EpetraPrecOp( Prec ) );
//
// *****Create parameter list for the LSQR solver manager*****
//
const int NumGlobalElements = B->GlobalLength();
if (maxiters == -1)
maxiters = NumGlobalElements/blocksize - 1; // maximum number of iterations to run
//
ParameterList belosList;
belosList.set( "Block Size", blocksize ); // Blocksize to be used by iterative solver
belosList.set( "Lambda", damp ); // Regularization parameter
belosList.set( "Rel RHS Err", relResTol ); // Relative convergence tolerance requested
belosList.set( "Rel Mat Err", relMatTol ); // Maximum number of restarts allowed
belosList.set( "Condition Limit", maxCond); // upper bound for cond(A)
belosList.set( "Maximum Iterations", maxiters );// Maximum number of iterations allowed
if (numrhs > 1) {
belosList.set( "Show Maximum Residual Norm Only", true ); // Show only the maximum residual norm
}
if (verbose) {
belosList.set( "Verbosity", Belos::Errors + Belos::Warnings +
Belos::TimingDetails + Belos::StatusTestDetails );
if (frequency > 0)
belosList.set( "Output Frequency", frequency );
}
else
belosList.set( "Verbosity", Belos::Errors + Belos::Warnings );
//
// *******Construct a preconditioned linear problem********
//
RCP<Belos::LinearProblem<double,MV,OP> > problem
= rcp( new Belos::LinearProblem<double,MV,OP>( A, X, B ) );
if (leftprec) {
problem->setLeftPrec( belosPrec );
}
else {
problem->setRightPrec( belosPrec );
}
bool set = problem->setProblem();
if (set == false) {
if (proc_verbose)
std::cout << std::endl << "ERROR: Belos::LinearProblem failed to set up correctly!" << std::endl;
return -1;
}
// Create an iterative solver manager.
RCP< Belos::LSQRSolMgr<double,MV,OP> > solver
= rcp( new Belos::LSQRSolMgr<double,MV,OP>(problem, rcp(&belosList,false)));
//
// *******************************************************************
// ******************Start the LSQR iteration*************************
// *******************************************************************
//
if (proc_verbose) {
std::cout << std::endl << std::endl;
std::cout << "Dimension of matrix: " << NumGlobalElements << std::endl;
std::cout << "Number of right-hand sides: " << numrhs << std::endl;
std::cout << "Block size used by solver: " << blocksize << std::endl;
std::cout << "Max number of Gmres iterations per restart cycle: " << maxiters << std::endl;
std::cout << "Relative residual tolerance: " << relResTol << std::endl;
std::cout << std::endl;
std::cout << "Solver's Description: " << std::endl;
std::cout << solver->description() << std::endl; // visually verify the parameter list
}
//
// Perform solve
//
Belos::ReturnType ret = solver->solve();
//
// Get the number of iterations for this solve.
//
std::vector<double> solNorm( numrhs ); // get solution norm
MVT::MvNorm( *X, solNorm );
int numIters = solver->getNumIters();
MT condNum = solver->getMatCondNum();
MT matrixNorm= solver->getMatNorm();
MT resNorm = solver->getResNorm();
MT lsResNorm = solver->getMatResNorm();
if (proc_verbose)
std::cout << "Number of iterations performed for this solve: " << numIters << std::endl
<< "matrix condition number: " << condNum << std::endl
<< "matrix norm: " << matrixNorm << std::endl
<< "residual norm: " << resNorm << std::endl
<< "solution norm: " << solNorm[0] << std::endl
<< "least squares residual Norm: " << lsResNorm << std::endl;
//
// Compute actual residuals.
//
bool badRes = false;
std::vector<double> actual_resids( numrhs );
std::vector<double> rhs_norm( numrhs );
Epetra_MultiVector resid(*Map, numrhs);
OPT::Apply( *A, *X, resid );
MVT::MvAddMv( -1.0, resid, 1.0, *B, resid );
MVT::MvNorm( resid, actual_resids );
MVT::MvNorm( *B, rhs_norm );
if (proc_verbose) {
std::cout<< "---------- Actual Residuals (normalized) ----------"<<std::endl<<std::endl;
for ( int i=0; i<numrhs; i++) {
double actRes = actual_resids[i]/rhs_norm[i];
std::cout<<"Problem "<<i<<" : \t"<< actRes <<std::endl;
if (actRes > relResTol * resGrowthFactor ) badRes = true;
}
}
if (ret!=Belos::Converged || badRes) {
success = false;
if (proc_verbose)
std::cout << std::endl << "ERROR: Belos did not converge!" << std::endl;
} else {
success = true;
if (proc_verbose)
std::cout << std::endl << "SUCCESS: Belos converged!" << std::endl;
}
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return success ? EXIT_SUCCESS : EXIT_FAILURE;
}