int AmesosGenOp::Apply(const Epetra_MultiVector& X, Epetra_MultiVector& Y ) const
{
if (!useTranspose_) {
// Storage for M*X
Epetra_MultiVector MX(X.Map(),X.NumVectors());
// Apply M*X
massMtx_->Apply(X, MX);
Y.PutScalar(0.0);
// Set the LHS and RHS
problem_->SetRHS(&MX);
problem_->SetLHS(&Y);
// Solve the linear system A*Y = MX
solver_->Solve();
}
else {
// Storage for A^{-T}*X
Epetra_MultiVector ATX(X.Map(),X.NumVectors());
Epetra_MultiVector tmpX = const_cast<Epetra_MultiVector&>(X);
// Set the LHS and RHS
problem_->SetRHS(&tmpX);
problem_->SetLHS(&ATX);
// Solve the linear system A^T*Y = X
solver_->Solve();
// Apply M*ATX
massMtx_->Apply(ATX, Y);
}
return 0;
}
int
main (int argc, char *argv[])
{
using namespace Anasazi;
using Teuchos::RCP;
using Teuchos::rcp;
using std::endl;
#ifdef HAVE_MPI
// Initialize MPI
MPI_Init (&argc, &argv);
#endif // HAVE_MPI
// Create an Epetra communicator
#ifdef HAVE_MPI
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif // HAVE_MPI
// Create an Anasazi output manager
BasicOutputManager<double> printer;
printer.stream(Errors) << Anasazi_Version() << std::endl << std::endl;
// Get the sorting std::string from the command line
std::string which ("LM");
Teuchos::CommandLineProcessor cmdp (false, true);
cmdp.setOption("sort", &which, "Targetted eigenvalues (SM or LM).");
if (cmdp.parse (argc, argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return -1;
}
// Dimension of the matrix
//
// Discretization points in any one direction.
const int nx = 10;
// Size of matrix nx*nx
const int NumGlobalElements = nx*nx;
// Construct a Map that puts approximately the same number of
// equations on each process.
Epetra_Map Map (NumGlobalElements, 0, Comm);
// Get update list and number of local equations from newly created Map.
int NumMyElements = Map.NumMyElements ();
std::vector<int> MyGlobalElements (NumMyElements);
Map.MyGlobalElements (&MyGlobalElements[0]);
// Create an integer vector NumNz that is used to build the Petra
// matrix. NumNz[i] is the number of OFF-DIAGONAL terms for the
// i-th global equation on this process.
std::vector<int> NumNz (NumMyElements);
/* We are building a matrix of block structure:
| T -I |
|-I T -I |
| -I T |
| ... -I|
| -I T|
where each block is dimension nx by nx and the matrix is on the order of
nx*nx. The block T is a tridiagonal matrix.
*/
for (int i=0; i<NumMyElements; ++i) {
if (MyGlobalElements[i] == 0 || MyGlobalElements[i] == NumGlobalElements-1 ||
MyGlobalElements[i] == nx-1 || MyGlobalElements[i] == nx*(nx-1) ) {
NumNz[i] = 3;
}
else if (MyGlobalElements[i] < nx || MyGlobalElements[i] > nx*(nx-1) ||
MyGlobalElements[i]%nx == 0 || (MyGlobalElements[i]+1)%nx == 0) {
NumNz[i] = 4;
}
else {
NumNz[i] = 5;
}
}
// Create an Epetra_Matrix
RCP<Epetra_CrsMatrix> A = rcp (new Epetra_CrsMatrix (Epetra_DataAccess::Copy, Map, &NumNz[0]));
// Compute coefficients for discrete convection-diffution operator
const double one = 1.0;
std::vector<double> Values(4);
std::vector<int> Indices(4);
double rho = 0.0;
double h = one /(nx+1);
double h2 = h*h;
double c = 5.0e-01*rho/ h;
Values[0] = -one/h2 - c; Values[1] = -one/h2 + c; Values[2] = -one/h2; Values[3]= -one/h2;
double diag = 4.0 / h2;
int NumEntries;
for (int i=0; i<NumMyElements; ++i) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
Indices[1] = nx;
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i] == nx*(nx-1)) {
Indices[0] = nx*(nx-1)+1;
Indices[1] = nx*(nx-2);
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i] == nx-1) {
Indices[0] = nx-2;
NumEntries = 1;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
Indices[0] = 2*nx-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i] == NumGlobalElements-1) {
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
Indices[0] = nx*(nx-1)-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i] < nx) {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i] > nx*(nx-1)) {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]-nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if (MyGlobalElements[i]%nx == 0) {
Indices[0] = MyGlobalElements[i]+1;
Indices[1] = MyGlobalElements[i]-nx;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else if ((MyGlobalElements[i]+1)%nx == 0) {
Indices[0] = MyGlobalElements[i]-nx;
Indices[1] = MyGlobalElements[i]+nx;
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
Indices[0] = MyGlobalElements[i]-1;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]-nx;
Indices[3] = MyGlobalElements[i]+nx;
NumEntries = 4;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
// Put in the diagonal entry
int info = A->InsertGlobalValues(MyGlobalElements[i], 1, &diag, &MyGlobalElements[i]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "InsertGlobalValues returned info = "
<< info << " != 0." );
}
// Finish up
int info = A->FillComplete ();
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "A->FillComplete() returned info = "
<< info << " != 0." );
A->SetTracebackMode (1); // Shutdown Epetra Warning tracebacks
// Create a identity matrix for the temporary mass matrix
RCP<Epetra_CrsMatrix> M = rcp (new Epetra_CrsMatrix (Epetra_DataAccess::Copy, Map, 1));
for (int i=0; i<NumMyElements; i++) {
Values[0] = one;
Indices[0] = i;
NumEntries = 1;
info = M->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "M->InsertGlobalValues() returned info = "
<< info << " != 0." );
}
// Finish up
info = M->FillComplete ();
TEUCHOS_TEST_FOR_EXCEPTION
(info != 0, std::runtime_error, "M->FillComplete() returned info = "
<< info << " != 0." );
M->SetTracebackMode (1); // Shutdown Epetra Warning tracebacks
//************************************
// Call the LOBPCG solver manager
//***********************************
//
// Variables used for the LOBPCG Method
const int nev = 10;
const int blockSize = 5;
const int maxIters = 500;
const double tol = 1.0e-8;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef MultiVecTraits<double, Epetra_MultiVector> MVT;
// Create an Epetra_MultiVector for an initial vector to start the
// solver. Note: This needs to have the same number of columns as
// the blocksize.
RCP<Epetra_MultiVector> ivec = rcp (new Epetra_MultiVector (Map, blockSize));
ivec->Random (); // fill the initial vector with random values
// Create the eigenproblem.
RCP<BasicEigenproblem<double, MV, OP> > MyProblem =
rcp (new BasicEigenproblem<double, MV, OP> (A, ivec));
// Inform the eigenproblem that the operator A is symmetric
MyProblem->setHermitian (true);
// Set the number of eigenvalues requested
MyProblem->setNEV (nev);
// Tell the eigenproblem that you are finishing passing it information.
const bool success = MyProblem->setProblem ();
if (! success) {
printer.print (Errors, "Anasazi::BasicEigenproblem::setProblem() reported an error.\n");
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return -1;
}
// Create parameter list to pass into the solver manager
Teuchos::ParameterList MyPL;
MyPL.set ("Which", which);
MyPL.set ("Block Size", blockSize);
MyPL.set ("Maximum Iterations", maxIters);
MyPL.set ("Convergence Tolerance", tol);
MyPL.set ("Full Ortho", true);
MyPL.set ("Use Locking", true);
// Create the solver manager
LOBPCGSolMgr<double, MV, OP> MySolverMan (MyProblem, MyPL);
// Solve the problem
ReturnType returnCode = MySolverMan.solve ();
// Get the eigenvalues and eigenvectors from the eigenproblem
Eigensolution<double,MV> sol = MyProblem->getSolution ();
std::vector<Value<double> > evals = sol.Evals;
RCP<MV> evecs = sol.Evecs;
// Compute residuals.
std::vector<double> normR (sol.numVecs);
if (sol.numVecs > 0) {
Teuchos::SerialDenseMatrix<int,double> T (sol.numVecs, sol.numVecs);
Epetra_MultiVector tempAevec (Map, sol.numVecs );
T.putScalar (0.0);
for (int i = 0; i < sol.numVecs; ++i) {
T(i,i) = evals[i].realpart;
}
A->Apply (*evecs, tempAevec);
MVT::MvTimesMatAddMv (-1.0, *evecs, T, 1.0, tempAevec);
MVT::MvNorm (tempAevec, normR);
}
// Print the results
std::ostringstream os;
os.setf (std::ios_base::right, std::ios_base::adjustfield);
os << "Solver manager returned "
<< (returnCode == Converged ? "converged." : "unconverged.") << endl;
os << endl;
os << "------------------------------------------------------" << endl;
os << std::setw(16) << "Eigenvalue"
<< std::setw(18) << "Direct Residual"
<< endl;
os << "------------------------------------------------------" << endl;
for (int i = 0; i < sol.numVecs; ++i) {
os << std::setw(16) << evals[i].realpart
<< std::setw(18) << normR[i] / evals[i].realpart
<< endl;
}
os << "------------------------------------------------------" << endl;
printer.print (Errors, os.str ());
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return 0;
}
int
main (int argc, char *argv[])
{
using Belos::OutputManager;
using Belos::SolverFactory;
using Teuchos::CommandLineProcessor;
using Teuchos::null;
using Teuchos::ParameterList;
using Teuchos::parameterList;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcp_dynamic_cast;
using Teuchos::rcp_implicit_cast;
using std::cerr;
using std::cout;
using std::endl;
typedef double scalar_type;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Belos::SolverManager<scalar_type, MV, OP> solver_base_type;
typedef Belos::SolverFactory<scalar_type, MV, OP> factory_type;
Teuchos::oblackholestream blackHole;
Teuchos::GlobalMPISession mpiSession (&argc, &argv, &blackHole);
RCP<Epetra_Comm> comm;
{
#ifdef EPETRA_MPI
RCP<Epetra_MpiComm> commSpecific (new Epetra_MpiComm (MPI_COMM_WORLD));
#else
RCP<Epetra_SerialComm> commSpecific (new Epetra_SerialComm);
#endif // EPETRA_MPI
comm = rcp_implicit_cast<Epetra_Comm> (commSpecific);
}
std::ostream& out = (comm->MyPID() == 0) ? std::cout : blackHole;
bool success = false;
bool verbose = false;
try {
int numRHS = 1;
bool debug = false;
// Define command-line arguments.
CommandLineProcessor cmdp (false, true);
cmdp.setOption ("numRHS", &numRHS, "Number of right-hand sides in the linear "
"system to solve.");
cmdp.setOption ("verbose", "quiet", &verbose, "Print messages and results.");
cmdp.setOption ("debug", "nodebug", &debug, "Print debugging information.");
// Parse the command-line arguments.
{
const CommandLineProcessor::EParseCommandLineReturn parseResult =
cmdp.parse (argc,argv);
if (parseResult == CommandLineProcessor::PARSE_HELP_PRINTED) {
if (comm->MyPID() == 0)
std::cout << "End Result: TEST PASSED" << endl;
return EXIT_SUCCESS;
}
TEUCHOS_TEST_FOR_EXCEPTION(parseResult != CommandLineProcessor::PARSE_SUCCESSFUL,
std::invalid_argument, "Failed to parse command-line arguments.");
}
// Declare an output manager for handling local output. Initialize,
// using the caller's desired verbosity level.
RCP<OutputManager<scalar_type> > outMan =
rcp (new OutputManager<scalar_type> (selectVerbosity (verbose, debug)));
// Stream for debug output. If debug output is not enabled, then
// this stream doesn't print anything sent to it (it's a "black
// hole" stream).
//std::ostream& debugOut = outMan->stream (Belos::Debug);
// Create the operator to test, with domain and range maps.
RCP<const Epetra_Map> domainMap, rangeMap;
RCP<Epetra_Operator> A = makeMatrix (comm, domainMap, rangeMap);
// "Solution" input/output multivector.
RCP<MV> X_exact = rcp (new MV (*domainMap, numRHS));
X_exact->Seed ();
X_exact->Random ();
RCP<MV> X = rcp (new MV (*domainMap, numRHS));
X->PutScalar (0.0);
// "Right-hand side" input multivector.
RCP<MV> B = rcp (new MV (*rangeMap, numRHS));
A->Apply (*X_exact, *B);
//
// Test creating solver instances using the solver factory.
//
factory_type factory;
//
// Test the canonical solver names.
//
{
typedef Belos::BlockGmresSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Block GMRES", out, verbose);
}
{
typedef Belos::PseudoBlockGmresSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Pseudoblock GMRES", out, verbose);
}
{
typedef Belos::BlockCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Block CG", out, verbose);
}
{
typedef Belos::PseudoBlockCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Pseudoblock CG", out, verbose);
}
{
typedef Belos::GCRODRSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "GCRODR", out, verbose);
}
{
typedef Belos::RCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "RCG", out, verbose);
}
{
typedef Belos::MinresSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "MINRES", out, verbose);
}
{
typedef Belos::LSQRSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "LSQR", out, verbose);
}
//
// Test aliases.
//
{
typedef Belos::BlockGmresSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Flexible GMRES", out, verbose);
}
{
typedef Belos::PseudoBlockCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "CG", out, verbose);
}
{
typedef Belos::RCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Recycling CG", out, verbose);
}
{
typedef Belos::GCRODRSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Recycling GMRES", out, verbose);
}
{
typedef Belos::PseudoBlockGmresSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Pseudo Block GMRES", out, verbose);
}
{
typedef Belos::PseudoBlockCGSolMgr<scalar_type, MV, OP> solver_impl_type;
testCreatingSolver<factory_type, solver_base_type,
solver_impl_type> (factory, "Pseudo Block CG", out, verbose);
}
success = true;
if (comm->MyPID() == 0) {
cout << "End Result: TEST PASSED" << endl;
}
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
return ( success ? EXIT_SUCCESS : EXIT_FAILURE );
}
// =============================================================================
int main (int argc, char *argv[])
{
Teuchos::GlobalMPISession(&argc, &argv, NULL);
// Create a communicator for Epetra objects
#ifdef HAVE_MPI
Epetra_MpiComm eComm(MPI_COMM_WORLD);
#else
Epetra_SerialComm eComm();
#endif
const RCP<Teuchos::FancyOStream> out =
Teuchos::VerboseObjectBase::getDefaultOStream();
bool success = true;
try {
// ===========================================================================
// handle command line arguments
Teuchos::CommandLineProcessor My_CLP;
My_CLP.setDocString("Linear solver testbed for the 1D Poisson matrix.\n");
std::string action("matvec");
My_CLP.setOption("action",
&action,
"Which action to perform with the operator (matvec, solve_cg, solve_minres, solve_gmres)"
);
std::string solver("cg");
// My_CLP.setOption("solver", &solver, "Krylov subspace method (cg, minres, gmres)");
// Operator op = JAC;
// Operator allOpts[] = {JAC, KEO, KEOREG, POISSON1D};
// std::string allOptNames[] = {"jac", "keo", "keoreg", "poisson1d"};
// My_CLP.setOption("operator", &op, 4, allOpts, allOptNames);
bool verbose = true;
My_CLP.setOption("verbose", "quiet",
&verbose,
"Print messages and results.");
int frequency = 10;
My_CLP.setOption("frequency",
&frequency,
"Solvers frequency for printing residuals (#iters).");
// Make sure this value is large enough to keep the cores busy for a while.
int n = 1000;
My_CLP.setOption("size",
&n,
"Size of the equation system (default: 1000).");
// print warning for unrecognized arguments
My_CLP.recogniseAllOptions(true);
My_CLP.throwExceptions(false);
// finally, parse the command line
TEUCHOS_ASSERT_EQUALITY(My_CLP.parse (argc, argv),
Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL
);
// =========================================================================
// Construct Epetra matrix.
RCP<Teuchos::Time> matrixConstructTime =
Teuchos::TimeMonitor::getNewTimer("Epetra matrix construction");
RCP<Epetra_CrsMatrix> A;
{
Teuchos::TimeMonitor tm(*matrixConstructTime);
A = contructEpetraMatrix(n, eComm);
}
// A->Print(std::cout);
// create initial guess and right-hand side
RCP<Epetra_Vector> x =
rcp(new Epetra_Vector(A->OperatorDomainMap()));
RCP<Epetra_MultiVector> b =
rcp(new Epetra_Vector(A->OperatorRangeMap()));
// b->Random();
TEUCHOS_ASSERT_EQUALITY(0, b->PutScalar(1.0));
if (action.compare("matvec") == 0) {
TEUCHOS_ASSERT_EQUALITY(0, x->PutScalar(1.0));
RCP<Teuchos::Time> mvTime = Teuchos::TimeMonitor::getNewTimer("Epetra operator apply");
{
Teuchos::TimeMonitor tm(*mvTime);
// Don't TEUCHOS_ASSERT_EQUALITY() here for speed.
A->Apply(*x, *b);
}
// print timing data
Teuchos::TimeMonitor::summarize();
} else {
// -----------------------------------------------------------------------
// Belos part
Teuchos::ParameterList belosList;
// Relative convergence tolerance requested
belosList.set("Convergence Tolerance", 1.0e-12);
if (verbose) {
belosList.set("Verbosity",
Belos::Errors +
Belos::Warnings +
Belos::IterationDetails +
Belos::FinalSummary +
Belos::Debug +
Belos::TimingDetails +
Belos::StatusTestDetails
);
if (frequency > 0)
belosList.set("Output Frequency", frequency);
} else {
belosList.set("Verbosity", Belos::Errors + Belos::Warnings);
}
// Belos::General, Belos::Brief
belosList.set("Output Style", static_cast<int>(Belos::Brief));
belosList.set("Maximum Iterations", 1000);
// Construct an unpreconditioned linear problem instance.
Belos::LinearProblem<double, MV, OP> problem(A, x, b);
bool set = problem.setProblem();
TEUCHOS_TEST_FOR_EXCEPTION(
!set,
std::logic_error,
"ERROR: Belos::LinearProblem failed to set up correctly!"
);
// -----------------------------------------------------------------------
// Create an iterative solver manager.
RCP<Belos::SolverManager<double, MV, OP> > newSolver;
if (action.compare("solve_cg") == 0) {
belosList.set("Assert Positive Definiteness", false);
newSolver =
rcp(new Belos::PseudoBlockCGSolMgr<double, MV, OP>(
rcp(&problem, false),
rcp(&belosList, false)
));
} else if (action.compare("solve_minres") == 0) {
newSolver =
rcp(new Belos::MinresSolMgr<double, MV, OP>(
rcp(&problem, false),
rcp(&belosList, false)
));
} else if (action.compare("solve_gmres") == 0) {
newSolver =
rcp(new Belos::PseudoBlockGmresSolMgr<double, MV, OP>(
rcp(&problem, false),
rcp(&belosList, false)
));
} else {
TEUCHOS_TEST_FOR_EXCEPT_MSG(true, "Unknown solver type \"" << solver << "\".");
}
// Perform solve
RCP<Teuchos::Time> solveTime =
Teuchos::TimeMonitor::getNewTimer("Linear system solve");
{
Teuchos::TimeMonitor tm(*solveTime);
Belos::ReturnType ret = newSolver->solve();
success = ret == Belos::Converged;
}
// *out << newSolver->getNumIters() << std::endl;
// // Compute actual residuals.
// bool badRes = false;
// Teuchos::Array<double> actual_resids(1);
// Teuchos::Array<double> rhs_norm(1);
// Epetra_Vector resid(keoMatrix->OperatorRangeMap());
// OPT::Apply(*keoMatrix, *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<1; i++) {
// double actRes = actual_resids[i]/rhs_norm[i];
// std::cout << "Problem " << i << " : \t" << actRes << std::endl;
// if (actRes > 1.0e-10) badRes = true;
// }
// }
}
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(true, *out, success);
return success ? EXIT_SUCCESS : EXIT_FAILURE;
}
int
main (int argc, char *argv[])
{
using Teuchos::RCP;
using Teuchos::rcp;
using std::cerr;
using std::cout;
using std::endl;
// Anasazi solvers have the following template parameters:
//
// - Scalar: The type of dot product results.
// - MV: The type of (multi)vectors.
// - OP: The type of operators (functions from multivector to
// multivector). A matrix (like Epetra_CrsMatrix) is an example
// of an operator; an Ifpack preconditioner is another example.
//
// Here, Scalar is double, MV is Epetra_MultiVector, and OP is
// Epetra_Operator.
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Anasazi::MultiVecTraits<double, MV> MVT;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init (&argc, &argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif // EPETRA_MPI
const int MyPID = Comm.MyPID ();
//
// Set up the test problem
//
// Dimensionality of the spatial domain to discretize
const int space_dim = 2;
// Size of each of the dimensions of the (discrete) domain
std::vector<double> brick_dim (space_dim);
brick_dim[0] = 1.0;
brick_dim[1] = 1.0;
// Number of elements in each of the dimensions of the domain
std::vector<int> elements (space_dim);
elements[0] = 10;
elements[1] = 10;
// Create the test problem.
RCP<ModalProblem> testCase =
rcp (new ModeLaplace2DQ2 (Comm, brick_dim[0], elements[0],
brick_dim[1], elements[1]));
// Get the stiffness and mass matrices.
//
// rcp (T*, false) returns a nonowning (doesn't deallocate) RCP.
RCP<Epetra_CrsMatrix> K =
rcp (const_cast<Epetra_CrsMatrix* > (testCase->getStiffness ()), false);
RCP<Epetra_CrsMatrix> M =
rcp (const_cast<Epetra_CrsMatrix* > (testCase->getMass ()), false);
//
// Create linear solver for linear systems with K
//
// Anasazi uses shift and invert, with a "shift" of zero, to find
// the eigenvalues of least magnitude. In this example, we
// implement the "invert" part of shift and invert by using an
// Amesos direct solver.
//
// Create Epetra linear problem class for solving linear systems
// with K. This implements the inverse operator for shift and
// invert.
Epetra_LinearProblem AmesosProblem;
// Tell the linear problem about the matrix K. Epetra_LinearProblem
// doesn't know about RCP, so we have to give it a raw pointer.
AmesosProblem.SetOperator (K.get ());
// Create Amesos factory and solver for solving linear systems with
// K. The solver uses the KLU library to do a sparse LU
// factorization.
//
// Note that the AmesosProblem object "absorbs" K. Anasazi doesn't
// see K, just the operator that implements K^{-1} M.
Amesos amesosFactory;
RCP<Amesos_BaseSolver> AmesosSolver;
// Amesos can interface to many different solvers. The following
// loop picks a solver that Amesos supports. The loop order
// reflects solver preference, only in the sense that using LAPACK
// here is a suboptimal fall-back. (With the LAPACK option, Amesos
// makes a dense version of the sparse matrix and uses LAPACK to
// compute the factorization. The other options are true sparse
// direct factorizations.)
const int numSolverNames = 9;
const char* solverNames[9] = {
"Klu", "Umfpack", "Superlu", "Superludist", "Mumps",
"Paradiso", "Taucs", "CSparse", "Lapack"
};
for (int k = 0; k < numSolverNames; ++k) {
const char* const solverName = solverNames[k];
if (amesosFactory.Query (solverName)) {
AmesosSolver = rcp (amesosFactory.Create (solverName, AmesosProblem));
if (MyPID == 0) {
cout << "Amesos solver: \"" << solverName << "\"" << endl;
}
}
}
if (AmesosSolver.is_null ()) {
throw std::runtime_error ("Amesos appears not to have any solvers enabled.");
}
// The AmesosGenOp class assumes that the symbolic and numeric
// factorizations have already been performed on the linear problem.
AmesosSolver->SymbolicFactorization ();
AmesosSolver->NumericFactorization ();
//
// Set parameters for the block Krylov-Schur eigensolver
//
double tol = 1.0e-8; // convergence tolerance
int nev = 10; // number of eigenvalues for which to solve
int blockSize = 3; // block size (number of eigenvectors processed at once)
int numBlocks = 3 * nev / blockSize; // restart length
int maxRestarts = 5; // maximum number of restart cycles
// We're looking for the largest-magnitude eigenvalues of the
// _inverse_ operator, thus, the smallest-magnitude eigenvalues of
// the original operator.
std::string which = "LM";
int verbosity = Anasazi::Errors + Anasazi::Warnings + Anasazi::FinalSummary;
// Create ParameterList to pass into eigensolver
Teuchos::ParameterList MyPL;
MyPL.set ("Verbosity", verbosity);
MyPL.set ("Which", which);
MyPL.set ("Block Size", blockSize);
MyPL.set ("Num Blocks", numBlocks);
MyPL.set ("Maximum Restarts", maxRestarts);
MyPL.set ("Convergence Tolerance", tol);
// Create an initial set of vectors to start the eigensolver. Note:
// This needs to have the same number of columns as the block size.
RCP<MV> ivec = rcp (new MV (K->Map (), blockSize));
ivec->Random ();
// Create the Epetra_Operator for the spectral transformation using
// the Amesos direct solver.
//
// The AmesosGenOp object is the operator we give to Anasazi. Thus,
// Anasazi just sees an operator that computes y = K^{-1} M x. The
// matrix K got absorbed into AmesosProblem (the
// Epetra_LinearProblem object). Later, when we set up the Anasazi
// eigensolver, we will need to tell it about M, so that it can
// orthogonalize basis vectors with respect to the inner product
// defined by M (since it is symmetric positive definite).
RCP<AmesosGenOp> Aop = rcp (new AmesosGenOp (AmesosSolver, M));
// Create the eigenproblem. This object holds all the stuff about
// your problem that Anasazi will see.
//
// Anasazi only needs M so that it can orthogonalize basis vectors
// with respect to the M inner product. Wouldn't it be nice if
// Anasazi didn't require M in two different places? Alas, this is
// not currently the case.
RCP<Anasazi::BasicEigenproblem<double,MV,OP> > MyProblem =
rcp (new Anasazi::BasicEigenproblem<double,MV,OP> (Aop, M, ivec));
// Tell the eigenproblem that the matrix pencil (K,M) is symmetric.
MyProblem->setHermitian (true);
// Set the number of eigenvalues requested
MyProblem->setNEV (nev);
// Tell the eigenproblem that you are finished passing it information.
const bool boolret = MyProblem->setProblem ();
if (boolret != true) {
if (MyPID == 0) {
cerr << "Anasazi::BasicEigenproblem::setProblem() returned with error." << endl;
}
#ifdef EPETRA_MPI
MPI_Finalize ();
#endif // EPETRA_MPI
return -1;
}
// Create the Block Krylov-Schur eigensolver.
Anasazi::BlockKrylovSchurSolMgr<double, MV, OP> MySolverMgr (MyProblem, MyPL);
// Solve the eigenvalue problem.
//
// Note that creating the eigensolver is separate from solving it.
// After creating the eigensolver, you may call solve() multiple
// times with different parameters or initial vectors. This lets
// you reuse intermediate state, like allocated basis vectors.
Anasazi::ReturnType returnCode = MySolverMgr.solve ();
if (returnCode != Anasazi::Converged && MyPID == 0) {
cout << "Anasazi eigensolver did not converge." << endl;
}
// Get the eigenvalues and eigenvectors from the eigenproblem.
Anasazi::Eigensolution<double,MV> sol = MyProblem->getSolution ();
// Anasazi returns eigenvalues as Anasazi::Value, so that if
// Anasazi's Scalar type is real-valued (as it is in this case), but
// some eigenvalues are complex, you can still access the
// eigenvalues correctly. In this case, there are no complex
// eigenvalues, since the matrix pencil is symmetric.
std::vector<Anasazi::Value<double> > evals = sol.Evals;
RCP<MV> evecs = sol.Evecs;
int numev = sol.numVecs;
if (numev > 0) {
// Reconstruct the eigenvalues. The ones that Anasazi gave back
// are the inverses of the original eigenvalues. Reconstruct the
// eigenvectors too.
MV tempvec (K->Map (), MVT::GetNumberVecs (*evecs));
K->Apply (*evecs, tempvec);
Teuchos::SerialDenseMatrix<int,double> dmatr (numev, numev);
MVT::MvTransMv (1.0, tempvec, *evecs, dmatr);
if (MyPID == 0) {
double compeval = 0.0;
cout.setf (std::ios_base::right, std::ios_base::adjustfield);
cout << "Actual Eigenvalues (obtained by Rayleigh quotient) : " << endl;
cout << "------------------------------------------------------" << endl;
cout << std::setw(16) << "Real Part"
<< std::setw(16) << "Rayleigh Error" << endl;
cout << "------------------------------------------------------" << endl;
for (int i = 0; i < numev; ++i) {
compeval = dmatr(i,i);
cout << std::setw(16) << compeval
<< std::setw(16)
<< std::fabs (compeval - 1.0/evals[i].realpart)
<< endl;
}
cout << "------------------------------------------------------" << endl;
}
}
#ifdef EPETRA_MPI
MPI_Finalize ();
#endif // EPETRA_MPI
return 0;
}
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[]) {
#include "MueLu_UseShortNames.hpp"
using Teuchos::RCP;
using Teuchos::rcp;
using namespace MueLuTests;
using namespace Teuchos;
typedef Xpetra::StridedMap<int,int> StridedMap;
typedef Xpetra::StridedMapFactory<int,int> StridedMapFactory;
oblackholestream blackhole;
GlobalMPISession mpiSession(&argc,&argv,&blackhole);
bool success = false;
bool verbose = true;
try {
RCP<const Comm<int> > comm = DefaultComm<int>::getComm();
RCP<FancyOStream> out = fancyOStream(rcpFromRef(std::cout));
out->setOutputToRootOnly(0);
*out << MueLu::MemUtils::PrintMemoryUsage() << std::endl;
// Timing
Time myTime("global");
TimeMonitor MM(myTime);
#ifndef HAVE_XPETRA_INT_LONG_LONG
*out << "Warning: scaling test was not compiled with long long int support" << std::endl;
#endif
// read in input parameters
// default parameters
LO BS_nSweeps = 100;
Scalar BS_omega = 1.7;
LO SC_nSweeps = 1;
Scalar SC_omega = 1.0;
int SC_bUseDirectSolver = 0;
// Note: use --help to list available options.
CommandLineProcessor clp(false);
clp.setOption("BraessSarazin_sweeps",&BS_nSweeps,"number of sweeps with BraessSarazin smoother");
clp.setOption("BraessSarazin_omega", &BS_omega, "scaling factor for BraessSarazin smoother");
clp.setOption("SchurComp_sweeps", &SC_nSweeps,"number of sweeps for BraessSarazin internal SchurComp solver/smoother (GaussSeidel)");
clp.setOption("SchurComp_omega", &SC_omega, "damping parameter for BraessSarazin internal SchurComp solver/smoother (GaussSeidel)");
clp.setOption("SchurComp_solver", &SC_bUseDirectSolver, "if 1: use direct solver for SchurComp equation, otherwise use GaussSeidel smoother (=default)");
switch (clp.parse(argc,argv)) {
case CommandLineProcessor::PARSE_HELP_PRINTED: return EXIT_SUCCESS; break;
case CommandLineProcessor::PARSE_ERROR:
case CommandLineProcessor::PARSE_UNRECOGNIZED_OPTION: return EXIT_FAILURE; break;
case CommandLineProcessor::PARSE_SUCCESSFUL: break;
}
int globalNumDofs = 1500; // used for the maps
//int nDofsPerNode = 3; // used for generating the fine level null-space
// build strided maps
// striding information: 2 velocity dofs and 1 pressure dof = 3 dofs per node
std::vector<size_t> stridingInfo;
stridingInfo.push_back(2);
stridingInfo.push_back(1);
/////////////////////////////////////// build strided maps
// build strided maps:
// xstridedfullmap: full map (velocity and pressure dof gids), continous
// xstridedvelmap: only velocity dof gid maps (i.e. 0,1,3,4,6,7...)
// xstridedpremap: only pressure dof gid maps (i.e. 2,5,8,...)
Xpetra::UnderlyingLib lib = Xpetra::UseEpetra;
RCP<StridedMap> xstridedfullmap = StridedMapFactory::Build(lib,globalNumDofs,0,stridingInfo,comm,-1);
RCP<StridedMap> xstridedvelmap = StridedMapFactory::Build(xstridedfullmap,0);
RCP<StridedMap> xstridedpremap = StridedMapFactory::Build(xstridedfullmap,1);
/////////////////////////////////////// transform Xpetra::Map objects to Epetra
// this is needed for our splitting routine
const RCP<const Epetra_Map> fullmap = rcpFromRef(Xpetra::toEpetra(*xstridedfullmap));
RCP<const Epetra_Map> velmap = rcpFromRef(Xpetra::toEpetra(*xstridedvelmap));
RCP<const Epetra_Map> premap = rcpFromRef(Xpetra::toEpetra(*xstridedpremap));
/////////////////////////////////////// import problem matrix and RHS from files (-> Epetra)
// read in problem
Epetra_CrsMatrix * ptrA = 0;
Epetra_Vector * ptrf = 0;
Epetra_MultiVector* ptrNS = 0;
*out << "Reading matrix market file" << std::endl;
EpetraExt::MatrixMarketFileToCrsMatrix("A_re1000_5932.txt",*fullmap,*fullmap,*fullmap,ptrA);
EpetraExt::MatrixMarketFileToVector("b_re1000_5932.txt",*fullmap,ptrf);
RCP<Epetra_CrsMatrix> epA = rcp(ptrA);
RCP<Epetra_Vector> epv = rcp(ptrf);
RCP<Epetra_MultiVector> epNS = rcp(ptrNS);
/////////////////////////////////////// split system into 2x2 block system
*out << "Split matrix into 2x2 block matrix" << std::endl;
// split fullA into A11,..., A22
RCP<Epetra_CrsMatrix> A11;
RCP<Epetra_CrsMatrix> A12;
RCP<Epetra_CrsMatrix> A21;
RCP<Epetra_CrsMatrix> A22;
if(SplitMatrix2x2(epA,*velmap,*premap,A11,A12,A21,A22)==false)
*out << "Problem with splitting matrix"<< std::endl;
/////////////////////////////////////// transform Epetra objects to Xpetra (needed for MueLu)
// build Xpetra objects from Epetra_CrsMatrix objects
RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA11 = rcp(new Xpetra::EpetraCrsMatrix(A11));
RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA12 = rcp(new Xpetra::EpetraCrsMatrix(A12));
RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA21 = rcp(new Xpetra::EpetraCrsMatrix(A21));
RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA22 = rcp(new Xpetra::EpetraCrsMatrix(A22));
/////////////////////////////////////// generate MapExtractor object
std::vector<RCP<const Xpetra::Map<LO,GO,Node> > > xmaps;
xmaps.push_back(xstridedvelmap);
xmaps.push_back(xstridedpremap);
RCP<const Xpetra::MapExtractor<Scalar,LO,GO,Node> > map_extractor = Xpetra::MapExtractorFactory<Scalar,LO,GO>::Build(xstridedfullmap,xmaps);
/////////////////////////////////////// build blocked transfer operator
// using the map extractor
RCP<Xpetra::BlockedCrsMatrix<Scalar,LO,GO,Node> > bOp = rcp(new Xpetra::BlockedCrsMatrix<Scalar,LO,GO>(map_extractor,map_extractor,10));
bOp->setMatrix(0,0,xA11);
bOp->setMatrix(0,1,xA12);
bOp->setMatrix(1,0,xA21);
bOp->setMatrix(1,1,xA22);
bOp->fillComplete();
//////////////////////////////////////////////////////// finest Level
RCP<MueLu::Level> Finest = rcp(new Level());
Finest->setDefaultVerbLevel(VERB_NONE);
Finest->Set("A",rcp_dynamic_cast<Matrix>(bOp));
///////////////////////////////////
// Test Braess Sarazin Smoother as a solver
*out << "Test: Creating Braess Sarazin Smoother" << std::endl;
*out << "Test: Omega for BraessSarazin = " << BS_omega << std::endl;
*out << "Test: Number of sweeps for BraessSarazin = " << BS_nSweeps << std::endl;
*out << "Test: Omega for Schur Complement solver= " << SC_omega << std::endl;
*out << "Test: Number of Schur Complement solver= " << SC_nSweeps << std::endl;
*out << "Test: Setting up Braess Sarazin Smoother" << std::endl;
// define BraessSarazin Smoother with BS_nSweeps and BS_omega as scaling factor
// AFact_ = null (= default) for the 2x2 blocked operator
RCP<BraessSarazinSmoother> BraessSarazinSm = rcp( new BraessSarazinSmoother() );
BraessSarazinSm->SetParameter("Sweeps", Teuchos::ParameterEntry(BS_nSweeps));
BraessSarazinSm->SetParameter("Damping factor", Teuchos::ParameterEntry(BS_omega));
RCP<SmootherFactory> smootherFact = rcp( new SmootherFactory(BraessSarazinSm) );
/*note that omega must be the same in the SchurComplementFactory and in the BraessSarazinSmoother*/
// define SchurComplement Factory
// SchurComp gets a RCP to AFact_ which has to be the 2x2 blocked operator
// and the scaling/damping factor omega that is used for BraessSarazin
// It stores the resulting SchurComplement operator as "A" generated by the SchurComplementFactory
// Instead of F^{-1} it uses the approximation \hat{F}^{-1} with \hat{F} = diag(F)
RCP<SchurComplementFactory> SFact = rcp(new SchurComplementFactory());
SFact->SetParameter("omega", ParameterEntry(BS_omega));
SFact->SetFactory("A",MueLu::NoFactory::getRCP());
// define smoother/solver for BraessSarazin
RCP<SmootherPrototype> smoProtoSC = null;
if(SC_bUseDirectSolver != 1) {
//Smoother Factory, using SFact as a factory for A
std::string ifpackSCType;
ParameterList ifpackSCList;
ifpackSCList.set("relaxation: sweeps", SC_nSweeps );
ifpackSCList.set("relaxation: damping factor", SC_omega );
ifpackSCType = "RELAXATION";
ifpackSCList.set("relaxation: type", "Gauss-Seidel");
smoProtoSC = rcp( new TrilinosSmoother(ifpackSCType, ifpackSCList, 0) );
smoProtoSC->SetFactory("A", SFact);
}
else {
ParameterList ifpackDSList;
std::string ifpackDSType;
smoProtoSC = rcp( new DirectSolver(ifpackDSType,ifpackDSList) ); smoProtoSC->SetFactory("A", SFact);
}
RCP<SmootherFactory> SmooSCFact = rcp( new SmootherFactory(smoProtoSC) );
// define temporary FactoryManager that is used as input for BraessSarazin smoother
RCP<FactoryManager> MB = rcp(new FactoryManager());
MB->SetFactory("A", SFact); // SchurComplement operator for correction step (defined as "A")
MB->SetFactory("Smoother", SmooSCFact); // solver/smoother for correction step
MB->SetFactory("PreSmoother", SmooSCFact);
MB->SetFactory("PostSmoother", SmooSCFact);
MB->SetIgnoreUserData(true); // always use data from factories defined in factory manager
BraessSarazinSm->AddFactoryManager(MB,0); // set temporary factory manager in BraessSarazin smoother
// setup main factory manager
RCP<FactoryManager> M = rcp(new FactoryManager());
M->SetFactory("A", MueLu::NoFactory::getRCP()); // this is the 2x2 blocked operator
M->SetFactory("Smoother", smootherFact); // BraessSarazin block smoother
M->SetFactory("PreSmoother", smootherFact);
M->SetFactory("PostSmoother", smootherFact);
MueLu::SetFactoryManager SFMCoarse(Finest, M);
Finest->Request(MueLu::TopSmootherFactory<Scalar,LocalOrdinal,GlobalOrdinal,Node>(M, "Smoother"));
// call setup (= extract blocks and extract diagonal of F)
BraessSarazinSm->Setup(*Finest);
RCP<MultiVector> xtest = MultiVectorFactory::Build(xstridedfullmap,1);
xtest->putScalar( (SC) 0.0);
RCP<Vector> xR = rcp(new Xpetra::EpetraVector(epv));
// calculate initial (absolute) residual
Array<ScalarTraits<SC>::magnitudeType> norms(1);
xR->norm2(norms);
*out << "Test: ||x_0|| = " << norms[0] << std::endl;
*out << "Test: Applying Braess-Sarazin Smoother" << std::endl;
*out << "Test: START DATA" << std::endl;
*out << "iterations\tVelocity_residual\tPressure_residual" << std::endl;
BraessSarazinSm->Apply(*xtest,*xR);
xtest->norm2(norms);
*out << "Test: ||x_1|| = " << norms[0] << std::endl;
Array<ScalarTraits<double>::magnitudeType> test = MueLu::Utils<double, int, int>::ResidualNorm(*bOp, *xtest, *xR);
*out << "residual norm: " << test[0] << std::endl;
success = true;
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
return ( success ? EXIT_SUCCESS : EXIT_FAILURE );
}
int main(int argc, char *argv[])
{
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm Comm;
#endif
// My MPI process rank.
const int MyPID = Comm.MyPID();
// "/Users/sakashitatatsuya/Downloads/barista_trunk_slepc/sample/hamiltonian_matrix.ip"
std::ifstream ifs(argv[1]);
alps::Parameters params(ifs);
Teuchos::oblackholestream blackHole;
std::ostream& out = (MyPID == 0) ? std::cout : blackHole;
barista::Hamiltonian<> hamiltonian(params);
matrix_type matrix(hamiltonian.dimension(), hamiltonian.dimension());
hamiltonian.fill<double>(matrix);
int m,n;
int N;
m = n = N = hamiltonian.dimension();
//std::cout << matrix << std::endl;
std::ofstream ofs;
if (MyPID==0) {
ofs.open("anasazi_time.txt");
if (!ofs) {
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return -1;
}
}
//Teuchos::ParameterList GaleriList;
using Teuchos::RCP;
using Teuchos::rcp;
typedef Teuchos::ScalarTraits<double> STS;
const double one = STS::one();
const double zero = STS::zero();
// The problem is defined on a 2D grid, global size is nx * nx.
//int nx = N;
//GaleriList.set("n", nx * nx);
//GaleriList.set("nx", nx);
//GaleriList.set("ny", nx);
//Teuchos::RCP<Epetra_Map> Map = Teuchos::rcp( Galeri::CreateMap("Linear", Comm, GaleriList) );
//Teuchos::RCP<Epetra_RowMatrix> A = Teuchos::rcp( Galeri::CreateCrsMatrix("Laplace2D", &*Map, GaleriList) );
// Construct a Map that puts approximately the same number of rows
// of the matrix A on each processor.
Epetra_Map RowMap (N, 0, Comm);
Epetra_Map ColMap (N, 0, Comm);
// Get update list and number of local equations from newly created Map.
const int NumMyRowElements = RowMap.NumMyElements ();
std::vector<int> MyGlobalRowElements (NumMyRowElements);
RowMap.MyGlobalElements (&MyGlobalRowElements[0]);
// Create an Epetra_CrsMatrix using the given row map.
RCP<Epetra_CrsMatrix> A = rcp (new Epetra_CrsMatrix (Copy, RowMap, n));
// We use info to catch any errors that may have happened during // matrix assembly, and report them globally. We do this so that // the MPI processes won't call FillComplete() unless they all // successfully filled their parts of the matrix.
int info = 0;
try {
//
// Compute coefficients for the discrete integral operator.
//
std::vector<double> Values (n);
std::vector<int> Indices (n);
//const double inv_mp1 = one / (m+1);
//const double inv_np1 = one / (n+1);
int count;
//for (int i = 0; i < n; ++i) {
// Indices[i] = i;
//}
for (int i = 0; i < NumMyRowElements; ++i) {
count =0;
for (int j = 0; j < n; ++j) {
if (matrix(MyGlobalRowElements[i],j)!=0) {
Values[count] = matrix(MyGlobalRowElements[i],j);
Indices[count] = j;
count++;
}
}
info = A->InsertGlobalValues (MyGlobalRowElements[i], count,
&Values[0], &Indices[0]);
// Make sure that the insertion succeeded. Teuchos'
// TEST_FOR_EXCEPTION macro gives a nice error message if the
// thrown exception isn't caught. We'll report this on the
// offending MPI process.
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error, "Failed to insert n="
<< n << " global value" << (n != 1 ? "s" : "")
<< " in row " << MyGlobalRowElements[i]
<< " of the matrix." );
*/
} // for i = 0...
// Call FillComplete on the matrix. Since the matrix isn't square,
// we have to give FillComplete the domain and range maps, which in
// this case are the column resp. row maps.
info = A->FillComplete (ColMap, RowMap);
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error,
"FillComplete failed with INFO = " << info << ".");
*/
info = A->OptimizeStorage();
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error,
"OptimizeStorage failed with INFO = " << info << ".");
*/
} catch (std::runtime_error& e) {
// If multiple MPI processes are reporting errors, sometimes
// forming the error message as a string and then writing it to
// the output stream prevents messages from different processes
// from being interleaved.
std::ostringstream os;
os << "*** Error on MPI process " << MyPID << ": " << e.what();
cerr << os.str() << endl;
if (info == 0)
info = -1; // All procs will share info later on.
}
// Variables used for the Block Davidson Method
const int nev = 5;
const int blockSize = 5;
const int numBlocks = 8;
const int maxRestarts = 500;
const double tol = 1.0e-8;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Anasazi::MultiVecTraits<double, Epetra_MultiVector> MVT;
// Create an Epetra_MultiVector for an initial vector to start the solver.
// Note: This needs to have the same number of columns as the blocksize.
//
//Teuchos::RCP<Epetra_MultiVector> ivec = Teuchos::rcp( new Epetra_MultiVector(*Map, blockSize) );
Teuchos::RCP<Epetra_MultiVector> ivec = Teuchos::rcp( new Epetra_MultiVector(ColMap, blockSize) );
ivec->Random();
// Create the eigenproblem.
Teuchos::RCP<Anasazi::BasicEigenproblem<double, MV, OP> > problem =
Teuchos::rcp( new Anasazi::BasicEigenproblem<double, MV, OP>(A, ivec) );
// Inform the eigenproblem that the operator A is symmetric
problem->setHermitian(true);
// Set the number of eigenvalues requested
problem->setNEV( nev );
// Inform the eigenproblem that you are finishing passing it information
bool boolret = problem->setProblem();
if (boolret != true) {
std::cout<<"Anasazi::BasicEigenproblem::setProblem() returned an error." << std::endl;
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return -1;
}
// Create parameter list to pass into the solver manager
Teuchos::ParameterList anasaziPL;
anasaziPL.set( "Which", "LM" );
anasaziPL.set( "Block Size", blockSize );
anasaziPL.set( "Maximum Iterations", 500 );
anasaziPL.set( "Convergence Tolerance", tol );
anasaziPL.set( "Verbosity", Anasazi::Errors+Anasazi::Warnings+Anasazi::TimingDetails+Anasazi::FinalSummary );
// Create the solver manager
Anasazi::LOBPCGSolMgr<double, MV, OP> anasaziSolver(problem, anasaziPL);
// Solve the problem
double start, end;
MPI_Barrier(MPI_COMM_WORLD);
start = MPI_Wtime();
Anasazi::ReturnType returnCode = anasaziSolver.solve();
MPI_Barrier(MPI_COMM_WORLD);
end = MPI_Wtime();
// Get the eigenvalues and eigenvectors from the eigenproblem
Anasazi::Eigensolution<double,MV> sol = problem->getSolution();
std::vector<Anasazi::Value<double> > evals = sol.Evals;
Teuchos::RCP<MV> evecs = sol.Evecs;
// Compute residuals.
std::vector<double> normR(sol.numVecs);
Teuchos::SerialDenseMatrix<int,double> T(sol.numVecs, sol.numVecs);
Epetra_MultiVector tempAevec( ColMap, sol.numVecs );
T.putScalar(0.0);
for (int i=0; i<sol.numVecs; i++) {
T(i,i) = evals[i].realpart;
}
A->Apply( *evecs, tempAevec );
MVT::MvTimesMatAddMv( -1.0, *evecs, T, 1.0, tempAevec );
MVT::MvNorm( tempAevec, normR );
if (MyPID == 0) {
// Print the results
std::cout<<"Solver manager returned " << (returnCode == Anasazi::Converged ? "converged." : "unconverged.") << std::endl;
std::cout<<std::endl;
std::cout<<"------------------------------------------------------"<<std::endl;
std::cout<<std::setw(16)<<"Eigenvalue"
<<std::setw(18)<<"Direct Residual"
<<std::endl;
std::cout<<"------------------------------------------------------"<<std::endl;
for (int i=0; i<sol.numVecs; i++) {
std::cout<<std::setw(16)<<evals[i].realpart
<<std::setw(18)<<normR[i]/evals[i].realpart
<<std::endl;
}
std::cout<<"------------------------------------------------------"<<std::endl;
}
// Print out the map and matrices
//ColMap.Print (out);
//A->Print (cout);
//RowMap.Print (cout);
double time;
int iter;
if (MyPID==0) {
iter = anasaziSolver.getNumIters();
Teuchos::Array<Teuchos::RCP<Teuchos::Time> > timer = anasaziSolver.getTimers();
Teuchos::RCP<Teuchos::Time> _timerSolve = timer[0];
cout << "timerSolve=" << _timerSolve << endl;
time = end - start;
cout << "time=" << time << endl;
ofs << "time=" << time << endl;
cout << "iter=" << iter << endl;
ofs << "iter=" << iter << endl;
}
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return 0;
}
int main(int argc, char *argv[]) {
#if defined(HAVE_MUELU_EPETRA)
typedef double Scalar;
typedef int LocalOrdinal;
typedef int GlobalOrdinal;
typedef LocalOrdinal LO;
typedef GlobalOrdinal GO;
typedef Xpetra::EpetraNode Node;
#include "MueLu_UseShortNames.hpp"
using Teuchos::RCP;
using Teuchos::rcp;
using namespace MueLuTests;
Teuchos::oblackholestream blackhole;
Teuchos::GlobalMPISession mpiSession(&argc,&argv,&blackhole);
bool success = false;
bool verbose = true;
try {
RCP<const Teuchos::Comm<int> > comm = Teuchos::DefaultComm<int>::getComm();
RCP<Teuchos::FancyOStream> out = Teuchos::fancyOStream(Teuchos::rcpFromRef(std::cout));
out->setOutputToRootOnly(0);
*out << MueLu::MemUtils::PrintMemoryUsage() << std::endl;
// Timing
Teuchos::Time myTime("global");
Teuchos::TimeMonitor MM(myTime);
// read in input parameters
// default parameters
LO SIMPLE_nSweeps = 600;
Scalar SIMPLE_omega = 0.5;
LO SC_nSweeps = 10;
Scalar SC_omega = 1.0;
LO PRED_nSweeps = 3;
Scalar PRED_omega = 1.0;
LO useSIMPLEC = 0;
int SC_bUseDirectSolver = 0;
// Note: use --help to list available options.
Teuchos::CommandLineProcessor clp(false);
clp.setOption("SIMPLE_sweeps",&SIMPLE_nSweeps,"number of sweeps with SIMPLE smoother");
clp.setOption("SIMPLE_omega", &SIMPLE_omega, "scaling factor for SIMPLE smoother");
clp.setOption("Predict_sweeps", &PRED_nSweeps, "number of sweeps for SIMPLE internal velocity prediction smoother (GaussSeidel)");
clp.setOption("Predict_omega", &PRED_omega, "damping parameter for SIMPLE internal velocity prediction smoother (GaussSeidel)");
clp.setOption("SchurComp_sweeps", &SC_nSweeps,"number of sweeps for SIMPLE internal SchurComp solver/smoother (GaussSeidel)");
clp.setOption("SchurComp_omega", &SC_omega, "damping parameter for SIMPLE internal SchurComp solver/smoother (GaussSeidel)");
clp.setOption("SchurComp_solver", &SC_bUseDirectSolver, "if 1: use direct solver for SchurComp equation, otherwise use GaussSeidel smoother");
clp.setOption("useSIMPLEC", &useSIMPLEC, "if 1: use SIMPLEC instead of SIMPLE (default = 0 (SIMPLE))");
switch (clp.parse(argc,argv)) {
case Teuchos::CommandLineProcessor::PARSE_HELP_PRINTED: return EXIT_SUCCESS; break;
case Teuchos::CommandLineProcessor::PARSE_ERROR:
case Teuchos::CommandLineProcessor::PARSE_UNRECOGNIZED_OPTION: return EXIT_FAILURE; break;
case Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL: break;
}
int globalNumDofs = 1500; // used for the maps
// build strided maps
// striding information: 2 velocity dofs and 1 pressure dof = 3 dofs per node
std::vector<size_t> stridingInfo;
stridingInfo.push_back(2);
stridingInfo.push_back(1);
/////////////////////////////////////// build strided maps
// build strided maps:
// xstridedfullmap: full map (velocity and pressure dof gids), continous
// xstridedvelmap: only velocity dof gid maps (i.e. 0,1,3,4,6,7...)
// xstridedpremap: only pressure dof gid maps (i.e. 2,5,8,...)
Xpetra::UnderlyingLib lib = Xpetra::UseEpetra;
RCP<const StridedMap> xstridedfullmap = StridedMapFactory::Build(lib,globalNumDofs,0,stridingInfo,comm,-1);
RCP<const StridedMap> xstridedvelmap = StridedMapFactory::Build(xstridedfullmap,0);
RCP<const StridedMap> xstridedpremap = StridedMapFactory::Build(xstridedfullmap,1);
/////////////////////////////////////// transform Xpetra::Map objects to Epetra
// this is needed for our splitting routine
const RCP<const Epetra_Map> fullmap = Teuchos::rcpFromRef(Xpetra::toEpetra(*xstridedfullmap));
RCP<const Epetra_Map> velmap = Teuchos::rcpFromRef(Xpetra::toEpetra(*xstridedvelmap));
RCP<const Epetra_Map> premap = Teuchos::rcpFromRef(Xpetra::toEpetra(*xstridedpremap));
/////////////////////////////////////// import problem matrix and RHS from files (-> Epetra)
// read in problem
Epetra_CrsMatrix * ptrA = 0;
Epetra_Vector * ptrf = 0;
Epetra_MultiVector* ptrNS = 0;
*out << "Reading matrix market file" << std::endl;
EpetraExt::MatrixMarketFileToCrsMatrix("A_re1000_5932.txt",*fullmap,*fullmap,*fullmap,ptrA);
EpetraExt::MatrixMarketFileToVector("b_re1000_5932.txt",*fullmap,ptrf);
RCP<Epetra_CrsMatrix> epA = Teuchos::rcp(ptrA);
RCP<Epetra_Vector> epv = Teuchos::rcp(ptrf);
RCP<Epetra_MultiVector> epNS = Teuchos::rcp(ptrNS);
/////////////////////////////////////// split system into 2x2 block system
*out << "Split matrix into 2x2 block matrix" << std::endl;
// split fullA into A11,..., A22
Teuchos::RCP<Epetra_CrsMatrix> A11;
Teuchos::RCP<Epetra_CrsMatrix> A12;
Teuchos::RCP<Epetra_CrsMatrix> A21;
Teuchos::RCP<Epetra_CrsMatrix> A22;
if(SplitMatrix2x2(epA,*velmap,*premap,A11,A12,A21,A22)==false)
*out << "Problem with splitting matrix"<< std::endl;
/////////////////////////////////////// transform Epetra objects to Xpetra (needed for MueLu)
// build Xpetra objects from Epetra_CrsMatrix objects
Teuchos::RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA11 = Teuchos::rcp(new Xpetra::EpetraCrsMatrixT<GO,Node>(A11));
Teuchos::RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA12 = Teuchos::rcp(new Xpetra::EpetraCrsMatrixT<GO,Node>(A12));
Teuchos::RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA21 = Teuchos::rcp(new Xpetra::EpetraCrsMatrixT<GO,Node>(A21));
Teuchos::RCP<Xpetra::CrsMatrix<Scalar,LO,GO,Node> > xA22 = Teuchos::rcp(new Xpetra::EpetraCrsMatrixT<GO,Node>(A22));
/////////////////////////////////////// generate MapExtractor object
std::vector<Teuchos::RCP<const Xpetra::Map<LO,GO,Node> > > xmaps;
xmaps.push_back(xstridedvelmap);
xmaps.push_back(xstridedpremap);
Teuchos::RCP<const Xpetra::MapExtractor<Scalar,LO,GO,Node> > map_extractor = Xpetra::MapExtractorFactory<Scalar,LO,GO,Node>::Build(xstridedfullmap,xmaps);
/////////////////////////////////////// build blocked transfer operator
// using the map extractor
Teuchos::RCP<Xpetra::BlockedCrsMatrix<Scalar,LO,GO,Node> > bOp = Teuchos::rcp(new Xpetra::BlockedCrsMatrix<Scalar,LO,GO,Node>(map_extractor,map_extractor,10));
bOp->setMatrix(0,0,Teuchos::rcp(new Xpetra::CrsMatrixWrap<Scalar,LocalOrdinal,GlobalOrdinal,Node>(xA11)));
bOp->setMatrix(0,1,Teuchos::rcp(new Xpetra::CrsMatrixWrap<Scalar,LocalOrdinal,GlobalOrdinal,Node>(xA12)));
bOp->setMatrix(1,0,Teuchos::rcp(new Xpetra::CrsMatrixWrap<Scalar,LocalOrdinal,GlobalOrdinal,Node>(xA21)));
bOp->setMatrix(1,1,Teuchos::rcp(new Xpetra::CrsMatrixWrap<Scalar,LocalOrdinal,GlobalOrdinal,Node>(xA22)));
bOp->fillComplete();
//////////////////////////////////////////////////////// finest Level
RCP<MueLu::Level> Finest = rcp(new Level());
Finest->setDefaultVerbLevel(Teuchos::VERB_NONE);
Finest->Set("A",Teuchos::rcp_dynamic_cast<Matrix>(bOp));
///////////////////////////////////
// Test Braess Sarazin Smoother as a solver
*out << "Test: Creating SIMPLE Smoother" << std::endl;
*out << "Test: Omega for SIMPLE = " << SIMPLE_omega << std::endl;
*out << "Test: Number of sweeps for SIMPLE = " << SIMPLE_nSweeps << std::endl;
*out << "Test: Omega for Schur Complement solver= " << SC_omega << std::endl;
*out << "Test: Number of Schur Complement solver= " << SC_nSweeps << std::endl;
*out << "Test: Setting up Braess Sarazin Smoother" << std::endl;
// define SIMPLE Smoother with SIMPLE_nSweeps and SIMPLE_omega as scaling factor
// AFact_ = Teuchos::null (= default) for the 2x2 blocked operator
RCP<SimpleSmoother> SimpleSm = rcp( new SimpleSmoother() );
SimpleSm->SetParameter("Sweeps", Teuchos::ParameterEntry(SIMPLE_nSweeps));
SimpleSm->SetParameter("Damping factor", Teuchos::ParameterEntry(SIMPLE_omega));
if(useSIMPLEC==1) SimpleSm->SetParameter("UseSIMPLEC", Teuchos::ParameterEntry(true));
RCP<SmootherFactory> smootherFact = rcp( new SmootherFactory(SimpleSm) );
// define smoother for velocity prediction
//RCP<SubBlockAFactory> A00Fact = Teuchos::rcp(new SubBlockAFactory(MueLu::NoFactory::getRCP(), 0, 0));
RCP<SubBlockAFactory> A00Fact = rcp(new SubBlockAFactory());
A00Fact->SetFactory("A",MueLu::NoFactory::getRCP());
A00Fact->SetParameter("block row",Teuchos::ParameterEntry(0));
A00Fact->SetParameter("block col",Teuchos::ParameterEntry(0));
RCP<SmootherPrototype> smoProtoPredict = Teuchos::null;
std::string ifpackPredictType;
Teuchos::ParameterList ifpackPredictList;
ifpackPredictList.set("relaxation: sweeps", PRED_nSweeps );
ifpackPredictList.set("relaxation: damping factor", PRED_omega );
ifpackPredictType = "RELAXATION";
ifpackPredictList.set("relaxation: type", "Gauss-Seidel");
smoProtoPredict = rcp( new TrilinosSmoother(ifpackPredictType, ifpackPredictList, 0) );
smoProtoPredict->SetFactory("A", A00Fact);
RCP<SmootherFactory> SmooPredictFact = rcp( new SmootherFactory(smoProtoPredict) );
// define temporary FactoryManager that is used as input for BraessSarazin smoother
RCP<FactoryManager> MPredict = rcp(new FactoryManager());
MPredict->SetFactory("A", A00Fact);
MPredict->SetFactory("Smoother", SmooPredictFact); // solver/smoother for correction step
MPredict->SetFactory("PreSmoother", SmooPredictFact);
MPredict->SetFactory("PostSmoother", SmooPredictFact);
MPredict->SetIgnoreUserData(true); // always use data from factories defined in factory manager
SimpleSm->SetVelocityPredictionFactoryManager(MPredict); // set temporary factory manager in BraessSarazin smoother
// define SchurComplement Factory
// SchurComp gets a RCP to AFact_ which has to be the 2x2 blocked operator
// It stores the resulting SchurComplement operator as "A" generated by the SchurComplementFactory
// Instead of F^{-1} it uses the approximation \hat{F}^{-1} with \hat{F} = diag(F)
RCP<SchurComplementFactory> SFact = Teuchos::rcp(new SchurComplementFactory());
SFact->SetParameter("omega", Teuchos::ParameterEntry(1.0)); // for Simple, omega is always 1.0 in the SchurComplement
if(useSIMPLEC == 1) SFact->SetParameter("lumping", Teuchos::ParameterEntry(true));
else SFact->SetParameter("lumping", Teuchos::ParameterEntry(false));
SFact->SetFactory("A",MueLu::NoFactory::getRCP());
// define smoother/solver for BraessSarazin
RCP<SmootherPrototype> smoProtoSC = Teuchos::null;
if(SC_bUseDirectSolver != 1) {
//Smoother Factory, using SFact as a factory for A
std::string ifpackSCType;
Teuchos::ParameterList ifpackSCList;
ifpackSCList.set("relaxation: sweeps", SC_nSweeps );
ifpackSCList.set("relaxation: damping factor", SC_omega );
ifpackSCType = "RELAXATION";
ifpackSCList.set("relaxation: type", "Gauss-Seidel");
smoProtoSC = rcp( new TrilinosSmoother(ifpackSCType, ifpackSCList, 0) );
smoProtoSC->SetFactory("A",SFact);
}
else {
Teuchos::ParameterList ifpackDSList;
std::string ifpackDSType;
smoProtoSC = rcp( new DirectSolver(ifpackDSType,ifpackDSList) ); smoProtoSC->SetFactory("A", SFact);
}
RCP<SmootherFactory> SmooSCFact = rcp( new SmootherFactory(smoProtoSC) );
// define temporary FactoryManager that is used as input for BraessSarazin smoother
RCP<FactoryManager> MB = rcp(new FactoryManager());
MB->SetFactory("A", SFact); // SchurComplement operator for correction step (defined as "A")
MB->SetFactory("Smoother", SmooSCFact); // solver/smoother for correction step
MB->SetFactory("PreSmoother", SmooSCFact);
MB->SetFactory("PostSmoother", SmooSCFact);
MB->SetIgnoreUserData(true); // always use data from factories defined in factory manager
SimpleSm->SetSchurCompFactoryManager(MB); // set temporary factory manager in BraessSarazin smoother
// setup main factory manager
RCP<FactoryManager> M = rcp(new FactoryManager());
M->SetFactory("A", MueLu::NoFactory::getRCP()); // this is the 2x2 blocked operator
M->SetFactory("Smoother", smootherFact); // BraessSarazin block smoother
M->SetFactory("PreSmoother", smootherFact);
M->SetFactory("PostSmoother", smootherFact);
MueLu::SetFactoryManager SFMCoarse(Finest, M);
Finest->Request(MueLu::TopSmootherFactory<Scalar,LO,GO,Node>(M, "Smoother"));
// call setup (= extract blocks and extract diagonal of F)
SimpleSm->Setup(*Finest);
RCP<MultiVector> xtest = MultiVectorFactory::Build(xstridedfullmap,1);
xtest->putScalar( (Scalar) 0.0);
RCP<Vector> xR = Teuchos::rcp(new Xpetra::EpetraVectorT<int,Node>(epv));
// calculate initial (absolute) residual
Teuchos::Array<Teuchos::ScalarTraits<Scalar>::magnitudeType> norms(1);
xR->norm2(norms);
*out << "Test: ||x_0|| = " << norms[0] << std::endl;
*out << "Test: Applying Simple Smoother" << std::endl;
*out << "Test: START DATA" << std::endl;
*out << "iterations\tVelocity_residual\tPressure_residual" << std::endl;
SimpleSm->Apply(*xtest,*xR);
xtest->norm2(norms);
*out << "Test: ||x_1|| = " << norms[0] << std::endl;
Teuchos::Array<Teuchos::ScalarTraits<Scalar>::magnitudeType> test = MueLu::Utilities<Scalar, LO, GO, Node>::ResidualNorm(*bOp, *xtest, *xR);
*out << "residual norm: " << test[0] << std::endl;
success = (test[0] < 1.0e-7);
if (!success)
*out << "no convergence" << std::endl;
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
return ( success ? EXIT_SUCCESS : EXIT_FAILURE );
#else
std::cout << "Epetra needs Serial node. Please recompile MueLu with the Serial node enabled." << std::endl;
return EXIT_SUCCESS;
#endif // #if defined(HAVE_MUELU_SERIAL) && defined(HAVE_MUELU_EPETRA)
}