//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( EpetraPointJacobiPreconditioner, tridiag_matrix )
{
typedef Epetra_RowMatrix MatrixType;
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 3 );
double diag_val = 2.0;
Teuchos::Array<double> values( 3, diag_val );
for ( int i = 1; i < global_num_rows-1; ++i )
{
global_columns[0] = i-1;
global_columns[1] = i;
global_columns[2] = i+1;
A->InsertGlobalValues( i, global_columns().size(),
&values[0], &global_columns[0] );
}
Teuchos::Array<int> single_col(1,0);
Teuchos::Array<double> diag_elem(1,diag_val);
A->InsertGlobalValues( 0, 1, diag_elem.getRawPtr(), single_col.getRawPtr() );
single_col[0] = global_num_rows-1;
A->InsertGlobalValues( global_num_rows-1, 1, diag_elem.getRawPtr(),
single_col.getRawPtr() );
A->FillComplete();
// Build the preconditioner.
Teuchos::RCP<MCLS::Preconditioner<MatrixType> > preconditioner =
Teuchos::rcp( new MCLS::EpetraPointJacobiPreconditioner() );
preconditioner->setOperator( A );
preconditioner->buildPreconditioner();
Teuchos::RCP<const MatrixType> M = preconditioner->getLeftPreconditioner();
// Check the preconditioner.
Teuchos::RCP<VectorType> X = MT::cloneVectorFromMatrixRows(*A);
MT::getLocalDiagCopy( *M, *X );
Teuchos::ArrayRCP<const double> X_view = VT::view( *X );
Teuchos::ArrayRCP<const double>::const_iterator view_iterator;
for ( view_iterator = X_view.begin();
view_iterator != X_view.end();
++view_iterator )
{
TEST_EQUALITY( *view_iterator, 1.0/diag_val );
}
}
Teuchos::RCP<Epetra_CrsMatrix> buildMatrix(int nx, Epetra_Comm & comm)
{
Epetra_Map map(nx*comm.NumProc(),0,comm);
Teuchos::RCP<Epetra_CrsMatrix> mat = Teuchos::rcp(new Epetra_CrsMatrix(Copy,map,3));
int offsets[3] = {-1, 0, 1 };
double values[3] = { -1, 2, -1};
int maxGid = map.MaxAllGID();
for(int lid=0;lid<nx;lid++) {
int gid = mat->GRID(lid);
int numEntries = 3, offset = 0;
int indices[3] = { gid+offsets[0],
gid+offsets[1],
gid+offsets[2] };
if(gid==0) { // left end point
numEntries = 2;
offset = 1;
} // right end point
else if(gid==maxGid)
numEntries = 2;
// insert rows
mat->InsertGlobalValues(gid,numEntries,values+offset,indices+offset);
}
mat->FillComplete();
return mat;
}
/// building the stencil
void PoissonSolver::StencilGeometry(Teuchos::RCP<Epetra_Vector> & RHS,
Teuchos::RCP<Epetra_CrsMatrix> & A)
{
const Epetra_BlockMap & MyMap = RHS->Map();
int NumMyElements = MyMap.NumMyElements();
int* MyGlobalElements = MyMap.MyGlobalElements();
double * rhsvalues = RHS->Values();
std::vector<double> Values(5);
std::vector<int> Indices(5);
const auto & inside = _bend.getInsideMask();
for (int lid = 0; lid < NumMyElements; ++ lid) {
size_t NumEntries = 0;
const size_t & gid = MyGlobalElements[lid];
cutoffStencil(Indices,
Values,
rhsvalues[lid],
NumEntries,
inside,
gid);
A->InsertGlobalValues(gid, NumEntries, &Values[0], &Indices[0]);
}
A->FillComplete();
A->OptimizeStorage();
}
Teuchos::RCP<Epetra_CrsMatrix>
buildH( const Teuchos::RCP<Epetra_CrsMatrix> &A)
{
Teuchos::RCP<Epetra_CrsMatrix> H = Teuchos::rcp(
new Epetra_CrsMatrix(Copy, A->RowMap(), A->GlobalMaxNumEntries() ) );
int N = A->NumGlobalRows();
std::vector<double> A_values( N );
std::vector<int> A_indices( N );
int A_size = 0;
double local_H;
bool found_diag = false;
for ( int i = 0; i < N; ++i )
{
A->ExtractGlobalRowCopy( i,
N,
A_size,
&A_values[0],
&A_indices[0] );
for ( int j = 0; j < A_size; ++j )
{
if ( i == A_indices[j] )
{
local_H = 1.0 - A_values[j];
H->InsertGlobalValues( i, 1, &local_H, &A_indices[j] );
found_diag = true;
}
else
{
local_H = -A_values[j];
H->InsertGlobalValues( i, 1, &local_H, &A_indices[j] );
}
}
if ( !found_diag )
{
local_H = 1.0;
H->InsertGlobalValues( i, 1, &local_H, &i );
}
}
H->FillComplete();
return H;
}
TEUCHOS_UNIT_TEST( OperatorTools, spectral_radius_test)
{
int problem_size = 100;
Epetra_SerialComm comm;
Epetra_Map map( problem_size, 0, comm );
// Build A.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, map, problem_size ) );
double lower_diag = -1.0;
double diag = 2.0;
double upper_diag = -1.0;
int global_row = 0;
int lower_row = 0;
int upper_row = 0;
for ( int i = 0; i < problem_size; ++i )
{
global_row = A->GRID(i);
lower_row = i-1;
upper_row = i+1;
if ( lower_row > -1 )
{
A->InsertGlobalValues( global_row, 1, &lower_diag, &lower_row );
}
A->InsertGlobalValues( global_row, 1, &diag, &global_row );
if ( upper_row < problem_size )
{
A->InsertGlobalValues( global_row, 1, &upper_diag, &upper_row );
}
}
A->FillComplete();
double spec_rad_A = HMCSA::OperatorTools::spectralRadius( A );
std::cout << spec_rad_A << std::endl;
}
Teuchos::RCP<Epetra_CrsMatrix> Epetra_Operator_to_Epetra_Matrix::constructInverseMatrix(const Epetra_Operator &op, const Epetra_Map &map)
{
int numEntriesPerRow = 0;
Teuchos::RCP<Epetra_FECrsMatrix> matrix = Teuchos::rcp(new Epetra_FECrsMatrix(::Copy, map, numEntriesPerRow));
int numRows = map.NumGlobalElements();
Epetra_Vector X(map);
Epetra_Vector Y(map);
double tol = 1e-15; // values below this will be considered 0
for (int rowIndex=0; rowIndex<numRows; rowIndex++)
{
int lid = map.LID(rowIndex);
if (lid != -1)
{
X[lid] = 1.0;
}
op.ApplyInverse(X, Y);
if (lid != -1)
{
X[lid] = 0.0;
}
std::vector<double> values;
std::vector<int> indices;
for (int i=0; i<map.NumMyElements(); i++)
{
if (abs(Y[i]) > tol)
{
values.push_back(Y[i]);
indices.push_back(map.GID(i));
}
}
matrix->InsertGlobalValues(rowIndex, values.size(), &values[0], &indices[0]);
}
matrix->GlobalAssemble();
return matrix;
}
// ****************************************************************************
void DG_Prob::DG_EI_Epetra_Outflow(const EDGE border,
Teuchos::RCP<Epetra_FECrsMatrix> A,
Teuchos::RCP<Epetra_FEVector> RHS)
{
MyElem e0;
e0=el[border.elemento[0]];
const int ntot = e0.show_ptr_stdel(sat)->nn_val() +
e0.show_ptr_stdel(pres)->nn_val();
double mx [ntot*ntot];
double B [ntot];
int indx[ntot];
const int ntot1 = DG_EI_Outflow(border,mx,B,indx);
if(ntot1 > ntot){
cout<<"DG_EI_Epetra_Outflow: ntot1 > ntot\n";
exit(1);
}
A->InsertGlobalValues(ntot1,&indx[0],&mx[0],Epetra_FECrsMatrix::ROW_MAJOR);
RHS->SumIntoGlobalValues(ntot1,&indx[0],&B[0]);
};
void
createEpetraProblem (const Teuchos::RCP<const Epetra_Comm>& epetraComm,
const std::string& filename,
Teuchos::RCP<Epetra_Map>& rowMap,
Teuchos::RCP<Epetra_CrsMatrix>& A,
Teuchos::RCP<Epetra_MultiVector>& B,
Teuchos::RCP<Epetra_MultiVector>& X,
int& numRHS)
{
using Teuchos::inOutArg;
using Teuchos::ptr;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::set_extra_data;
const int MyPID = epetraComm->MyPID();
int n_nonzeros, N_update;
int *bindx = NULL, *update = NULL, *col_inds = NULL;
double *val = NULL, *row_vals = NULL;
double *xguess = NULL, *b = NULL, *xexact = NULL;
//
// Set up the problem to be solved
//
int NumGlobalElements; // total # of rows in matrix
try {
// Read in matrix from HB file
Trilinos_Util_read_hb (const_cast<char *> (filename.c_str()), MyPID,
&NumGlobalElements, &n_nonzeros, &val, &bindx,
&xguess, &b, &xexact);
// Distribute data among processors
Trilinos_Util_distrib_msr_matrix (*epetraComm, &NumGlobalElements,
&n_nonzeros, &N_update, &update, &val,
&bindx, &xguess, &b, &xexact);
//
// Construct the matrix
//
int NumMyElements = N_update; // # local rows of matrix on processor
//
// Create an int array 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 processor.
//
std::vector<int> NumNz (NumMyElements);
for (int i = 0; i < NumMyElements; ++i) {
NumNz[i] = bindx[i+1] - bindx[i] + 1;
}
rowMap = rcp (new Epetra_Map (NumGlobalElements, NumMyElements,
update, 0, *epetraComm));
//set_extra_data (epetraComm, "Map::Comm", inOutArg (rowMap));
// Create an Epetra sparse matrix.
if (NumMyElements == 0) {
A = rcp (new Epetra_CrsMatrix (Copy, *rowMap, static_cast<int*>(NULL)));
} else {
A = rcp (new Epetra_CrsMatrix (Copy, *rowMap, &NumNz[0]));
}
//set_extra_data (rowMap, "Operator::Map", ptr (A));
// Add rows to the sparse matrix one at a time.
int NumEntries;
for (int i = 0; i < NumMyElements; ++i) {
row_vals = val + bindx[i];
col_inds = bindx + bindx[i];
NumEntries = bindx[i+1] - bindx[i];
int info = A->InsertGlobalValues (update[i], NumEntries, row_vals, col_inds);
TEUCHOS_TEST_FOR_EXCEPTION( info != 0, std::logic_error,
"Failed to insert global value into A." );
info = A->InsertGlobalValues (update[i], 1, val+i, update+i);
TEUCHOS_TEST_FOR_EXCEPTION( info != 0, std::logic_error,
"Failed to insert global value into A." );
}
// Finish initializing the sparse matrix.
int info = A->FillComplete();
TEUCHOS_TEST_FOR_EXCEPTION( info != 0, std::logic_error,
"FillComplete() failed on the sparse matrix A.");
info = A->OptimizeStorage();
TEUCHOS_TEST_FOR_EXCEPTION( info != 0, std::logic_error,
"OptimizeStorage() failed on the sparse matrix A.");
A->SetTracebackMode (1); // Shut down Epetra warning tracebacks
//
// Construct the right-hand side and solution multivectors.
//
if (false && b != NULL) {
B = rcp (new Epetra_MultiVector (::Copy, *rowMap, b, NumMyElements, 1));
numRHS = 1;
} else {
B = rcp (new Epetra_MultiVector (*rowMap, numRHS));
B->Random ();
}
X = rcp (new Epetra_MultiVector (*rowMap, numRHS));
X->PutScalar (0.0);
//set_extra_data (rowMap, "X::Map", Teuchos::ptr (X));
//set_extra_data (rowMap, "B::Map", Teuchos::ptr (B));
}
catch (std::exception& e) {
// Free up memory before rethrowing the exception.
if (update) free(update);
if (val) free(val);
if (bindx) free(bindx);
if (xexact) free(xexact);
if (xguess) free(xguess);
if (b) free(b);
throw e; // Rethrow the exception.
}
catch (...) { // Epetra sometimes throws an int "object."
// Free up memory before rethrowing the exception.
if (update) free(update);
if (val) free(val);
if (bindx) free(bindx);
if (xexact) free(xexact);
if (xguess) free(xguess);
if (b) free(b);
TEUCHOS_TEST_FOR_EXCEPTION(true, std::logic_error,
"Generating the Epetra problem to solve failed "
"with an unknown error.");
}
//if (false) {
//
// Create workspace
//
//Teuchos::set_default_workspace_store (rcp (new Teuchos::WorkspaceStoreInitializeable(static_cast<size_t>(2e+6))));
//}
//
// Free up memory.
//
if (update) free(update);
if (val) free(val);
if (bindx) free(bindx);
if (xexact) free(xexact);
if (xguess) free(xguess);
if (b) free(b);
}
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( UniformForwardSource, nh_set )
{
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
typedef MCLS::ForwardHistory<int> HistoryType;
typedef std::mt19937 rng_type;
typedef MCLS::ForwardDomain<VectorType,MatrixType,rng_type> DomainType;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear system.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 1 );
Teuchos::Array<double> values( 1 );
for ( int i = 1; i < global_num_rows; ++i )
{
global_columns[0] = i-1;
values[0] = -0.5;
A->InsertGlobalValues( i, global_columns().size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-1;
values[0] = -0.5;
A->InsertGlobalValues( global_num_rows-1, global_columns().size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> A_T = MT::copyTranspose(*A);
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *A );
Teuchos::RCP<VectorType> b = MT::cloneVectorFromMatrixRows( *A );
VT::putScalar( *b, -1.0 );
// Build the forward domain.
Teuchos::ParameterList plist;
plist.set<int>( "Overlap Size", 0 );
Teuchos::RCP<DomainType> domain = Teuchos::rcp( new DomainType( A_T, x, plist ) );
// History setup.
HistoryType::setByteSize();
// Create the forward source with a set number of histories.
int mult = 10;
double cutoff = 1.0e-8;
plist.set<double>("Sample Ratio",mult);
plist.set<double>("Weight Cutoff", cutoff);
MCLS::UniformForwardSource<DomainType>
source( b, domain, comm, comm->getSize(), comm->getRank(), plist );
TEST_ASSERT( source.empty() );
TEST_EQUALITY( source.numToTransport(), 0 );
TEST_EQUALITY( source.numToTransportInSet(), mult*global_num_rows );
TEST_EQUALITY( source.numRequested(), mult*global_num_rows );
TEST_EQUALITY( source.numLeft(), 0 );
TEST_EQUALITY( source.numEmitted(), 0 );
// Build the source.
source.buildSource();
TEST_ASSERT( !source.empty() );
TEST_EQUALITY( source.numToTransport(), mult*local_num_rows );
TEST_EQUALITY( source.numToTransportInSet(), mult*global_num_rows );
TEST_EQUALITY( source.numRequested(), mult*global_num_rows );
TEST_EQUALITY( source.numLeft(), mult*local_num_rows );
TEST_EQUALITY( source.numEmitted(), 0 );
// Sample the source.
Teuchos::RCP<MCLS::PRNG<rng_type> > rng = Teuchos::rcp(
new MCLS::PRNG<rng_type>(comm->getRank()) );
source.setRNG( rng );
for ( int i = 0; i < mult*local_num_rows; ++i )
{
TEST_ASSERT( !source.empty() );
TEST_EQUALITY( source.numLeft(), mult*local_num_rows-i );
TEST_EQUALITY( source.numEmitted(), i );
Teuchos::RCP<HistoryType> history = source.getHistory();
TEST_EQUALITY( history->weight(), 1.0 );
TEST_ASSERT( domain->isGlobalState( history->globalState() ) );
TEST_ASSERT( history->alive() );
TEST_ASSERT( VT::isGlobalRow( *x, history->globalState() ) );
}
TEST_ASSERT( source.empty() );
TEST_EQUALITY( source.numLeft(), 0 );
TEST_EQUALITY( source.numEmitted(), mult*local_num_rows );
}
int main(int argc, char** argv) {
#if defined(HAVE_MPI) && defined(HAVE_EPETRA)
int numProcs = 1;
int localProc = 0;
//first, set up our MPI environment...
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &localProc);
MPI_Comm_size(MPI_COMM_WORLD, &numProcs);
// This program can only run on 3 processors, to make things
// work out easy.
//
if (numProcs != 3) {
std::cout << "num-procs="<<numProcs<<". This program can only "
<< "run on 3 procs. Exiting."<<std::endl;
MPI_Finalize();
return(0);
}
//Consider the following mesh of 4 2-D quad elements:
//
// *-------*-------*
// 8| 7| 6|
// | E2 | E3 |
// *-------*-------*
// 3| 2| 5|
// | E0 | E1 |
// *-------*-------*
// 0 1 4
//
// Node-ids are to the lower-left of each node (*).
//
// Mimicing a finite-element application, we will say that
// each node has 1 scalar degree-of-freedom, and assemble
// a matrix which would have 9 global rows and columns.
//
// Each processor will have 3 rows. We'll set up a strange
// initial map, where nodes are distributed as follows:
//
// proc 0: nodes 0,3,8,
// proc 1: nodes 1,2,7
// proc 2: nodes 4,5,6.
//
// After we assemble our matrix, we'll create another matrix
// and populate it with graph edge weights such that the
// partitioner repartitions the problem so that nodes are
// laid out as follows:
//
// proc 0: nodes 0, 1, 4
// proc 1: nodes 3, 2, 5
// proc 2: nodes 8, 7, 6
//
int nodesPerElem = 4;
int global_n = 9;
//First, set up the initial map:
std::vector<int> mynodes(3);
if (localProc == 0) {
mynodes[0] = 0; mynodes[1] = 3; mynodes[2] = 8;
}
if (localProc == 1) {
mynodes[0] = 1; mynodes[1] = 2; mynodes[2] = 7;
}
if (localProc == 2) {
mynodes[0] = 4; mynodes[1] = 5; mynodes[2] = 6;
}
Epetra_MpiComm comm(MPI_COMM_WORLD);
Epetra_Map origmap(global_n, 3, &mynodes[0], 0, comm);
Teuchos::RCP<Epetra_FECrsMatrix> matrix =
Teuchos::rcp(new Epetra_FECrsMatrix(Copy, origmap, 0));
//We'll assemble elements E0 and E1 on proc 0,
// element E2 or proc 1,
// element E3 on proc 2.
std::vector<int> indices(nodesPerElem);
std::vector<double> coefs(nodesPerElem*nodesPerElem,2.0);
if (localProc == 0) {
//element E0:
indices[0] = 0; indices[1] = 1; indices[2] = 2; indices[3] = 3;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
//element E1:
indices[0] = 1; indices[1] = 4; indices[2] = 5; indices[3] = 2;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
else if (localProc == 1) {
//element E2:
indices[0] = 3; indices[1] = 2; indices[2] = 7; indices[3] = 8;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
else { //localProc==2
//element E3:
indices[0] = 2; indices[1] = 5; indices[2] = 6; indices[3] = 7;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
int err = matrix->GlobalAssemble();
if (err != 0) {
std::cout << "err="<<err<<" returned from matrix->GlobalAssemble()"
<< std::endl;
}
// std::cout << "matrix: " << std::endl;
// std::cout << *matrix << std::endl;
//We'll need a Teuchos::ParameterList object to pass to the
//Isorropia::Epetra::Partitioner class.
Teuchos::ParameterList paramlist;
#ifdef HAVE_ISORROPIA_ZOLTAN
// If Zoltan is available, we'll specify that the Zoltan package be
// used for the partitioning operation, by creating a parameter
// sublist named "Zoltan".
// In the sublist, we'll set parameters that we want sent to Zoltan.
paramlist.set("PARTITIONING METHOD", "GRAPH");
paramlist.set("PRINT ZOLTAN METRICS", "2");
Teuchos::ParameterList& sublist = paramlist.sublist("Zoltan");
sublist.set("GRAPH_PACKAGE", "PHG");
//sublist.set("DEBUG_LEVEL", "1"); // Zoltan will print out parameters
//sublist.set("DEBUG_LEVEL", "5"); // proc 0 will trace Zoltan calls
//sublist.set("DEBUG_MEMORY", "2"); // Zoltan will trace alloc & free
#else
// If Zoltan is not available, a simple linear partitioner will be
// used to partition such that the number of nonzeros is equal (or
// close to equal) on each processor. No parameter is necessary to
// specify this.
#endif
Teuchos::RCP<Isorropia::Epetra::CostDescriber> costs =
Teuchos::rcp(new Isorropia::Epetra::CostDescriber);
//Next create a matrix which is a copy of the matrix we just
//assembled, but we'll replace the values with graph edge weights.
Teuchos::RCP<Epetra_FECrsMatrix> ge_weights =
Teuchos::rcp(new Epetra_FECrsMatrix(*matrix));
Teuchos::RCP<Epetra_CrsMatrix> crs_ge_weights;
crs_ge_weights = ge_weights;
//Fill the matrix with a "default" weight of 1.0.
crs_ge_weights->PutScalar(1.0);
//Now we'll put a "large" weight on edges that connect nodes
//0 and 1, 1 and 4,
//3 and 2, 2 and 5,
//8 and 7, 7 and 6.
double weight = 500.0;
if (localProc == 0) {
//row 0, edge 1
indices[0] = 1;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(0, 1, &coefs[0], &indices[0]);
//row 3, edge 2
indices[0] = 2;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(3, 1, &coefs[0], &indices[0]);
//row 8, edge 7
indices[0] = 7;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(8, 1, &coefs[0], &indices[0]);
}
if (localProc == 1) {
//row 1, edges 0 and 4
indices[0] = 0; indices[1] = 4;
coefs[0] = weight; coefs[1] = weight;
crs_ge_weights->ReplaceGlobalValues(1, 2, &coefs[0], &indices[0]);
//row 2, edges 3 and 5
indices[0] = 3; indices[1] = 5;
coefs[0] = weight; coefs[1] = weight;
crs_ge_weights->ReplaceGlobalValues(2, 2, &coefs[0], &indices[0]);
//row 7, edges 6 and 8
indices[0] = 6; indices[1] = 8;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(7, 2, &coefs[0], &indices[0]);
}
if (localProc == 2) {
//row 4, edge 1
indices[0] = 1;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(4, 1, &coefs[0], &indices[0]);
//row 5, edge 2
indices[0] = 2;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(5, 1, &coefs[0], &indices[0]);
//row 6, edge 7
indices[0] = 7;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(6, 1, &coefs[0], &indices[0]);
}
// std::cout << "crs_ge_weights: " << std::endl
// << *crs_ge_weights << std::endl;
//Now give the graph edge weights to the CostDescriber:
costs->setGraphEdgeWeights(crs_ge_weights);
Teuchos::RCP<const Epetra_RowMatrix> rowmatrix;
rowmatrix = matrix;
//Now create the partitioner object using an Isorropia factory-like
//function...
Teuchos::RCP<Isorropia::Epetra::Partitioner> partitioner =
Teuchos::rcp(new Isorropia::Epetra::Partitioner(rowmatrix, costs, paramlist));
//Next create a Redistributor object and use it to create a
//repartitioned copy of the matrix
Isorropia::Epetra::Redistributor rd(partitioner);
Teuchos::RCP<Epetra_CrsMatrix> bal_matrix;
//Use a try-catch block because Isorropia will throw an exception
//if it encounters an error.
if (localProc == 0) {
std::cout << " calling Isorropia::Epetra::Redistributor::redistribute..."
<< std::endl;
}
try {
bal_matrix = rd.redistribute(*rowmatrix);
}
catch(std::exception& exc) {
std::cout << "linsys example: Isorropia::Epetra::Redistributor threw "
<< "exception '" << exc.what() << "' on proc "
<< localProc << std::endl;
MPI_Finalize();
return(-1);
}
// Results
double bal0, bal1, cutn0, cutn1, cutl0, cutl1, cutWgt0, cutWgt1;
int numCuts0, numCuts1;
#if 1
// Balance and cut quality before partitioning
double goalWeight = 1.0 / (double)numProcs;
ispatest::compute_graph_metrics(*rowmatrix, *costs, goalWeight,
bal0, numCuts0, cutWgt0, cutn0, cutl0);
// Balance and cut quality after partitioning
Teuchos::RCP<Epetra_CrsMatrix> new_weights = rd.redistribute(*crs_ge_weights);
Isorropia::Epetra::CostDescriber new_costs;
new_costs.setGraphEdgeWeights(new_weights);
ispatest::compute_graph_metrics(*bal_matrix, new_costs, goalWeight,
bal1, numCuts1, cutWgt1, cutn1, cutl1);
#else
std::vector<double> bal(2), cutwgt(2), cutn(2), cutl(2);
std::vector<int >ncuts(2);
Epetra_Import &importer = rd.get_importer();
costs->compareBeforeAndAfterGraph(*rowmatrix, *bal_matrix, importer,
bal, ncuts, cutwgt, cutn, cutl);
bal0 = bal[0]; cutn0 = cutn[0]; cutl0 = cutl[0]; cutWgt0 = cutwgt[0]; numCuts0 = ncuts[0];
bal1 = bal[1]; cutn1 = cutn[1]; cutl1 = cutl[1]; cutWgt1 = cutwgt[1]; numCuts1 = ncuts[1];
#endif
bal_matrix.release();
if (localProc == 0){
std::cout << "Before partitioning: Number of cuts " << numCuts0 << " Cut weight " << cutWgt0 << std::endl;
std::cout << " Balance " << bal0 << " cutN " << cutn0 << " cutL " << cutl0;
std::cout << std::endl;
std::cout << "After partitioning: Number of cuts " << numCuts1 << " Cut weight " << cutWgt1 << std::endl;
std::cout << " Balance " << bal1 << " cutN " << cutn1 << " cutL " << cutl1;
std::cout << std::endl;
}
MPI_Finalize();
#else
std::cout << "part_redist: must have both MPI and EPETRA. Make sure Trilinos "
<< "is configured with --enable-mpi and --enable-epetra." << std::endl;
#endif
return(0);
}
int main(int argc, char *argv[])
{
using Teuchos::rcp_implicit_cast;
int i, ierr, gerr;
gerr = 0;
#ifdef HAVE_MPI
// Initialize MPI and setup an Epetra communicator
MPI_Init(&argc,&argv);
Teuchos::RCP<Epetra_MpiComm> Comm = Teuchos::rcp( new Epetra_MpiComm(MPI_COMM_WORLD) );
#else
// If we aren't using MPI, then setup a serial communicator.
Teuchos::RCP<Epetra_SerialComm> Comm = Teuchos::rcp( new Epetra_SerialComm() );
#endif
// number of global elements
int dim = 100;
int blockSize = 3;
// PID info
int MyPID = Comm->MyPID();
bool verbose = 0;
if (argc>1) {
if (argv[1][0]=='-' && argv[1][1]=='v') {
verbose = true;
}
}
// Construct a Map that puts approximately the same number of
// equations on each processor.
Teuchos::RCP<Epetra_Map> Map = Teuchos::rcp( new Epetra_Map(dim, 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 std::vector NumNz that is used to build the Petra Matrix.
// NumNz[i] is the Number of OFF-DIAGONAL term for the ith global equation
// on this processor
std::vector<int> NumNz(NumMyElements);
// We are building a tridiagonal matrix where each row has (-1 2 -1)
// So we need 2 off-diagonal terms (except for the first and last equation)
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0 || MyGlobalElements[i] == dim-1) {
NumNz[i] = 2;
}
else {
NumNz[i] = 3;
}
}
// Create an Epetra_Matrix
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(Copy, *Map, &NumNz[0]) );
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
std::vector<double> Values(2);
Values[0] = -1.0; Values[1] = -1.0;
std::vector<int> Indices(2);
double two = 2.0;
int NumEntries;
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
NumEntries = 1;
}
else if (MyGlobalElements[i] == dim-1) {
Indices[0] = dim-2;
NumEntries = 1;
}
else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
NumEntries = 2;
}
ierr = A->InsertGlobalValues(MyGlobalElements[i],NumEntries,&Values[0],&Indices[0]);
assert(ierr==0);
// Put in the diagonal entry
ierr = A->InsertGlobalValues(MyGlobalElements[i],1,&two,&MyGlobalElements[i]);
assert(ierr==0);
}
// Finish building the epetra matrix A
ierr = A->FillComplete();
assert(ierr==0);
// Create an Belos::EpetraOp from this Epetra_CrsMatrix
Teuchos::RCP<Belos::EpetraOp> op = Teuchos::rcp(new Belos::EpetraOp(A));
// Issue several useful typedefs;
typedef Belos::MultiVec<double> EMV;
typedef Belos::Operator<double> EOP;
// Create an Epetra_MultiVector for an initial std::vector to start the solver.
// Note that this needs to have the same number of columns as the blocksize.
Teuchos::RCP<Belos::EpetraMultiVec> ivec = Teuchos::rcp( new Belos::EpetraMultiVec(*Map, blockSize) );
ivec->Random();
// Create an output manager to handle the I/O from the solver
Teuchos::RCP<Belos::OutputManager<double> > MyOM = Teuchos::rcp( new Belos::OutputManager<double>( MyPID ) );
if (verbose) {
MyOM->setVerbosity( Belos::Errors + Belos::Warnings );
}
#ifdef HAVE_EPETRA_THYRA
typedef Thyra::MultiVectorBase<double> TMVB;
typedef Thyra::LinearOpBase<double> TLOB;
// create thyra objects from the epetra objects
// first, a Thyra::VectorSpaceBase
Teuchos::RCP<const Thyra::VectorSpaceBase<double> > epetra_vs =
Thyra::create_VectorSpace(Map);
// then, a MultiVectorBase (from the Epetra_MultiVector)
Teuchos::RCP<Thyra::MultiVectorBase<double> > thyra_ivec =
Thyra::create_MultiVector(rcp_implicit_cast<Epetra_MultiVector>(ivec),epetra_vs);
// then, a LinearOpBase (from the Epetra_CrsMatrix)
Teuchos::RCP<Thyra::LinearOpBase<double> > thyra_op =
Teuchos::rcp( new Thyra::EpetraLinearOp(A) );
// test the Thyra adapter multivector
ierr = Belos::TestMultiVecTraits<double,TMVB>(MyOM,thyra_ivec);
gerr |= ierr;
switch (ierr) {
case Belos::Ok:
if ( verbose && MyPID==0 ) {
std::cout << "*** ThyraAdapter PASSED TestMultiVecTraits()" << std::endl;
}
break;
case Belos::Error:
if ( verbose && MyPID==0 ) {
std::cout << "*** ThyraAdapter FAILED TestMultiVecTraits() ***"
<< std::endl << std::endl;
}
break;
}
// test the Thyra adapter operator
ierr = Belos::TestOperatorTraits<double,TMVB,TLOB>(MyOM,thyra_ivec,thyra_op);
gerr |= ierr;
switch (ierr) {
case Belos::Ok:
if ( verbose && MyPID==0 ) {
std::cout << "*** ThyraAdapter PASSED TestOperatorTraits()" << std::endl;
}
break;
case Belos::Error:
if ( verbose && MyPID==0 ) {
std::cout << "*** ThyraAdapter FAILED TestOperatorTraits() ***"
<< std::endl << std::endl;
}
break;
}
#endif
#ifdef HAVE_MPI
MPI_Finalize();
#endif
if (gerr) {
if (verbose && MyPID==0)
std::cout << "End Result: TEST FAILED" << std::endl;
return -1;
}
//
// Default return value
//
if (verbose && MyPID==0)
std::cout << "End Result: TEST PASSED" << std::endl;
return 0;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
bool testFailed;
bool boolret;
int MyPID = Comm.MyPID();
bool verbose = true;
bool debug = false;
std::string which("SM");
Teuchos::CommandLineProcessor cmdp(false,true);
cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
cmdp.setOption("debug","nodebug",&debug,"Print debugging information.");
cmdp.setOption("sort",&which,"Targetted eigenvalues (SM,LM,SR,LR,SI,or LI).");
if (cmdp.parse(argc,argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return -1;
}
typedef double ScalarType;
typedef Teuchos::ScalarTraits<ScalarType> ScalarTypeTraits;
typedef ScalarTypeTraits::magnitudeType MagnitudeType;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Anasazi::MultiVecTraits<ScalarType,MV> MVTraits;
typedef Anasazi::OperatorTraits<ScalarType,MV,OP> OpTraits;
// Dimension of the matrix
int nx = 10; // Discretization points in any one direction.
int NumGlobalElements = nx*nx; // Size of matrix nx*nx
// Construct a Map that puts approximately the same number of
// equations on each processor.
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 term for the ith global equation
// on this processor
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
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(Copy, Map, &NumNz[0]) );
// Diffusion coefficient, can be set by user.
// When rho*h/2 <= 1, the discrete convection-diffusion operator has real eigenvalues.
// When rho*h/2 > 1, the operator has complex eigenvalues.
double rho = 2*(nx+1);
// Compute coefficients for discrete convection-diffution operator
const double one = 1.0;
std::vector<double> Values(4);
std::vector<int> Indices(4);
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, info;
for (int i=0; i<NumMyElements; i++)
{
if (MyGlobalElements[i]==0)
{
Indices[0] = 1;
Indices[1] = nx;
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx*(nx-1))
{
Indices[0] = nx*(nx-1)+1;
Indices[1] = nx*(nx-2);
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx-1)
{
Indices[0] = nx-2;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = 2*nx-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == NumGlobalElements-1)
{
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = nx*(nx-1)-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] < nx)
{
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if ((MyGlobalElements[i]+1)%nx == 0)
{
Indices[0] = MyGlobalElements[i]-nx;
Indices[1] = MyGlobalElements[i]+nx;
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
Indices[0] = MyGlobalElements[i]-1;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
// Put in the diagonal entry
info = A->InsertGlobalValues(MyGlobalElements[i], 1, &diag, &MyGlobalElements[i]);
assert( info==0 );
}
// Finish up
info = A->FillComplete();
assert( info==0 );
A->SetTracebackMode(1); // Shutdown Epetra Warning tracebacks
//************************************
// Start the block Davidson iteration
//***********************************
//
// Variables used for the Generalized Davidson Method
//
int nev = 4;
int blockSize = 1;
int maxDim = 50;
int restartDim = 10;
int maxRestarts = 500;
double tol = 1e-10;
// Set verbosity level
int verbosity = Anasazi::Errors + Anasazi::Warnings;
if (verbose) {
verbosity += Anasazi::FinalSummary + Anasazi::TimingDetails;
}
if (debug) {
verbosity += Anasazi::Debug;
}
//
// Create parameter list to pass into solver manager
//
Teuchos::ParameterList MyPL;
MyPL.set( "Verbosity", verbosity );
MyPL.set( "Which", which );
MyPL.set( "Block Size", blockSize );
MyPL.set( "Maximum Subspace Dimension", maxDim);
MyPL.set( "Restart Dimension", restartDim);
MyPL.set( "Maximum Restarts", maxRestarts );
MyPL.set( "Convergence Tolerance", tol );
MyPL.set( "Relative Convergence Tolerance", true );
MyPL.set( "Initial Guess", "User" );
// 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) );
ivec->Random();
// Create the eigenproblem.
Teuchos::RCP<Anasazi::BasicEigenproblem<double, MV, OP> > MyProblem = Teuchos::rcp(
new Anasazi::BasicEigenproblem<double,MV,OP>() );
MyProblem->setA(A);
MyProblem->setInitVec(ivec);
// Inform the eigenproblem that the operator A is non-Hermitian
MyProblem->setHermitian(false);
// Set the number of eigenvalues requested
MyProblem->setNEV( nev );
// Inform the eigenproblem that you are finishing passing it information
boolret = MyProblem->setProblem();
if (boolret != true) {
if (verbose && MyPID == 0) {
std::cout << "Anasazi::BasicEigenproblem::setProblem() returned with error." << std::endl;
}
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return -1;
}
// Initialize the Block Arnoldi solver
Anasazi::GeneralizedDavidsonSolMgr<double, MV, OP> MySolverMgr(MyProblem, MyPL);
// Solve the problem to the specified tolerances or length
Anasazi::ReturnType returnCode = MySolverMgr.solve();
testFailed = false;
if (returnCode != Anasazi::Converged && MyPID==0 && verbose) {
testFailed = true;
}
// Get the eigenvalues and eigenvectors from the eigenproblem
Anasazi::Eigensolution<ScalarType,MV> sol = MyProblem->getSolution();
std::vector<Anasazi::Value<ScalarType> > evals = sol.Evals;
Teuchos::RCP<MV> evecs = sol.Evecs;
std::vector<int> index = sol.index;
int numev = sol.numVecs;
// Output computed eigenvalues and their direct residuals
if (verbose && MyPID==0) {
int numritz = (int)evals.size();
std::cout.setf(std::ios_base::right, std::ios_base::adjustfield);
std::cout<<std::endl<< "Computed Ritz Values"<< std::endl;
std::cout<< std::setw(16) << "Real Part"
<< std::setw(16) << "Imag Part"
<< std::endl;
std::cout<<"-----------------------------------------------------------"<<std::endl;
for (int i=0; i<numritz; i++) {
std::cout<< std::setw(16) << evals[i].realpart
<< std::setw(16) << evals[i].imagpart
<< std::endl;
}
std::cout<<"-----------------------------------------------------------"<<std::endl;
}
if (numev > 0) {
// Compute residuals.
Teuchos::LAPACK<int,double> lapack;
std::vector<double> normA(numev);
// The problem is non-Hermitian.
int i=0;
std::vector<int> curind(1);
std::vector<double> resnorm(1), tempnrm(1);
Teuchos::RCP<MV> tempAevec;
Teuchos::RCP<const MV> evecr, eveci;
Epetra_MultiVector Aevec(Map,numev);
// Compute A*evecs
OpTraits::Apply( *A, *evecs, Aevec );
Teuchos::SerialDenseMatrix<int,double> Breal(1,1), Bimag(1,1);
while (i<numev) {
if (index[i]==0) {
// Get a view of the current eigenvector (evecr)
curind[0] = i;
evecr = MVTraits::CloneView( *evecs, curind );
// Get a copy of A*evecr
tempAevec = MVTraits::CloneCopy( Aevec, curind );
// Compute A*evecr - lambda*evecr
Breal(0,0) = evals[i].realpart;
MVTraits::MvTimesMatAddMv( -1.0, *evecr, Breal, 1.0, *tempAevec );
// Compute the norm of the residual and increment counter
MVTraits::MvNorm( *tempAevec, resnorm );
normA[i] = resnorm[0] / Teuchos::ScalarTraits<MagnitudeType>::magnitude( evals[i].realpart );
i++;
} else {
// Get a view of the real part of the eigenvector (evecr)
curind[0] = i;
evecr = MVTraits::CloneView( *evecs, curind );
// Get a copy of A*evecr
tempAevec = MVTraits::CloneCopy( Aevec, curind );
// Get a view of the imaginary part of the eigenvector (eveci)
curind[0] = i+1;
eveci = MVTraits::CloneView( *evecs, curind );
// Set the eigenvalue into Breal and Bimag
Breal(0,0) = evals[i].realpart;
Bimag(0,0) = evals[i].imagpart;
// Compute A*evecr - evecr*lambdar + eveci*lambdai
MVTraits::MvTimesMatAddMv( -1.0, *evecr, Breal, 1.0, *tempAevec );
MVTraits::MvTimesMatAddMv( 1.0, *eveci, Bimag, 1.0, *tempAevec );
MVTraits::MvNorm( *tempAevec, tempnrm );
// Get a copy of A*eveci
tempAevec = MVTraits::CloneCopy( Aevec, curind );
// Compute A*eveci - eveci*lambdar - evecr*lambdai
MVTraits::MvTimesMatAddMv( -1.0, *evecr, Bimag, 1.0, *tempAevec );
MVTraits::MvTimesMatAddMv( -1.0, *eveci, Breal, 1.0, *tempAevec );
MVTraits::MvNorm( *tempAevec, resnorm );
// Compute the norms and scale by magnitude of eigenvalue
normA[i] = lapack.LAPY2( tempnrm[0], resnorm[0] ) /
lapack.LAPY2( evals[i].realpart, evals[i].imagpart );
normA[i+1] = normA[i];
i=i+2;
}
}
// Output computed eigenvalues and their direct residuals
if (verbose && MyPID==0) {
std::cout.setf(std::ios_base::right, std::ios_base::adjustfield);
std::cout<<std::endl<< "Actual Residuals"<<std::endl;
std::cout<< std::setw(16) << "Real Part"
<< std::setw(16) << "Imag Part"
<< std::setw(20) << "Direct Residual"<< std::endl;
std::cout<<"-----------------------------------------------------------"<<std::endl;
for (int j=0; j<numev; j++) {
std::cout<< std::setw(16) << evals[j].realpart
<< std::setw(16) << evals[j].imagpart
<< std::setw(20) << normA[j] << std::endl;
if ( normA[j] > tol ) {
testFailed = true;
}
}
std::cout<<"-----------------------------------------------------------"<<std::endl;
}
}
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
if (testFailed) {
if (verbose && MyPID==0) {
std::cout << "End Result: TEST FAILED" << std::endl;
}
return -1;
}
//
// Default return value
//
if (verbose && MyPID==0) {
std::cout << "End Result: TEST PASSED" << std::endl;
}
return 0;
}
int extract_matrices
(
Epetra_CrsMatrix *A, // i/p: A matrix
shylu_symbolic *ssym, // symbolic structure
shylu_data *data, // numeric structure, TODO: Required ?
shylu_config *config, // i/p: library configuration
bool insertValues // true implies values will be inserted and fill
// complete will be called. false implies values
// will be replaced.
)
{
Teuchos::RCP<Epetra_CrsMatrix> D = ssym->D;
Teuchos::RCP<Epetra_CrsMatrix> C = ssym->C;
Teuchos::RCP<Epetra_CrsMatrix> R = ssym->R;
Teuchos::RCP<Epetra_CrsMatrix> G = ssym->G;
Teuchos::RCP<Epetra_CrsGraph> Sg = ssym->Sg;
int *DColElems = data->DColElems;
int *gvals = data->gvals;
double Sdiagfactor = config->Sdiagfactor;
int *LeftIndex = new int[data->lmax];
double *LeftValues = new double[data->lmax];
int *RightIndex = new int[data->rmax];
double *RightValues = new double[data->rmax];
int err;
int lcnt, rcnt ;
int gcid;
int gid;
int *Ai;
double *Ax;
int nrows = A->RowMap().NumMyElements();
int *rows = A->RowMap().MyGlobalElements();
for (int i = 0; i < nrows ; i++)
{
int NumEntries;
err = A->ExtractMyRowView(i, NumEntries, Ax, Ai);
lcnt = 0; rcnt = 0;
// Place the entry in the correct sub matrix, Works only for sym
gid = rows[i];
int lcid;
for (int j = 0 ; j < NumEntries ; j++)
{ // O(nnz) ! Careful what you do inside
// Row permutation does not matter here
gcid = A->GCID(Ai[j]);
assert(gcid != -1);
//Either in D or R
if ((gvals[gid] != 1 && gvals[gcid] == 1)
|| (gvals[gid] == 1 && A->LRID(gcid) != -1 && gvals[gcid] == 1))
{
assert(lcnt < data->lmax);
if (insertValues)
LeftIndex[lcnt] = gcid;
else
{
//local column id
lcid = (gvals[gid] == 1 ? D->LCID(gcid) : R->LCID(gcid));
assert(lcid != -1);
LeftIndex[lcnt] = lcid;
}
LeftValues[lcnt++] = Ax[j];
}
else
{
assert(rcnt < data->rmax);
if (insertValues)
RightIndex[rcnt] = gcid;
else
{
//local column id
lcid = (gvals[gid] == 1 ? C->LCID(gcid) : G->LCID(gcid));
assert(lcid != -1);
RightIndex[rcnt] = lcid;
}
RightValues[rcnt++] = Ax[j];
}
}
if (gvals[gid] == 1)
{ // D or C row
if (insertValues)
{
err = D->InsertGlobalValues(gid, lcnt, LeftValues, LeftIndex);
assert(err == 0);
err = C->InsertGlobalValues(gid, rcnt, RightValues, RightIndex);
assert(err == 0);
}
else
{
err = D->ReplaceMyValues(D->LRID(gid), lcnt, LeftValues,
LeftIndex);
assert(err == 0);
err = C->ReplaceMyValues(C->LRID(gid), rcnt, RightValues,
RightIndex);
assert(err == 0);
}
}
else
{ // R or S row
//assert(lcnt > 0); // TODO: Enable this once using narrow sep.
if (insertValues)
{
assert(rcnt > 0);
err = R->InsertGlobalValues(gid, lcnt, LeftValues, LeftIndex);
assert(err == 0);
err = G->InsertGlobalValues(gid, rcnt, RightValues, RightIndex);
assert(err == 0);
if (config->schurApproxMethod == 1)
{
err = Sg->InsertGlobalIndices(gid, rcnt, RightIndex);
assert(err == 0);
}
}
else
{
assert(rcnt > 0);
err = R->ReplaceMyValues(R->LRID(gid), lcnt, LeftValues,
LeftIndex);
assert(err == 0);
err = G->ReplaceMyValues(G->LRID(gid), rcnt, RightValues,
RightIndex);
assert(err == 0);
}
}
}
if (insertValues)
{
/* ------------- Create the maps for the DBBD form ------------------ */
Epetra_Map *DRowMap, *SRowMap, *DColMap;
Epetra_SerialComm LComm;
if (config->sep_type != 1)
{
DRowMap = new Epetra_Map(-1, data->Dnr, data->DRowElems, 0,
A->Comm());
SRowMap = new Epetra_Map(-1, data->Snr, data->SRowElems, 0,
A->Comm());
DColMap = new Epetra_Map(-1, data->Dnc, DColElems, 0,
A->Comm());
}
else
{
DRowMap = new Epetra_Map(-1, data->Dnr, data->DRowElems, 0, LComm);
SRowMap = new Epetra_Map(-1, data->Snr, data->SRowElems, 0, LComm);
DColMap = new Epetra_Map(-1, data->Dnc, DColElems, 0, LComm);
}
D->FillComplete();
//config->dm.print(5, "Done D fillcomplete");
G->FillComplete();
//config->dm.print(5, "Done G fillcomplete");
C->FillComplete(*SRowMap, *DRowMap); //TODO:Won't work if permutation is
// unsymmetric SRowMap
//config->dm.print(5, "Done C fillcomplete");
R->FillComplete(*DColMap, *SRowMap);
//config->dm.print(5, "Done R fillcomplete");
int Sdiag = (int) data->Snr * Sdiagfactor;
Sdiag = MIN(Sdiag, data->Snr-1);
Sdiag = MAX(Sdiag, 0);
// Add the diagonals to Sg
for (int i = 0; config->schurApproxMethod == 1 && i < nrows ; i++)
{
gid = rows[i];
if (gvals[gid] == 1) continue; // not a row in S
if (data->Snr == 0) assert(0 == 1);
rcnt = 0;
//TODO Will be trouble if SNumGlobalCols != Snc
//assert(SNumGlobalCols == Snc);
//for (int j = MAX(i-Sdiag,0) ; j<MIN(SNumGlobalCols, i+Sdiag); j++)
for (int j = MAX(i-Sdiag, 0) ; j < MIN(data->Snr, i+Sdiag); j++)
{
// find the adjacent columns from the row map of S
//assert (j >= 0 && j < Snr);
RightIndex[rcnt++] = data->SRowElems[j];
}
err = Sg->InsertGlobalIndices(gid, rcnt, RightIndex);
assert(err == 0);
// Always insert the diagonals, if it is added twice that is fine.
err = Sg->InsertGlobalIndices(gid, 1, &gid);
assert(err == 0);
}
if (config->schurApproxMethod == 1)
Sg->FillComplete();
delete DRowMap;
delete SRowMap;
delete DColMap;
}
#if 0
if (insertValues)
{
#ifdef TIMING_OUTPUT
Teuchos::Time ttime("transpose time");
ttime.start();
#endif
bool MakeDataContiguous = true;
ssym->transposer = Teuchos::RCP<EpetraExt::RowMatrix_Transpose>(new EpetraExt::RowMatrix_Transpose(MakeDataContiguous));
ssym->DT = Teuchos::rcp( dynamic_cast<Epetra_CrsMatrix *>(&(*ssym->transposer)(*D)), false);
#ifdef TIMING_OUTPUT
ttime.stop();
cout << "Transpose Time" << ttime.totalElapsedTime() << endl;
ttime.reset();
#endif
}
else
{
ssym->transposer->fwd();
//ssym->ReIdx_LP->fwd(); // TODO: Needed ?
}
#endif
// A is no longer needed
delete[] LeftIndex;
delete[] LeftValues;
delete[] RightIndex;
delete[] RightValues;
//cout << msg << "S rows=" << S.NumGlobalRows() << " S cols=" <<
//S.NumGlobalCols() << "#cols in column map="<<
//S.ColMap().NumMyElements() << endl;
//cout << msg << "C rows=" << Cptr->NumGlobalRows() << " C cols=" <<
//Cptr->NumGlobalCols() << "#cols in column map="<<
//Cptr->ColMap().NumMyElements() << endl;
//cout << msg << "D rows=" << D.NumGlobalRows() << " D cols=" <<
//D.NumGlobalCols() << "#cols in column map="<<
//D.ColMap().NumMyElements() << endl;
//cout << msg << "R rows=" << Rptr->NumGlobalRows() << " R cols=" <<
//Rptr->NumGlobalCols() << "#cols in column map="<<
//Rptr->ColMap().NumMyElements() << endl;
// ]
return 0;
}
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( DomainTransporter, Boundary )
{
typedef Epetra_Vector VectorType;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
typedef MCLS::AdjointHistory<int> HistoryType;
typedef std::mt19937 rng_type;
typedef MCLS::AdjointDomain<VectorType,MatrixType,rng_type> DomainType;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int comm_rank = comm->getRank();
// This test really needs a decomposed domain such that we can check
// hitting the local domain boundary.
if ( comm_size > 1 )
{
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear operator and solution vector. This operator will
// be assymetric so we quickly move the histories out of the domain
// before they hit the low weight cutoff.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 3 );
Teuchos::Array<double> values( 3 );
global_columns[0] = 0;
global_columns[1] = 1;
global_columns[2] = 2;
values[0] = 1.0;
values[1] = 0.49;
values[2] = 0.49;
A->InsertGlobalValues( 0, global_columns.size(),
&values[0], &global_columns[0] );
for ( int i = 1; i < global_num_rows-1; ++i )
{
global_columns[0] = i-1;
global_columns[1] = i;
global_columns[2] = i+1;
values[0] = 0.49;
values[1] = 1.0;
values[2] = 0.49;
A->InsertGlobalValues( i, global_columns.size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-3;
global_columns[1] = global_num_rows-2;
global_columns[2] = global_num_rows-1;
values[0] = 0.49;
values[1] = 0.49;
values[2] = 1.0;
A->InsertGlobalValues( global_num_rows-1, global_columns().size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> B = MT::copyTranspose(*A);
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *A );
// Build the adjoint domain.
Teuchos::ParameterList plist;
plist.set<int>( "Overlap Size", 2 );
Teuchos::RCP<DomainType> domain = Teuchos::rcp( new DomainType( B, x, plist ) );
Teuchos::RCP<MCLS::PRNG<rng_type> > rng = Teuchos::rcp(
new MCLS::PRNG<rng_type>( comm->getRank() ) );
domain->setRNG( rng );
// Build the domain transporter.
MCLS::DomainTransporter<DomainType> transporter( domain, plist );
domain->setCutoff( 1.0e-12 );
// Transport histories through the domain until they hit a boundary.
double weight = 3.0;
for ( int i = 0; i < global_num_rows-1; ++i )
{
if ( comm_rank == comm_size - 1 )
{
if ( i >= local_num_rows*comm_rank && i < local_num_rows*(comm_rank+1) )
{
HistoryType history( i, i, weight );
history.live();
transporter.transport( history );
TEST_ASSERT( history.event() ==
MCLS::Event::BOUNDARY || MCLS::Event::CUTOFF );
TEST_ASSERT( !history.alive() );
}
}
else
{
if ( i >= local_num_rows*comm_rank && i < 2+local_num_rows*(comm_rank+1) )
{
HistoryType history( i, i, weight );
history.live();
transporter.transport( history );
TEST_ASSERT( history.event() ==
MCLS::Event::BOUNDARY || MCLS::Event::CUTOFF );
TEST_ASSERT( !history.alive() );
}
}
}
}
}
int main(int argc, char *argv[]) {
int i, j, info;
const double one = 1.0;
const double zero = 0.0;
Teuchos::LAPACK<int,double> lapack;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
// Dimension of the matrix
int m = 500;
int n = 100;
// Construct a Map that puts approximately the same number of
// equations on each processor.
Epetra_Map RowMap(m, 0, Comm);
Epetra_Map ColMap(n, 0, Comm);
// Get update list and number of local equations from newly created Map.
int NumMyRowElements = RowMap.NumMyElements();
std::vector<int> MyGlobalRowElements(NumMyRowElements);
RowMap.MyGlobalElements(&MyGlobalRowElements[0]);
/* We are building an m by n matrix with entries
A(i,j) = k*(si)*(tj - 1) if i <= j
= k*(tj)*(si - 1) if i > j
where si = i/(m+1) and tj = j/(n+1) and k = 1/(n+1).
*/
// Create an Epetra_Matrix
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(Copy, RowMap, n) );
// Compute coefficients for discrete integral operator
std::vector<double> Values(n);
std::vector<int> Indices(n);
double inv_mp1 = one/(m+1);
double inv_np1 = one/(n+1);
for (i=0; i<n; i++) { Indices[i] = i; }
for (i=0; i<NumMyRowElements; i++) {
//
for (j=0; j<n; j++) {
//
if ( MyGlobalRowElements[i] <= j ) {
Values[j] = inv_np1 * ( (MyGlobalRowElements[i]+one)*inv_mp1 ) * ( (j+one)*inv_np1 - one ); // k*(si)*(tj-1)
}
else {
Values[j] = inv_np1 * ( (j+one)*inv_np1 ) * ( (MyGlobalRowElements[i]+one)*inv_mp1 - one ); // k*(tj)*(si-1)
}
}
info = A->InsertGlobalValues(MyGlobalRowElements[i], n, &Values[0], &Indices[0]);
assert( info==0 );
}
// Finish up
info = A->FillComplete(ColMap, RowMap);
assert( info==0 );
info = A->OptimizeStorage();
assert( info==0 );
A->SetTracebackMode(1); // Shutdown Epetra Warning tracebacks
//************************************
// Start the block Arnoldi iteration
//***********************************
//
// Variables used for the Block Arnoldi Method
//
int nev = 4;
int blockSize = 1;
int numBlocks = 10;
int maxRestarts = 20;
int verbosity = Anasazi::Errors + Anasazi::Warnings + Anasazi::FinalSummary;
double tol = lapack.LAMCH('E');
std::string which = "LM";
//
// Create parameter list to pass into solver
//
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 );
typedef Anasazi::MultiVec<double> MV;
typedef Anasazi::Operator<double> OP;
// Create an Anasazi::EpetraMultiVec for an initial vector to start the solver.
// Note: This needs to have the same number of columns as the blocksize.
Teuchos::RCP<Anasazi::EpetraMultiVec> ivec = Teuchos::rcp( new Anasazi::EpetraMultiVec(ColMap, blockSize) );
ivec->MvRandom();
// Call the constructor for the (A^T*A) operator
Teuchos::RCP<Anasazi::EpetraSymOp> Amat = Teuchos::rcp( new Anasazi::EpetraSymOp(A) );
Teuchos::RCP<Anasazi::BasicEigenproblem<double, MV, OP> > MyProblem =
Teuchos::rcp( new Anasazi::BasicEigenproblem<double, MV, OP>(Amat, ivec) );
// Inform the eigenproblem that the matrix A is symmetric
MyProblem->setHermitian(true);
// Set the number of eigenvalues requested and the blocksize the solver should use
MyProblem->setNEV( nev );
// Inform the eigenproblem that you are finished passing it information
bool boolret = MyProblem->setProblem();
if (boolret != true) {
if (MyPID == 0) {
cout << "Anasazi::BasicEigenproblem::setProblem() returned with error." << endl;
}
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return -1;
}
// Initialize the Block Arnoldi solver
Anasazi::BlockKrylovSchurSolMgr<double, MV, OP> MySolverMgr(MyProblem, MyPL);
// Solve the problem to the specified tolerances or length
Anasazi::ReturnType returnCode = MySolverMgr.solve();
if (returnCode != Anasazi::Converged && MyPID==0) {
cout << "Anasazi::EigensolverMgr::solve() returned unconverged." << endl;
}
// Get the eigenvalues and eigenvectors from the eigenproblem
Anasazi::Eigensolution<double,MV> sol = MyProblem->getSolution();
std::vector<Anasazi::Value<double> > evals = sol.Evals;
int numev = sol.numVecs;
if (numev > 0) {
// Compute singular values/vectors and direct residuals.
//
// Compute singular values which are the square root of the eigenvalues
if (MyPID==0) {
cout<<"------------------------------------------------------"<<endl;
cout<<"Computed Singular Values: "<<endl;
cout<<"------------------------------------------------------"<<endl;
}
for (i=0; i<numev; i++) { evals[i].realpart = Teuchos::ScalarTraits<double>::squareroot( evals[i].realpart ); }
//
// Compute left singular vectors : u = Av/sigma
//
std::vector<double> tempnrm(numev), directnrm(numev);
std::vector<int> index(numev);
for (i=0; i<numev; i++) { index[i] = i; }
Anasazi::EpetraMultiVec Av(RowMap,numev), u(RowMap,numev);
Anasazi::EpetraMultiVec* evecs = dynamic_cast<Anasazi::EpetraMultiVec* >(sol.Evecs->CloneViewNonConst( index ));
Teuchos::SerialDenseMatrix<int,double> S(numev,numev);
A->Apply( *evecs, Av );
Av.MvNorm( tempnrm );
for (i=0; i<numev; i++) { S(i,i) = one/tempnrm[i]; };
u.MvTimesMatAddMv( one, Av, S, zero );
//
// Compute direct residuals : || Av - sigma*u ||
//
for (i=0; i<numev; i++) { S(i,i) = evals[i].realpart; }
Av.MvTimesMatAddMv( -one, u, S, one );
Av.MvNorm( directnrm );
if (MyPID==0) {
cout.setf(std::ios_base::right, std::ios_base::adjustfield);
cout<<std::setw(16)<<"Singular Value"
<<std::setw(20)<<"Direct Residual"
<<endl;
cout<<"------------------------------------------------------"<<endl;
for (i=0; i<numev; i++) {
cout<<std::setw(16)<<evals[i].realpart
<<std::setw(20)<<directnrm[i]
<<endl;
}
cout<<"------------------------------------------------------"<<endl;
}
}
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0;
}
/* Apply an identity matrix to the Schur complement operator. Drop the entries
entries using a relative threshold. Assemble the result in a Crs Matrix
which will be our approximate Schur complement.
*/
Teuchos::RCP<Epetra_CrsMatrix> computeApproxSchur(shylu_config *config,
shylu_symbolic *sym,
Epetra_CrsMatrix *G, Epetra_CrsMatrix *R,
Epetra_LinearProblem *LP, Amesos_BaseSolver *solver,
Ifpack_Preconditioner *ifSolver, Epetra_CrsMatrix *C,
Epetra_Map *localDRowMap)
{
double relative_thres = config->relative_threshold;
int nvectors = 16;
ShyLU_Probing_Operator probeop(config, sym, G, R, LP, solver, ifSolver, C,
localDRowMap, nvectors);
// Get row map
Epetra_Map rMap = G->RowMap();
int *rows = rMap.MyGlobalElements();
int totalElems = rMap.NumGlobalElements();
int localElems = rMap.NumMyElements();
//cout << " totalElems in Schur Complement" << totalElems << endl;
//cout << myPID << " localElems" << localElems << endl;
// **************** Two collectives here *********************
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
ftime.start();
#endif
int prefixSum;
G->Comm().ScanSum(&localElems, &prefixSum, 1);
//cout << " prefixSum" << prefixSum << endl;
// Start the index in prefixSum-localElems
int *mySGID = new int[totalElems]; // vector of size Schur complement !
int *allSGID = new int[totalElems]; // vector of size Schur complement !
int i, j;
for (i = 0, j = 0; i < totalElems ; i++)
{
if (i >= prefixSum - localElems && i < prefixSum)
{
mySGID[i] = rows[j];
j++;
}
else
{
mySGID[i] = 0;
}
allSGID[i] = 0;
}
C->Comm().SumAll(mySGID, allSGID, totalElems);
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time to Compute RowIDS" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
// Now everyone knows the GIDs in the Schur complement
//cout << rMap << endl;
j = 0;
Teuchos::RCP<Epetra_CrsMatrix> Sbar = Teuchos::rcp(new Epetra_CrsMatrix(
Copy, rMap, localElems));
int nentries;
double *values = new double[localElems]; // Need to adjust this for more
int *indices = new int[localElems]; // than one vector
double *vecvalues;
int dropped = 0;
double *maxvalue = new double[nvectors];
#ifdef TIMING_OUTPUT
ftime.start();
#endif
#ifdef TIMING_OUTPUT
Teuchos::Time app_time("Apply time");
#endif
int findex = totalElems / nvectors ;
for (i = 0 ; i < findex*nvectors ; i+=nvectors)
{
Epetra_MultiVector probevec(rMap, nvectors);
Epetra_MultiVector Scol(rMap, nvectors);
probevec.PutScalar(0.0);
int cindex;
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
if (cindex >= prefixSum - localElems && cindex < prefixSum)
{
probevec.ReplaceGlobalValue(allSGID[cindex], k, 1.0);
}
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
Scol.MaxValue(maxvalue);
for (int k = 0; k < nvectors; k++) //TODO:Need to switch these loops
{
cindex = k+i;
vecvalues = Scol[k];
//cout << "MAX" << maxvalue << endl;
for (j = 0 ; j < localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (allSGID[cindex] == rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[cindex];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else if (abs(vecvalues[j]/maxvalue[k]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[cindex];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else
{
if (vecvalues[j] != 0.0) dropped++;
}
}
}
}
probeop.ResetTempVectors(1);
for ( ; i < totalElems ; i++)
{
Epetra_MultiVector probevec(rMap, 1); // TODO: Try doing more than one
Epetra_MultiVector Scol(rMap, 1); // vector at a time
probevec.PutScalar(0.0);
if (i >= prefixSum - localElems && i < prefixSum)
{
probevec.ReplaceGlobalValue(allSGID[i], 0, 1.0);
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
vecvalues = Scol[0];
Scol.MaxValue(maxvalue);
//cout << "MAX" << maxvalue << endl;
for (j = 0 ; j < localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (allSGID[i] == rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[i];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else if (abs(vecvalues[j]/maxvalue[0]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = allSGID[i];
nentries++;
Sbar->InsertGlobalValues(rows[j], nentries, values, indices);
}
else
{
if (vecvalues[j] != 0.0) dropped++;
}
}
}
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time in finding and dropping entries" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
#ifdef TIMING_OUTPUT
cout << "Time in Apply of probing" << app_time.totalElapsedTime() << endl;
#endif
Sbar->FillComplete();
cout << "#dropped entries" << dropped << endl;
delete[] allSGID;
delete[] mySGID;
delete[] values;
delete[] indices;
delete[] maxvalue;
return Sbar;
}
//---------------------------------------------------------------------------//
// Test templates
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( SolverFactory, mcsa_two_by_two )
{
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
int comm_size = comm->getSize();
int comm_rank = comm->getRank();
// This is a 4 processor test.
if ( comm_size == 4 )
{
// Build the set-constant communicator.
Teuchos::Array<int> ranks(2);
if ( comm_rank < 2 )
{
ranks[0] = 0;
ranks[1] = 1;
}
else
{
ranks[0] = 2;
ranks[1] = 3;
}
Teuchos::RCP<const Teuchos::Comm<int> > comm_set =
comm->createSubcommunicator( ranks() );
int set_size = comm_set->getSize();
// Declare the linear problem in the global scope.
Teuchos::RCP<MCLS::LinearProblem<VectorType,MatrixType> > linear_problem;
// Build the linear system on set 0.
if ( comm_rank < 2 )
{
int local_num_rows = 10;
int global_num_rows = local_num_rows*set_size;
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm_set );
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear system. This operator is symmetric with a spectral
// radius less than 1.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 3 );
Teuchos::Array<double> values( 3 );
global_columns[0] = 0;
global_columns[1] = 1;
global_columns[2] = 2;
values[0] = 1.0;
values[1] = 0.05;
values[2] = 0.05;
A->InsertGlobalValues( 0, global_columns.size(),
&values[0], &global_columns[0] );
for ( int i = 1; i < global_num_rows-1; ++i )
{
global_columns[0] = i-1;
global_columns[1] = i;
global_columns[2] = i+1;
values[0] = 0.05;
values[1] = 1.0;
values[2] = 0.05;
A->InsertGlobalValues( i, global_columns.size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-3;
global_columns[1] = global_num_rows-2;
global_columns[2] = global_num_rows-1;
values[0] = 0.05;
values[1] = 0.05;
values[2] = 1.0;
A->InsertGlobalValues( global_num_rows-1, global_columns.size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> B = A;
// Build the LHS. Put a large positive number here to be sure we are
// clear the vector before solving.
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *x, 0.0 );
// Build the RHS with negative numbers. this gives us a negative
// solution.
Teuchos::RCP<VectorType> b = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *b, -1.0 );
// Create the linear problem.
linear_problem = Teuchos::rcp(
new MCLS::LinearProblem<VectorType,MatrixType>(B, x, b) );
}
comm->barrier();
// Solver parameters.
Teuchos::RCP<Teuchos::ParameterList> plist =
Teuchos::rcp( new Teuchos::ParameterList() );
double cutoff = 1.0e-4;
plist->set<std::string>("MC Type", "Adjoint");
plist->set<double>("Convergence Tolerance", 1.0e-8);
plist->set<int>("Maximum Iterations", 10);
plist->set<double>("Weight Cutoff", cutoff);
plist->set<int>("Iteration Print Frequency", 1);
plist->set<int>("MC Check Frequency", 50);
plist->set<bool>("Reproducible MC Mode",true);
plist->set<int>("Overlap Size", 2);
plist->set<int>("Number of Sets", 2);
plist->set<double>("Sample Ratio", 10.0);
plist->set<std::string>("Transport Type", "Global" );
// Create the solver.
MCLS::SolverFactory<VectorType,MatrixType> factory;
Teuchos::RCP<MCLS::SolverManager<VectorType,MatrixType> > solver_manager =
factory.create( "MCSA", comm, plist );
solver_manager->setProblem( linear_problem );
// Solve the problem.
bool converged_status = solver_manager->solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager->getConvergedStatus() );
TEST_ASSERT( solver_manager->getNumIters() < 10 );
if ( comm_rank < 2 )
{
TEST_ASSERT( solver_manager->achievedTol() > 0.0 );
}
else
{
TEST_ASSERT( solver_manager->achievedTol() == 0.0 );
}
if ( comm_rank < 2 )
{
// Check that we got a negative solution.
Teuchos::ArrayRCP<const double> x_view =
VT::view( *linear_problem->getLHS() );
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it < Teuchos::ScalarTraits<double>::zero() );
}
}
comm->barrier();
// Now solve the problem with a positive source.
if ( comm_rank < 2 )
{
Teuchos::RCP<VectorType> b =
MT::cloneVectorFromMatrixRows( *linear_problem->getOperator() );
VT::putScalar( *b, 2.0 );
linear_problem->setRHS( b );
VT::putScalar( *linear_problem->getLHS(), 0.0 );
}
comm->barrier();
converged_status = solver_manager->solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager->getConvergedStatus() );
TEST_ASSERT( solver_manager->getNumIters() < 10 );
if ( comm_rank < 2 )
{
TEST_ASSERT( solver_manager->achievedTol() > 0.0 );
}
else
{
TEST_ASSERT( solver_manager->achievedTol() == 0.0 );
}
if ( comm_rank < 2 )
{
Teuchos::ArrayRCP<const double> x_view =
VT::view( *linear_problem->getLHS() );
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it > Teuchos::ScalarTraits<double>::zero() );
}
}
comm->barrier();
// Reset the domain and solve again with a positive source.
if ( comm_rank < 2 )
{
VT::putScalar( *linear_problem->getLHS(), 0.0 );
}
comm->barrier();
solver_manager->setProblem( linear_problem );
converged_status = solver_manager->solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager->getConvergedStatus() );
TEST_ASSERT( solver_manager->getNumIters() < 10 );
if ( comm_rank < 2 )
{
TEST_ASSERT( solver_manager->achievedTol() > 0.0 );
}
else
{
TEST_ASSERT( solver_manager->achievedTol() == 0.0 );
}
if ( comm_rank < 2 )
{
Teuchos::ArrayRCP<const double> x_view =
VT::view( *linear_problem->getLHS() );
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it > Teuchos::ScalarTraits<double>::zero() );
}
}
comm->barrier();
// Reset both and solve with a negative source.
if ( comm_rank < 2 )
{
Teuchos::RCP<VectorType> b =
MT::cloneVectorFromMatrixRows( *linear_problem->getOperator() );
VT::putScalar( *b, -2.0 );
linear_problem->setRHS( b );
VT::putScalar( *linear_problem->getLHS(), 0.0 );
}
comm->barrier();
converged_status = solver_manager->solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager->getConvergedStatus() );
TEST_ASSERT( solver_manager->getNumIters() < 10 );
if ( comm_rank < 2 )
{
TEST_ASSERT( solver_manager->achievedTol() > 0.0 );
}
else
{
TEST_ASSERT( solver_manager->achievedTol() == 0.0 );
}
if ( comm_rank < 2 )
{
Teuchos::ArrayRCP<const double> x_view =
VT::view( *linear_problem->getLHS() );
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it < Teuchos::ScalarTraits<double>::zero() );
}
}
comm->barrier();
}
}
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( SourceTransporter, transport )
{
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
typedef MCLS::AdjointHistory<int> HistoryType;
typedef std::mt19937 rng_type;
typedef MCLS::AdjointDomain<VectorType,MatrixType,rng_type> DomainType;
typedef MCLS::UniformAdjointSource<DomainType> SourceType;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear system. This operator is symmetric with a spectral
// radius less than 1.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 3 );
Teuchos::Array<double> values( 3 );
global_columns[0] = 0;
global_columns[1] = 1;
global_columns[2] = 2;
values[0] = 1.0/comm_size;
values[1] = 0.12/comm_size;
values[2] = 0.0/comm_size;
A->InsertGlobalValues( 0, global_columns.size(),
&values[0], &global_columns[0] );
for ( int i = 1; i < global_num_rows-1; ++i )
{
global_columns[0] = i-1;
global_columns[1] = i;
global_columns[2] = i+1;
values[0] = 0.12/comm_size;
values[1] = 1.0/comm_size;
values[2] = 0.12/comm_size;
A->InsertGlobalValues( i, global_columns.size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-3;
global_columns[1] = global_num_rows-2;
global_columns[2] = global_num_rows-1;
values[0] = 0.0/comm_size;
values[1] = 0.12/comm_size;
values[2] = 1.0/comm_size;
A->InsertGlobalValues( global_num_rows-1, global_columns.size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> B = MT::copyTranspose(*A);
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *x, 0.0 );
Teuchos::RCP<VectorType> b = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *b, -1.0 );
// Build the adjoint domain.
Teuchos::ParameterList plist;
plist.set<int>( "Overlap Size", 2 );
Teuchos::RCP<DomainType> domain = Teuchos::rcp( new DomainType( B, x, plist ) );
Teuchos::RCP<MCLS::PRNG<rng_type> > rng = Teuchos::rcp(
new MCLS::PRNG<rng_type>( comm->getRank() ) );
domain->setRNG( rng );
// History setup.
HistoryType::setByteSize();
// Create the adjoint source with a set number of histories.
int mult = 100;
double cutoff = 1.0e-6;
plist.set<double>("Sample Ratio", mult);
plist.set<double>("Weight Cutoff", cutoff);
Teuchos::RCP<SourceType> source = Teuchos::rcp(
new SourceType( b, domain, comm,
comm->getSize(), comm->getRank(), plist ) );
source->setRNG( rng );
source->buildSource();
// Create the source transporter.
plist.set<int>("MC Check Frequency", 10);
double relative_cutoff = source->sourceWeight()*cutoff;
MCLS::SourceTransporter<SourceType> source_transporter(
comm, domain, plist );
source_transporter.assignSource( source, relative_cutoff );
// Do transport.
source_transporter.transport();
domain->domainTally()->combineSetTallies( comm );
// Check that we got a negative solution.
Teuchos::ArrayRCP<const double> x_view =
VT::view( *x );
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it < 0.0 );
}
}
int main(int argc, char *argv[]) {
using std::cout;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
bool boolret;
int MyPID = Comm.MyPID();
bool verbose = true;
bool debug = false;
bool dynXtraNev = false;
std::string which("SM");
int nx = 10; // Discretization points in any one direction.
int nev = 10;
int blockSize = 3;
int numBlocks = 10;
double tol = 1e-5;
Teuchos::CommandLineProcessor cmdp(false,true);
cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
cmdp.setOption("debug","nodebug",&debug,"Print debugging information.");
cmdp.setOption("sort",&which,"Targetted eigenvalues (SM,LM,SR,LR,SI,or LI).");
cmdp.setOption("nx",&nx,"Number of discretization points in each direction; n=nx*nx.");
cmdp.setOption("nev",&nev,"Number of eigenvalues to compute.");
cmdp.setOption("blocksize",&blockSize,"Block Size.");
cmdp.setOption("tol",&tol,"Solver Tolerance.");
cmdp.setOption("numblocks",&numBlocks,"Number of blocks for the Krylov-Schur form.");
cmdp.setOption("dynrestart","nodynrestart",&dynXtraNev,"Use dynamic restart boundary to accelerate convergence.");
if (cmdp.parse(argc,argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return -1;
}
typedef double ScalarType;
typedef Teuchos::ScalarTraits<ScalarType> SCT;
typedef SCT::magnitudeType MagnitudeType;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Anasazi::MultiVecTraits<ScalarType,MV> MVT;
typedef Anasazi::OperatorTraits<ScalarType,MV,OP> OPT;
// Dimension of the matrix
int NumGlobalElements = nx*nx; // Size of matrix nx*nx
// Construct a Map that puts approximately the same number of
// equations on each processor.
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 term for the ith global equation
// on this processor
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
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(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 h = one /(nx+1);
double h2 = h*h;
Values[0] = -one/h2; Values[1] = -one/h2; Values[2] = -one/h2; Values[3]= -one/h2;
double diag = 4.0 / h2;
int NumEntries, info;
for (int i=0; i<NumMyElements; i++)
{
if (MyGlobalElements[i]==0)
{
Indices[0] = 1;
Indices[1] = nx;
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx*(nx-1))
{
Indices[0] = nx*(nx-1)+1;
Indices[1] = nx*(nx-2);
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx-1)
{
Indices[0] = nx-2;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = 2*nx-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == NumGlobalElements-1)
{
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = nx*(nx-1)-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] < nx)
{
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if ((MyGlobalElements[i]+1)%nx == 0)
{
Indices[0] = MyGlobalElements[i]-nx;
Indices[1] = MyGlobalElements[i]+nx;
NumEntries = 2;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
Indices[0] = MyGlobalElements[i]-1;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( 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;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
// Put in the diagonal entry
info = A->InsertGlobalValues(MyGlobalElements[i], 1, &diag, &MyGlobalElements[i]);
assert( info==0 );
}
// Finish up
info = A->FillComplete();
assert( info==0 );
A->SetTracebackMode(1); // Shutdown Epetra Warning tracebacks
//************************************
// Start the block Arnoldi iteration
//***********************************
//
// Variables used for the Block Krylov Schur Method
//
int maxRestarts = 500;
// Create a sort manager to pass into the block Krylov-Schur solver manager
// --> Make sure the reference-counted pointer is of type Anasazi::SortManager<>
// --> The block Krylov-Schur solver manager uses Anasazi::BasicSort<> by default,
// so you can also pass in the parameter "Which", instead of a sort manager.
Teuchos::RCP<Anasazi::SortManager<MagnitudeType> > MySort =
Teuchos::rcp( new Anasazi::BasicSort<MagnitudeType>( which ) );
// Set verbosity level
int verbosity = Anasazi::Errors + Anasazi::Warnings;
if (verbose) {
verbosity += Anasazi::FinalSummary + Anasazi::TimingDetails;
}
if (debug) {
verbosity += Anasazi::Debug;
}
//
// Create parameter list to pass into solver manager
//
Teuchos::ParameterList MyPL;
MyPL.set( "Verbosity", verbosity );
MyPL.set( "Sort Manager", MySort );
MyPL.set( "Block Size", blockSize );
MyPL.set( "Num Blocks", numBlocks );
MyPL.set( "Maximum Restarts", maxRestarts );
MyPL.set( "Convergence Tolerance", tol );
MyPL.set("Dynamic Extra NEV",dynXtraNev);
// 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) );
ivec->Random();
// Create the eigenproblem.
Teuchos::RCP<Anasazi::BasicEigenproblem<double, MV, OP> > MyProblem =
Teuchos::rcp( new Anasazi::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 );
// Inform the eigenproblem that you are finishing passing it information
boolret = MyProblem->setProblem();
if (boolret != true) {
if (verbose && MyPID == 0) {
std::cout << "Anasazi::BasicEigenproblem::setProblem() returned with error." << std::endl;
}
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return -1;
}
// Initialize the Block Arnoldi solver
Anasazi::BlockKrylovSchurSolMgr<double, MV, OP> MySolverMgr(MyProblem, MyPL);
// Solve the problem to the specified tolerances or length
bool testFailed = false;
Anasazi::ReturnType returnCode = MySolverMgr.solve();
if (returnCode != Anasazi::Converged) {
testFailed = true;
if (MyPID==0 && verbose)
std::cout << "Anasazi::EigensolverMgr::solve() returned unconverged." << std::endl;
}
// Get the Ritz values from the eigensolver
std::vector<Anasazi::Value<double> > ritzValues = MySolverMgr.getRitzValues();
// Output computed eigenvalues and their direct residuals
if (verbose && MyPID==0) {
int numritz = (int)ritzValues.size();
std::cout.setf(std::ios_base::right, std::ios_base::adjustfield);
std::cout<<std::endl<< "Computed Ritz Values"<< std::endl;
std::cout<< std::setw(16) << "Real Part"
<< std::endl;
std::cout<<"-----------------------------------------------------------"<<std::endl;
for (int i=0; i<numritz; i++) {
std::cout<< std::setw(16) << ritzValues[i].realpart
<< std::endl;
}
std::cout<<"-----------------------------------------------------------"<<std::endl;
}
// Get the eigenvalues and eigenvectors from the eigenproblem
Anasazi::Eigensolution<ScalarType,MV> sol = MyProblem->getSolution();
std::vector<Anasazi::Value<ScalarType> > evals = sol.Evals;
Teuchos::RCP<MV> evecs = sol.Evecs;
std::vector<int> index = sol.index;
int numev = sol.numVecs;
if (numev > 0) {
// Compute residuals.
std::vector<double> normA(numev);
// Get storage
Epetra_MultiVector Aevecs(Map,numev);
Teuchos::SerialDenseMatrix<int,double> B(numev,numev);
B.putScalar(0.0);
for (int i=0; i<numev; i++) {B(i,i) = evals[i].realpart;}
// Compute A*evecs
OPT::Apply( *A, *evecs, Aevecs );
// Compute A*evecs - lambda*evecs and its norm
MVT::MvTimesMatAddMv( -1.0, *evecs, B, 1.0, Aevecs );
MVT::MvNorm( Aevecs, normA );
// Scale the norms by the eigenvalue
for (int i=0; i<numev; i++) {
normA[i] /= Teuchos::ScalarTraits<double>::magnitude( evals[i].realpart );
if ( normA[i] > tol ) {
testFailed = true;
}
}
// Output computed eigenvalues and their direct residuals
if (verbose && MyPID==0) {
std::cout.setf(std::ios_base::right, std::ios_base::adjustfield);
std::cout<<std::endl<< "Actual Residuals"<<std::endl;
std::cout<< std::setw(16) << "Real Part"
<< std::setw(20) << "Direct Residual"<< std::endl;
std::cout<<"-----------------------------------------------------------"<<std::endl;
for (int i=0; i<numev; i++) {
std::cout<< std::setw(16) << evals[i].realpart
<< std::setw(20) << normA[i] << std::endl;
}
std::cout<<"-----------------------------------------------------------"<<std::endl;
}
}
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
if (testFailed) {
if (verbose && MyPID==0) {
std::cout << "End Result: TEST FAILED" << std::endl;
}
return -1;
}
//
// Default return value
//
if (verbose && MyPID==0) {
std::cout << "End Result: TEST PASSED" << std::endl;
}
return 0;
}
TEUCHOS_UNIT_TEST( MCSA, MCSA_test)
{
int problem_size = 16;
Epetra_SerialComm comm;
Epetra_Map map( problem_size, 0, comm );
std::vector<double> x_vector( problem_size );
Epetra_Vector x( View, map, &x_vector[0] );
std::vector<double> x_aztec_vector( problem_size );
Epetra_Vector x_aztec( View, map, &x_aztec_vector[0] );
std::vector<double> b_vector( problem_size, 0.4 );
Epetra_Vector b( View, map, &b_vector[0] );
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, map, problem_size ) );
double lower_diag = -0.1;
double diag = 2.4;
double upper_diag = -0.1;
int global_row = 0;
int lower_row = 0;
int upper_row = 0;
for ( int i = 0; i < problem_size; ++i )
{
global_row = A->GRID(i);
lower_row = i-1;
upper_row = i+1;
if ( lower_row > -1 )
{
A->InsertGlobalValues( global_row, 1, &lower_diag, &lower_row );
}
A->InsertGlobalValues( global_row, 1, &diag, &global_row );
if ( upper_row < problem_size )
{
A->InsertGlobalValues( global_row, 1, &upper_diag, &upper_row );
}
}
A->FillComplete();
double spec_rad_A = HMCSA::OperatorTools::spectralRadius( A );
std::cout << std::endl <<
"Operator spectral radius: " << spec_rad_A << std::endl;
Teuchos::RCP<Epetra_LinearProblem> linear_problem = Teuchos::rcp(
new Epetra_LinearProblem( A.getRawPtr(), &x, &b ) );
HMCSA::JacobiPreconditioner preconditioner( linear_problem );
preconditioner.precondition();
HMCSA::MCSA mcsa_solver( linear_problem );
mcsa_solver.iterate( 100, 1.0e-8, 100, 1.0e-8 );
std::cout << "MCSA ITERS: " << mcsa_solver.getNumIters() << std::endl;
Epetra_LinearProblem aztec_linear_problem( A.getRawPtr(), &x_aztec, &b );
AztecOO aztec_solver( aztec_linear_problem );
aztec_solver.SetAztecOption( AZ_solver, AZ_gmres );
aztec_solver.Iterate( 100, 1.0e-8 );
std::vector<double> error_vector( problem_size );
Epetra_Vector error( View, map, &error_vector[0] );
for (int i = 0; i < problem_size; ++i)
{
error[i] = x[i] - x_aztec[i];
}
double error_norm;
error.Norm2( &error_norm );
std::cout << std::endl <<
"Aztec GMRES vs. MCSA absolute error L2 norm: " <<
error_norm << std::endl;
}
/* Computes the approximate Schur complement for the wide separator */
Teuchos::RCP<Epetra_CrsMatrix> computeApproxWideSchur(shylu_config *config,
shylu_symbolic *ssym, // symbolic structure
Epetra_CrsMatrix *G, Epetra_CrsMatrix *R,
Epetra_LinearProblem *LP, Amesos_BaseSolver *solver,
Ifpack_Preconditioner *ifSolver, Epetra_CrsMatrix *C,
Epetra_Map *localDRowMap)
{
int i;
double relative_thres = config->relative_threshold;
// Need to create local G (block diagonal portion) , R, C
// Get row map of G
//Epetra_Map CrMap = C->RowMap();
//int *c_rows = CrMap.MyGlobalElements();
//int *c_cols = (C->ColMap()).MyGlobalElements();
//int c_totalElems = CrMap.NumGlobalElements();
//int c_localElems = CrMap.NumMyElements();
//int c_localcolElems = (C->ColMap()).NumMyElements();
Epetra_Map GrMap = G->RowMap();
int *g_rows = GrMap.MyGlobalElements();
//int g_totalElems = GrMap.NumGlobalElements();
int g_localElems = GrMap.NumMyElements();
//Epetra_Map RrMap = R->RowMap();
//int *r_rows = RrMap.MyGlobalElements();
//int *r_cols = (R->ColMap()).MyGlobalElements();
//int r_totalElems = RrMap.NumGlobalElements();
//int r_localElems = RrMap.NumMyElements();
//int r_localcolElems = (R->ColMap()).NumMyElements();
Epetra_SerialComm LComm;
Epetra_Map G_localRMap (-1, g_localElems, g_rows, 0, LComm);
int nentries1, gid;
// maxentries is the maximum of all three possible matrices as the arrays
// are reused between the three
int maxentries = max(C->MaxNumEntries(), R->MaxNumEntries());
maxentries = max(maxentries, G->MaxNumEntries());
double *values1 = new double[maxentries];
double *values2 = new double[maxentries];
double *values3 = new double[maxentries];
int *indices1 = new int[maxentries];
int *indices2 = new int[maxentries];
int *indices3 = new int[maxentries];
// Sbar - Approximate Schur complement
Teuchos::RCP<Epetra_CrsMatrix> Sbar = Teuchos::rcp(new Epetra_CrsMatrix(
Copy, GrMap, g_localElems));
// Include only the block diagonal elements of G in localG
Epetra_CrsMatrix localG(Copy, G_localRMap, G->MaxNumEntries(), false);
int cnt, scnt;
for (i = 0; i < g_localElems ; i++)
{
gid = g_rows[i];
G->ExtractGlobalRowCopy(gid, maxentries, nentries1, values1, indices1);
cnt = 0;
scnt = 0;
for (int j = 0 ; j < nentries1 ; j++)
{
if (G->LRID(indices1[j]) != -1)
{
values2[cnt] = values1[j];
indices2[cnt++] = indices1[j];
}
else
{
// Add it to Sbar immediately
values3[scnt] = values1[j];
indices3[scnt++] = indices1[j];
}
}
localG.InsertGlobalValues(gid, cnt, values2, indices2);
Sbar->InsertGlobalValues(gid, scnt, values3, indices3);
}
localG.FillComplete();
//cout << "Created local G matrix" << endl;
int nvectors = 16;
/*ShyLU_Probing_Operator probeop(&localG, &localR, LP, solver, &localC,
localDRowMap, nvectors);*/
ShyLU_Local_Schur_Operator probeop(config, ssym, &localG, R, LP, solver,
ifSolver, C, localDRowMap, nvectors);
#ifdef DUMP_MATRICES
//ostringstream fnamestr;
//fnamestr << "localC" << C->Comm().MyPID() << ".mat";
//string Cfname = fnamestr.str();
//EpetraExt::RowMatrixToMatlabFile(Cfname.c_str(), localC);
//Epetra_Map defMapg(-1, g_localElems, 0, localG.Comm());
//EpetraExt::ViewTransform<Epetra_CrsMatrix> * ReIdx_MatTransg =
//new EpetraExt::CrsMatrix_Reindex( defMapg );
//Epetra_CrsMatrix t2G = (*ReIdx_MatTransg)( localG );
//ReIdx_MatTransg->fwd();
//EpetraExt::RowMatrixToMatlabFile("localG.mat", t2G);
#endif
//cout << " totalElems in Schur Complement" << totalElems << endl;
//cout << myPID << " localElems" << localElems << endl;
// **************** Two collectives here *********************
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
ftime.start();
#endif
#ifdef TIMING_OUTPUT
Teuchos::Time app_time("setup time");
#endif
int nentries;
// size > maxentries as there could be fill
// TODO: Currently the size of the two arrays can be one, Even if we switch
// the loop below the size of the array required is nvectors. Fix it
double *values = new double[nvectors];
int *indices = new int[nvectors];
double *vecvalues;
#ifdef SHYLU_DEBUG
// mfh 25 May 2015: Don't declare this variable if it's not used.
// It's only used if SHYLU_DEBUG is defined.
int dropped = 0;
#endif // SHYLU_DEBUG
double *maxvalue = new double[nvectors];
#ifdef TIMING_OUTPUT
ftime.start();
#endif
int findex = g_localElems / nvectors ;
int cindex;
// int mypid = C->Comm().MyPID(); // unused
Epetra_MultiVector probevec (G_localRMap, nvectors);
Epetra_MultiVector Scol (G_localRMap, nvectors);
probevec.PutScalar(0.0);
for (i = 0 ; i < findex*nvectors ; i+=nvectors)
{
// Set the probevec to find block columns of S.
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
// TODO: Can do better than this, just need to go to the column map
// of C, there might be null columns in C
probevec.ReplaceGlobalValue(g_rows[cindex], k, 1.0);
//if (mypid == 0)
//cout << "Changing row to 1.0 " << g_rows[cindex] << endl;
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
// Reset the probevec to all zeros.
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
probevec.ReplaceGlobalValue(g_rows[cindex], k, 0.0);
}
Scol.MaxValue(maxvalue);
nentries = 0;
for (int j = 0 ; j < g_localElems ; j++)
{
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
vecvalues = Scol[k];
if ((g_rows[cindex] == g_rows[j]) ||
(abs(vecvalues[j]/maxvalue[k]) > relative_thres))
// diagonal entry or large entry.
{
values[nentries] = vecvalues[j];
indices[nentries++] = g_rows[cindex];
}
#ifdef SHYLU_DEBUG
else if (vecvalues[j] != 0.0)
{
dropped++;
}
#endif // SHYLU_DEBUG
}
Sbar->InsertGlobalValues(g_rows[j], nentries, values,
indices);
nentries = 0;
}
}
if (i < g_localElems)
{
nvectors = g_localElems - i;
probeop.ResetTempVectors(nvectors);
Epetra_MultiVector probevec1 (G_localRMap, nvectors);
Epetra_MultiVector Scol1 (G_localRMap, nvectors);
probevec1.PutScalar(0.0);
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
// TODO: Can do better than this, just need to go to the column map
// of C, there might be null columns in C
probevec1.ReplaceGlobalValue(g_rows[cindex], k, 1.0);
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec1, Scol1);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
Scol1.MaxValue(maxvalue);
nentries = 0;
for (int j = 0 ; j < g_localElems ; j++)
{
//cout << "MAX" << maxvalue << endl;
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
vecvalues = Scol1[k];
//nentries = 0; // inserting one entry in each row for now
if ((g_rows[cindex] == g_rows[j]) ||
(abs(vecvalues[j]/maxvalue[k]) > relative_thres))
// diagonal entry or large entry.
{
values[nentries] = vecvalues[j];
indices[nentries++] = g_rows[cindex];
}
#ifdef SHYLU_DEBUG
else if (vecvalues[j] != 0.0)
{
dropped++;
}
#endif // SHYLU_DEBUG
}
Sbar->InsertGlobalValues(g_rows[j], nentries, values,
indices);
nentries = 0;
}
}
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time in finding and dropping entries" << ftime.totalElapsedTime()
<< endl;
ftime.reset();
cout << "Time in Apply of probing" << app_time.totalElapsedTime() << endl;
probeop.PrintTimingInfo();
#endif
Sbar->FillComplete();
#ifdef DUMP_MATRICES
Epetra_Map defMap2(-1, g_localElems, 0, C->Comm());
EpetraExt::ViewTransform<Epetra_CrsMatrix> * ReIdx_MatTrans2 =
new EpetraExt::CrsMatrix_Reindex( defMap2 );
Epetra_CrsMatrix t2S = (*ReIdx_MatTrans2)( *Sbar );
ReIdx_MatTrans2->fwd();
EpetraExt::RowMatrixToMatlabFile("Schur.mat", t2S);
#endif
#ifdef SHYLU_DEBUG
cout << "#dropped entries" << dropped << endl;
#endif
delete[] values;
delete[] indices;
delete[] values1;
delete[] indices1;
delete[] values2;
delete[] indices2;
delete[] values3;
delete[] indices3;
delete[] maxvalue;
return Sbar;
}
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( DomainTransporter, Cutoff2 )
{
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
typedef MCLS::AdjointHistory<int> HistoryType;
typedef std::mt19937 rng_type;
typedef MCLS::AdjointDomain<VectorType,MatrixType,rng_type> DomainType;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int comm_rank = comm->getRank();
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear operator and solution vector.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 1 );
Teuchos::Array<double> values( 1 );
for ( int i = 1; i < global_num_rows; ++i )
{
global_columns[0] = i-1;
values[0] = -0.5;
A->InsertGlobalValues( i, global_columns().size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-1;
values[0] = -0.5;
A->InsertGlobalValues( global_num_rows-1, global_columns.size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> B = MT::copyTranspose(*A);
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *A );
// Build the adjoint domain.
Teuchos::ParameterList plist;
plist.set<int>( "Overlap Size", 2 );
Teuchos::RCP<DomainType> domain = Teuchos::rcp( new DomainType( B, x, plist ) );
Teuchos::RCP<MCLS::PRNG<rng_type> > rng = Teuchos::rcp(
new MCLS::PRNG<rng_type>( comm->getRank() ) );
domain->setRNG( rng );
// Build the domain transporter.
double weight = 3.0;
double relative_cutoff = weight / 4 + 0.01;
MCLS::DomainTransporter<DomainType> transporter( domain, plist );
domain->setCutoff( relative_cutoff );
// Transport histories through the domain.
for ( int i = 0; i < global_num_rows-2; ++i )
{
if ( comm_rank == comm_size - 1 )
{
if ( i >= local_num_rows*comm_rank && i < local_num_rows*(comm_rank+1) )
{
HistoryType history( i, i, weight );
history.live();
transporter.transport( history );
TEST_EQUALITY( history.globalState(), i+2 );
TEST_EQUALITY( history.weight(), weight / 4 );
TEST_EQUALITY( history.event(), MCLS::Event::CUTOFF );
TEST_ASSERT( !history.alive() );
}
}
else
{
if ( i >= local_num_rows*comm_rank && i < 1+local_num_rows*(comm_rank+1) )
{
HistoryType history( i, i, weight );
history.live();
transporter.transport( history );
TEST_EQUALITY( history.globalState(), i+2 );
TEST_EQUALITY( history.weight(), weight / 4 );
TEST_EQUALITY( history.event(), MCLS::Event::CUTOFF );
TEST_ASSERT( !history.alive() );
}
}
}
// Check the tally.
domain->domainTally()->combineSetTallies( comm );
Teuchos::ArrayRCP<const double> x_view =
VT::view( *x );
double x_val = weight;
for ( int i = 0; i < local_num_rows; ++i )
{
if ( comm_rank == comm_size-1 && i == local_num_rows-1 )
{
TEST_EQUALITY( x_view[i], 0.0 );
}
else if ( comm_rank == comm_size-1 && i == local_num_rows-2 )
{
TEST_EQUALITY( x_view[i], 3.0/2.0 );
}
else if ( comm_rank == 0 && i == 0 )
{
TEST_EQUALITY( x_view[i], x_val );
}
else if ( comm_rank == 0 || i > 2 )
{
TEST_EQUALITY( x_view[i], 3.0*x_val/2.0 );
}
else
{
TEST_EQUALITY( x_view[i], 3.0*x_val );
}
}
}
/* Computes the approximate Schur complement for the wide separator
using guided probing*/
Teuchos::RCP<Epetra_CrsMatrix> computeSchur_GuidedProbing
(
shylu_config *config,
shylu_symbolic *ssym, // symbolic structure
shylu_data *data, // numeric structure
Epetra_Map *localDRowMap
)
{
int i;
double relative_thres = config->relative_threshold;
Epetra_CrsMatrix *G = ssym->G.getRawPtr();
Epetra_CrsMatrix *R = ssym->R.getRawPtr();
Epetra_LinearProblem *LP = ssym->LP.getRawPtr();
Amesos_BaseSolver *solver = ssym->Solver.getRawPtr();
Ifpack_Preconditioner *ifSolver = ssym->ifSolver.getRawPtr();
Epetra_CrsMatrix *C = ssym->C.getRawPtr();
// Need to create local G (block diagonal portion) , R, C
// Get row map of G
Epetra_Map CrMap = C->RowMap();
int *c_rows = CrMap.MyGlobalElements();
int *c_cols = (C->ColMap()).MyGlobalElements();
//int c_totalElems = CrMap.NumGlobalElements();
int c_localElems = CrMap.NumMyElements();
int c_localcolElems = (C->ColMap()).NumMyElements();
Epetra_Map GrMap = G->RowMap();
int *g_rows = GrMap.MyGlobalElements();
//int g_totalElems = GrMap.NumGlobalElements();
int g_localElems = GrMap.NumMyElements();
Epetra_Map RrMap = R->RowMap();
int *r_rows = RrMap.MyGlobalElements();
int *r_cols = (R->ColMap()).MyGlobalElements();
//int r_totalElems = RrMap.NumGlobalElements();
int r_localElems = RrMap.NumMyElements();
int r_localcolElems = (R->ColMap()).NumMyElements();
Epetra_SerialComm LComm;
Epetra_Map C_localRMap (-1, c_localElems, c_rows, 0, LComm);
Epetra_Map C_localCMap (-1, c_localcolElems, c_cols, 0, LComm);
Epetra_Map G_localRMap (-1, g_localElems, g_rows, 0, LComm);
Epetra_Map R_localRMap (-1, r_localElems, r_rows, 0, LComm);
Epetra_Map R_localCMap (-1, r_localcolElems, r_cols, 0, LComm);
//cout << "#local rows" << g_localElems << "#non zero local cols" << c_localcolElems << endl;
#ifdef DEBUG
cout << "DEBUG MODE" << endl;
int nrows = C->RowMap().NumMyElements();
assert(nrows == localDRowMap->NumGlobalElements());
int gids[nrows], gids1[nrows];
C_localRMap.MyGlobalElements(gids);
localDRowMap->MyGlobalElements(gids1);
cout << "Comparing R's domain map with D's row map" << endl;
for (int i = 0; i < nrows; i++)
{
assert(gids[i] == gids1[i]);
}
#endif
int nentries1, gid;
// maxentries is the maximum of all three possible matrices as the arrays
// are reused between the three
int maxentries = max(C->MaxNumEntries(), R->MaxNumEntries());
maxentries = max(maxentries, G->MaxNumEntries());
double *values1 = new double[maxentries];
double *values2 = new double[maxentries];
double *values3 = new double[maxentries];
int *indices1 = new int[maxentries];
int *indices2 = new int[maxentries];
int *indices3 = new int[maxentries];
//cout << "Creating local matrices" << endl;
int err;
Epetra_CrsMatrix localC(Copy, C_localRMap, C->MaxNumEntries(), false);
for (i = 0; i < c_localElems ; i++)
{
gid = c_rows[i];
err = C->ExtractGlobalRowCopy(gid, maxentries, nentries1, values1,
indices1);
assert (err == 0);
//if (nentries1 > 0) // TODO: Later
//{
err = localC.InsertGlobalValues(gid, nentries1, values1, indices1);
assert (err == 0);
//}
}
localC.FillComplete(G_localRMap, C_localRMap);
//cout << "Created local C matrix" << endl;
Epetra_CrsMatrix localR(Copy, R_localRMap, R->MaxNumEntries(), false);
for (i = 0; i < r_localElems ; i++)
{
gid = r_rows[i];
R->ExtractGlobalRowCopy(gid, maxentries, nentries1, values1, indices1);
localR.InsertGlobalValues(gid, nentries1, values1, indices1);
}
localR.FillComplete(*localDRowMap, R_localRMap);
//cout << "Created local R matrix" << endl;
// Sbar - Approximate Schur complement
Teuchos::RCP<Epetra_CrsMatrix> Sbar = Teuchos::rcp(new Epetra_CrsMatrix(
Copy, GrMap, g_localElems));
// Include only the block diagonal elements of G in localG
Epetra_CrsMatrix localG(Copy, G_localRMap, G->MaxNumEntries(), false);
int cnt, scnt;
for (i = 0; i < g_localElems ; i++)
{
gid = g_rows[i];
G->ExtractGlobalRowCopy(gid, maxentries, nentries1, values1, indices1);
cnt = 0;
scnt = 0;
for (int j = 0 ; j < nentries1 ; j++)
{
if (G->LRID(indices1[j]) != -1)
{
values2[cnt] = values1[j];
indices2[cnt++] = indices1[j];
}
else
{
// Add it to Sbar immediately
values3[scnt] = values1[j];
indices3[scnt++] = indices1[j];
}
}
localG.InsertGlobalValues(gid, cnt, values2, indices2);
Sbar->InsertGlobalValues(gid, scnt, values3, indices3);
}
localG.FillComplete();
cout << "Created local G matrix" << endl;
int nvectors = 16;
ShyLU_Probing_Operator probeop(config, ssym, &localG, &localR, LP, solver,
ifSolver, &localC, localDRowMap, nvectors);
#ifdef DUMP_MATRICES
//ostringstream fnamestr;
//fnamestr << "localC" << C->Comm().MyPID() << ".mat";
//string Cfname = fnamestr.str();
//EpetraExt::RowMatrixToMatlabFile(Cfname.c_str(), localC);
//Epetra_Map defMapg(-1, g_localElems, 0, localG.Comm());
//EpetraExt::ViewTransform<Epetra_CrsMatrix> * ReIdx_MatTransg =
//new EpetraExt::CrsMatrix_Reindex( defMapg );
//Epetra_CrsMatrix t2G = (*ReIdx_MatTransg)( localG );
//ReIdx_MatTransg->fwd();
//EpetraExt::RowMatrixToMatlabFile("localG.mat", t2G);
#endif
//cout << " totalElems in Schur Complement" << totalElems << endl;
//cout << myPID << " localElems" << localElems << endl;
// **************** Two collectives here *********************
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
#endif
#ifdef TIMING_OUTPUT
Teuchos::Time app_time("setup time");
#endif
Teuchos::RCP<Epetra_CrsGraph> lSGraph = Teuchos::RCP<Epetra_CrsGraph> (
new Epetra_CrsGraph(Copy, G_localRMap, maxentries));
if (data->num_compute % config->reset_iter == 0)
{
int nentries;
// size > maxentries as there could be fill
// TODO: Currently the size of the two arrays can be one, Even if we switch
// the loop below the size of the array required is nvectors. Fix it
double *values = new double[g_localElems];
int *indices = new int[g_localElems];
double *vecvalues;
int dropped = 0;
double *maxvalue = new double[nvectors];
#ifdef TIMING_OUTPUT
ftime.start();
#endif
int findex = g_localElems / nvectors ;
int cindex;
// int mypid = C->Comm().MyPID(); // unused
Epetra_MultiVector probevec(G_localRMap, nvectors);
Epetra_MultiVector Scol(G_localRMap, nvectors);
for (i = 0 ; i < findex*nvectors ; i+=nvectors)
{
probevec.PutScalar(0.0); // TODO: Move it out
for (int k = 0; k < nvectors; k++)
{
cindex = k+i;
// TODO: Can do better than this, just need to go to the column map
// of C, there might be null columns in C
// Not much of use for Shasta 2x2 .. Later.
probevec.ReplaceGlobalValue(g_rows[cindex], k, 1.0);
//if (mypid == 0)
//cout << "Changing row to 1.0 " << g_rows[cindex] << endl;
}
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
Scol.MaxValue(maxvalue);
for (int k = 0; k < nvectors; k++) //TODO:Need to switch these loops
{
cindex = k+i;
vecvalues = Scol[k];
//cout << "MAX" << maxvalue << endl;
for (int j = 0 ; j < g_localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (g_rows[cindex] == g_rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = g_rows[cindex];
nentries++;
err = Sbar->InsertGlobalValues(g_rows[j], nentries, values,
indices);
assert(err >= 0);
err = lSGraph->InsertGlobalIndices(g_rows[j], nentries,
indices);
assert(err >= 0);
}
else if (abs(vecvalues[j]/maxvalue[k]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = g_rows[cindex];
nentries++;
err = Sbar->InsertGlobalValues(g_rows[j], nentries, values,
indices);
assert(err >= 0);
err = lSGraph->InsertGlobalIndices(g_rows[j], nentries,
indices);
assert(err >= 0);
}
else
{
if (vecvalues[j] != 0.0)
{
dropped++;
//cout << "vecvalues[j]" << vecvalues[j] <<
// " max" << maxvalue[k] << endl;
}
}
}
}
}
probeop.ResetTempVectors(1);
for ( ; i < g_localElems ; i++)
{
// TODO: Can move the next two decalarations outside the loop
Epetra_MultiVector probevec(G_localRMap, 1);
Epetra_MultiVector Scol(G_localRMap, 1);
probevec.PutScalar(0.0);
// TODO: Can do better than this, just need to go to the column map
// of C, there might be null columns in C
probevec.ReplaceGlobalValue(g_rows[i], 0, 1.0);
#ifdef TIMING_OUTPUT
app_time.start();
#endif
probeop.Apply(probevec, Scol);
#ifdef TIMING_OUTPUT
app_time.stop();
#endif
vecvalues = Scol[0];
Scol.MaxValue(maxvalue);
//cout << "MAX" << maxvalue << endl;
for (int j = 0 ; j < g_localElems ; j++)
{
nentries = 0; // inserting one entry in each row for now
if (g_rows[i] == g_rows[j]) // diagonal entry
{
values[nentries] = vecvalues[j];
indices[nentries] = g_rows[i];
nentries++;
err = Sbar->InsertGlobalValues(g_rows[j], nentries, values, indices);
assert(err >= 0);
err = lSGraph->InsertGlobalIndices(g_rows[j], nentries, indices);
assert(err >= 0);
}
else if (abs(vecvalues[j]/maxvalue[0]) > relative_thres)
{
values[nentries] = vecvalues[j];
indices[nentries] = g_rows[i];
nentries++;
err = Sbar->InsertGlobalValues(g_rows[j], nentries, values, indices);
assert(err >= 0);
err = lSGraph->InsertGlobalIndices(g_rows[j], nentries, indices);
assert(err >= 0);
}
else
{
if (vecvalues[j] != 0.0) dropped++;
}
}
}
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time in finding and dropping entries" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
#ifdef TIMING_OUTPUT
cout << "Time in Apply of probing" << app_time.totalElapsedTime() << endl;
#endif
probeop.PrintTimingInfo();
Sbar->FillComplete();
lSGraph->FillComplete();
data->localSbargraph = lSGraph;
#ifdef DUMP_MATRICES
Epetra_Map defMap2(-1, g_localElems, 0, C->Comm());
EpetraExt::ViewTransform<Epetra_CrsMatrix> * ReIdx_MatTrans2 =
new EpetraExt::CrsMatrix_Reindex( defMap2 );
Epetra_CrsMatrix t2S = (*ReIdx_MatTrans2)( *Sbar );
ReIdx_MatTrans2->fwd();
EpetraExt::RowMatrixToMatlabFile("Schur.mat", t2S);
#endif
cout << "#dropped entries" << dropped << endl;
delete[] values;
delete[] indices;
delete[] maxvalue;
}
else
{
if (((data->num_compute-1) % config->reset_iter) == 0)
{
// We recomputed the Schur complement with dropping for the last
// compute. Reset the prober with the new orthogonal vectors for
// the Sbar from the previous iteration.
Teuchos::ParameterList pList;
Teuchos::RCP<Isorropia::Epetra::Prober> gprober =
Teuchos::RCP<Isorropia::Epetra::Prober> (new
Isorropia::Epetra::Prober(
data->localSbargraph.getRawPtr(), pList, false));
gprober->color();
data->guided_prober = gprober;
}
// Use the prober to probe the probeop for the sparsity pattern
// add that to Sbar and call Fill complete
int nvectors = data->guided_prober->getNumOrthogonalVectors();
cout << "Number of Orthogonal Vectors for guided probing" << nvectors
<< endl;
probeop.ResetTempVectors(nvectors);
Teuchos::RCP<Epetra_CrsMatrix> blockdiag_Sbar =
data->guided_prober->probe(probeop);
int maxentries = blockdiag_Sbar->GlobalMaxNumEntries();
int *indices = new int[maxentries];
double *values = new double[maxentries];
int numentries;
for (int i = 0; i < blockdiag_Sbar->NumGlobalRows() ; i++)
{
int gid = blockdiag_Sbar->GRID(i);
blockdiag_Sbar->ExtractGlobalRowCopy(gid, maxentries, numentries,
values, indices);
Sbar->InsertGlobalValues(gid, numentries, values, indices);
}
Sbar->FillComplete();
delete[] indices;
delete[] values;
}
delete[] values1;
delete[] indices1;
delete[] values2;
delete[] indices2;
delete[] values3;
delete[] indices3;
return Sbar;
}
// helper routines
bool SplitMatrix2x2(Teuchos::RCP<const Epetra_CrsMatrix> A,
const Epetra_Map& A11rowmap,
const Epetra_Map& A22rowmap,
Teuchos::RCP<Epetra_CrsMatrix>& A11,
Teuchos::RCP<Epetra_CrsMatrix>& A12,
Teuchos::RCP<Epetra_CrsMatrix>& A21,
Teuchos::RCP<Epetra_CrsMatrix>& A22)
{
if (A==Teuchos::null)
{
std::cout << "ERROR: SplitMatrix2x2: A==null on entry" << std::endl;
return false;
}
const Epetra_Comm& Comm = A->Comm();
const Epetra_Map& A22map = A22rowmap;
const Epetra_Map& A11map = A11rowmap;
//----------------------------- create a parallel redundant map of A22map
std::map<int,int> a22gmap;
{
std::vector<int> a22global(A22map.NumGlobalElements());
int count=0;
for (int proc=0; proc<Comm.NumProc(); ++proc)
{
int length = 0;
if (proc==Comm.MyPID())
{
for (int i=0; i<A22map.NumMyElements(); ++i)
{
a22global[count+length] = A22map.GID(i);
++length;
}
}
Comm.Broadcast(&length,1,proc);
Comm.Broadcast(&a22global[count],length,proc);
count += length;
}
if (count != A22map.NumGlobalElements())
{
std::cout << "ERROR SplitMatrix2x2: mismatch in dimensions" << std::endl;
return false;
}
// create the map
for (int i=0; i<count; ++i)
a22gmap[a22global[i]] = 1;
a22global.clear();
}
//--------------------------------------------------- create matrix A22
A22 = Teuchos::rcp(new Epetra_CrsMatrix(Copy,A22map,100));
{
std::vector<int> a22gcindices(100);
std::vector<double> a22values(100);
for (int i=0; i<A->NumMyRows(); ++i)
{
const int grid = A->GRID(i);
if (A22map.MyGID(grid)==false)
continue;
int numentries;
double* values;
int* cindices;
int err = A->ExtractMyRowView(i,numentries,values,cindices);
if (err)
{
std::cout << "ERROR: SplitMatrix2x2: A->ExtractMyRowView returned " << err << std::endl;
return false;
}
if (numentries>(int)a22gcindices.size())
{
a22gcindices.resize(numentries);
a22values.resize(numentries);
}
int count=0;
for (int j=0; j<numentries; ++j)
{
const int gcid = A->ColMap().GID(cindices[j]);
// see whether we have gcid in a22gmap
std::map<int,int>::iterator curr = a22gmap.find(gcid);
if (curr==a22gmap.end()) continue;
//std::cout << gcid << " ";
a22gcindices[count] = gcid;
a22values[count] = values[j];
++count;
}
//std::cout << std::endl; fflush(stdout);
// add this filtered row to A22
err = A22->InsertGlobalValues(grid,count,&a22values[0],&a22gcindices[0]);
if (err<0)
{
std::cout << "ERROR: SplitMatrix2x2: A->InsertGlobalValues returned " << err << std::endl;
return false;
}
} //for (int i=0; i<A->NumMyRows(); ++i)
a22gcindices.clear();
a22values.clear();
}
A22->FillComplete();
A22->OptimizeStorage();
//----------------------------------------------------- create matrix A11
A11 = Teuchos::rcp(new Epetra_CrsMatrix(Copy,A11map,100));
{
std::vector<int> a11gcindices(100);
std::vector<double> a11values(100);
for (int i=0; i<A->NumMyRows(); ++i)
{
const int grid = A->GRID(i);
if (A11map.MyGID(grid)==false) continue;
int numentries;
double* values;
int* cindices;
int err = A->ExtractMyRowView(i,numentries,values,cindices);
if (err)
{
std::cout << "ERROR: SplitMatrix2x2: A->ExtractMyRowView returned " << err << std::endl;
return false;
}
if (numentries>(int)a11gcindices.size())
{
a11gcindices.resize(numentries);
a11values.resize(numentries);
}
int count=0;
for (int j=0; j<numentries; ++j)
{
const int gcid = A->ColMap().GID(cindices[j]);
// see whether we have gcid as part of a22gmap
std::map<int,int>::iterator curr = a22gmap.find(gcid);
if (curr!=a22gmap.end()) continue;
a11gcindices[count] = gcid;
a11values[count] = values[j];
++count;
}
err = A11->InsertGlobalValues(grid,count,&a11values[0],&a11gcindices[0]);
if (err<0)
{
std::cout << "ERROR: SplitMatrix2x2: A->InsertGlobalValues returned " << err << std::endl;
return false;
}
} // for (int i=0; i<A->NumMyRows(); ++i)
a11gcindices.clear();
a11values.clear();
}
A11->FillComplete();
A11->OptimizeStorage();
//---------------------------------------------------- create matrix A12
A12 = Teuchos::rcp(new Epetra_CrsMatrix(Copy,A11map,100));
{
std::vector<int> a12gcindices(100);
std::vector<double> a12values(100);
for (int i=0; i<A->NumMyRows(); ++i)
{
const int grid = A->GRID(i);
if (A11map.MyGID(grid)==false) continue;
int numentries;
double* values;
int* cindices;
int err = A->ExtractMyRowView(i,numentries,values,cindices);
if (err)
{
std::cout << "ERROR: SplitMatrix2x2: A->ExtractMyRowView returned " << err << std::endl;
return false;
}
if (numentries>(int)a12gcindices.size())
{
a12gcindices.resize(numentries);
a12values.resize(numentries);
}
int count=0;
for (int j=0; j<numentries; ++j)
{
const int gcid = A->ColMap().GID(cindices[j]);
// see whether we have gcid as part of a22gmap
std::map<int,int>::iterator curr = a22gmap.find(gcid);
if (curr==a22gmap.end()) continue;
a12gcindices[count] = gcid;
a12values[count] = values[j];
++count;
}
err = A12->InsertGlobalValues(grid,count,&a12values[0],&a12gcindices[0]);
if (err<0)
{
std::cout << "ERROR: SplitMatrix2x2: A->InsertGlobalValues returned " << err << std::endl;
return false;
}
} // for (int i=0; i<A->NumMyRows(); ++i)
a12values.clear();
a12gcindices.clear();
}
A12->FillComplete(A22map,A11map);
A12->OptimizeStorage();
//----------------------------------------------------------- create A21
A21 = Teuchos::rcp(new Epetra_CrsMatrix(Copy,A22map,100));
{
std::vector<int> a21gcindices(100);
std::vector<double> a21values(100);
for (int i=0; i<A->NumMyRows(); ++i)
{
const int grid = A->GRID(i);
if (A22map.MyGID(grid)==false) continue;
int numentries;
double* values;
int* cindices;
int err = A->ExtractMyRowView(i,numentries,values,cindices);
if (err)
{
std::cout << "ERROR: SplitMatrix2x2: A->ExtractMyRowView returned " << err << std::endl;
return false;
}
if (numentries>(int)a21gcindices.size())
{
a21gcindices.resize(numentries);
a21values.resize(numentries);
}
int count=0;
for (int j=0; j<numentries; ++j)
{
const int gcid = A->ColMap().GID(cindices[j]);
// see whether we have gcid as part of a22gmap
std::map<int,int>::iterator curr = a22gmap.find(gcid);
if (curr!=a22gmap.end()) continue;
a21gcindices[count] = gcid;
a21values[count] = values[j];
++count;
}
err = A21->InsertGlobalValues(grid,count,&a21values[0],&a21gcindices[0]);
if (err<0)
{
std::cout << "ERROR: SplitMatrix2x2: A->InsertGlobalValues returned " << err << std::endl;
return false;
}
} // for (int i=0; i<A->NumMyRows(); ++i)
a21values.clear();
a21gcindices.clear();
}
A21->FillComplete(A11map,A22map);
A21->OptimizeStorage();
//-------------------------------------------------------------- tidy up
a22gmap.clear();
return true;
}
//---------------------------------------------------------------------------//
// Test templates
//---------------------------------------------------------------------------//
TEUCHOS_UNIT_TEST( FixedPointSolverManager, one_by_one )
{
typedef Epetra_Vector VectorType;
typedef MCLS::VectorTraits<VectorType> VT;
typedef Epetra_RowMatrix MatrixType;
typedef MCLS::MatrixTraits<VectorType,MatrixType> MT;
Teuchos::RCP<const Teuchos::Comm<int> > comm =
Teuchos::DefaultComm<int>::getComm();
Teuchos::RCP<Epetra_Comm> epetra_comm = getEpetraComm( comm );
int comm_size = comm->getSize();
int local_num_rows = 10;
int global_num_rows = local_num_rows*comm_size;
Teuchos::RCP<Epetra_Map> map = Teuchos::rcp(
new Epetra_Map( global_num_rows, 0, *epetra_comm ) );
// Build the linear system.
Teuchos::RCP<Epetra_CrsMatrix> A =
Teuchos::rcp( new Epetra_CrsMatrix( Copy, *map, 0 ) );
Teuchos::Array<int> global_columns( 3 );
Teuchos::Array<double> values( 3 );
global_columns[0] = 0;
global_columns[1] = 1;
global_columns[2] = 2;
values[0] = 0.14;
values[1] = 0.14;
values[2] = 1.0;
A->InsertGlobalValues( 0, global_columns.size(),
&values[0], &global_columns[0] );
for ( int i = 1; i < global_num_rows-1; ++i )
{
global_columns[0] = i-1;
global_columns[1] = i;
global_columns[2] = i+1;
values[0] = 0.14;
values[1] = 1.0;
values[2] = 0.14;
A->InsertGlobalValues( i, global_columns.size(),
&values[0], &global_columns[0] );
}
global_columns[0] = global_num_rows-3;
global_columns[1] = global_num_rows-2;
global_columns[2] = global_num_rows-1;
values[0] = 0.14;
values[1] = 0.14;
values[2] = 1.0;
A->InsertGlobalValues( global_num_rows-1, global_columns.size(),
&values[0], &global_columns[0] );
A->FillComplete();
Teuchos::RCP<MatrixType> B = A;
// Build the LHS. Put a large positive number here to be sure we are
// clear the vector before solving.
Teuchos::RCP<VectorType> x = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *x, 0.0 );
// Build the RHS with negative numbers. this gives us a negative
// solution.
Teuchos::RCP<VectorType> b = MT::cloneVectorFromMatrixRows( *B );
VT::putScalar( *b, -1.0 );
// Solver parameters.
Teuchos::RCP<Teuchos::ParameterList> plist =
Teuchos::rcp( new Teuchos::ParameterList() );
plist->set<int>("Iteration Print Frequency", 1);
plist->set<double>("Convergence Tolerance", 1.0e-8);
plist->set<int>("Maximum Iterations", 100);
plist->set<double>("Richardson Relaxation", 1.0);
// Create the linear problem.
Teuchos::RCP<MCLS::LinearProblem<VectorType,MatrixType> > linear_problem =
Teuchos::rcp( new MCLS::LinearProblem<VectorType,MatrixType>(
B, x, b ) );
// Create the solver.
MCLS::FixedPointSolverManager<VectorType,MatrixType>
solver_manager( linear_problem, comm, plist );
// Solve the problem.
bool converged_status = solver_manager.solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager.getConvergedStatus() );
TEST_EQUALITY( solver_manager.getNumIters(), 15 );
TEST_ASSERT( solver_manager.achievedTol() > 0.0 );
// Check that we got a negative solution.
Teuchos::ArrayRCP<const double> x_view = VT::view(*x);
Teuchos::ArrayRCP<const double>::const_iterator x_view_it;
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it < Teuchos::ScalarTraits<double>::zero() );
}
// Now solve the problem with a positive source.
VT::putScalar( *b, 2.0 );
VT::putScalar( *x, 0.0 );
linear_problem->setLHS(x);
converged_status = solver_manager.solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager.getConvergedStatus() );
TEST_EQUALITY( solver_manager.getNumIters(), 15 );
TEST_ASSERT( solver_manager.achievedTol() > 0.0 );
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it > Teuchos::ScalarTraits<double>::zero() );
}
// Reset the domain and solve again with a positive source.
VT::putScalar( *x, 0.0 );
solver_manager.setProblem( linear_problem );
linear_problem->setLHS(x);
converged_status = solver_manager.solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager.getConvergedStatus() );
TEST_EQUALITY( solver_manager.getNumIters(), 15 );
TEST_ASSERT( solver_manager.achievedTol() > 0.0 );
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it > Teuchos::ScalarTraits<double>::zero() );
}
// Reset both and solve with a negative source.
VT::putScalar( *b, -2.0 );
VT::putScalar( *x, 0.0 );
linear_problem->setLHS(x);
converged_status = solver_manager.solve();
TEST_ASSERT( converged_status );
TEST_ASSERT( solver_manager.getConvergedStatus() );
TEST_EQUALITY( solver_manager.getNumIters(), 15 );
TEST_ASSERT( solver_manager.achievedTol() > 0.0 );
for ( x_view_it = x_view.begin(); x_view_it != x_view.end(); ++x_view_it )
{
TEST_ASSERT( *x_view_it < Teuchos::ScalarTraits<double>::zero() );
}
}
int
main(int argc, char *argv[])
{
using Teuchos::CommandLineProcessor;
using Teuchos::ParameterList;
using Teuchos::RCP;
using Teuchos::rcp;
using std::cout;
using std::endl;
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;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
const int MyPID = Comm.MyPID();
#else // No EPETRA_MPI
Epetra_SerialComm Comm;
const int MyPID = 0;
#endif // EPETRA_MPI
bool verbose = false, debug = false;
int frequency = -1; // frequency of status test output.
int numrhs = 1; // number of right-hand sides to solve for
int maxiters = -1; // maximum number of iterations allowed per linear system
MT tol = 1.0e-8; // relative residual tolerance
// Define command-line arguments
CommandLineProcessor cmdp(false,true);
cmdp.setOption ("verbose", "quiet", &verbose, "Print messages and results.");
cmdp.setOption ("debug", "nodebug", &debug,
"Print debugging information from the solver.");
cmdp.setOption ("frequency", &frequency,
"Solver's frequency for printing residuals (#iters).");
cmdp.setOption ("tol", &tol,
"Relative residual tolerance used by MINRES solver.");
cmdp.setOption ("num-rhs", &numrhs,
"Number of right-hand sides for which to solve (> 0).");
cmdp.setOption ("max-iters", &maxiters,
"Maximum number of iterations per linear system. -1 means "
"we choose, based on problem size.");
// Parse command-line arguments and fetch values
if (cmdp.parse(argc,argv) != CommandLineProcessor::PARSE_SUCCESSFUL)
{
std::cout << "Failed to parse command-line arguments!" << std::endl;
std::cout << "End Result: TEST FAILED" << std::endl;
return -1;
}
// **********************************************************************
// ******************Set up the problem to be solved*********************
// construct diagonal matrix
const int NumGlobalElements = 1e2;
const int m = 4; // number of negative eigenvalues
// Create diagonal matrix with n-m positive and m negative eigenvalues.
Epetra_Map epetraMap( NumGlobalElements, 0, Comm );
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix( Copy, epetraMap, 1 ) );
for ( int k=0; k<epetraMap.NumMyElements(); k++ )
{
int GIDk = epetraMap.GID(k);
double val = 2*(GIDk-m) + 1;
TEUCHOS_ASSERT_EQUALITY( 0, A->InsertGlobalValues( GIDk, 1, &val, &GIDk ) );
}
TEUCHOS_ASSERT_EQUALITY( 0, A->FillComplete() );
TEUCHOS_ASSERT_EQUALITY( 0, A->OptimizeStorage() );
//
// Make some (multi)vectors, for use in testing: an exact solution,
// its corresponding right-hand side, and an initial guess.
//
// Make a random exact solution.
Teuchos::RCP<Epetra_MultiVector> X_exact (new Epetra_MultiVector (epetraMap, numrhs));
X_exact->Seed();
MVT::MvRandom (*X_exact);
// Compute the right-hand side as B = A*X.
Teuchos::RCP<Epetra_MultiVector> B = MVT::Clone (*X_exact, numrhs);
OPT::Apply (*A, *X_exact, *B);
// Choose an initial guess of all zeros.
Teuchos::RCP<Epetra_MultiVector> X = MVT::Clone (*X_exact, numrhs);
MVT::MvInit (*X, ST(0.0));
// **********************************************************************
//
// Set parameters that the user asked us to pick intelligently.
//
// Maximum number of iterations: set according to problem size.
// maxiters=numGlobalElements-1 may not always converge, so we
// set to numGlobalElements+1 for good measure.
if (maxiters == -1)
maxiters = NumGlobalElements + 1;
// In a nonverbose test, the frequency should be set to -1, which
// Belos interprets as "no intermediate status output." Override
// whatever the user may have specified.
if (! verbose)
frequency = -1;
// Silently fix a bad frequency value.
else if (frequency < 0 && frequency != -1)
frequency = -1;
// Validate command-line arguments
TEUCHOS_TEST_FOR_EXCEPTION( tol < 0, std::invalid_argument,
"Relative residual tolerance must be nonnegative, but "
"you supplied tol = " << tol << "." );
TEUCHOS_TEST_FOR_EXCEPTION( numrhs < 1, std::invalid_argument,
"MINRES test requires at least one right-hand side, but "
"you set the number of right-hand sides to "
<< numrhs << "." );
TEUCHOS_TEST_FOR_EXCEPTION( maxiters < 1, std::invalid_argument,
"MINRES test requires at least one iteration, but you "
"set the maximum number of iterations to "
<< maxiters << "." );
const bool proc_verbose = verbose && (MyPID==0); /* Only print on the zero processor */
//
// ********Other information used by block solver***********
// *****************(can be user specified)******************
//
//
ParameterList belosList;
belosList.set( "Maximum Iterations", maxiters ); // Maximum number of iterations allowed
belosList.set( "Convergence Tolerance", tol ); // Relative convergence tolerance requested
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", (int) verbosity );
//
// Construct an unpreconditioned linear problem instance.
//
Belos::LinearProblem< ST, MV, OP > problem (A, X, B);
{
const bool set = problem.setProblem();
TEUCHOS_TEST_FOR_EXCEPTION( set == false, std::logic_error,
"Belos::LinearProblem failed to set up correctly (setP"
"roblem() returned false)! This probably means we imp"
"lemented our test incorrectly." );
}
//
// *******************************************************************
// ****************Start the MINRES iteration*************************
// *******************************************************************
//
// Create an iterative solver manager.
RCP< Belos::SolverManager<ST,MV,OP> > newSolver
= rcp( new Belos::MinresSolMgr<ST,MV,OP>(rcp(&problem,false), rcp(&belosList,false)));
//
// **********Print out information about problem*******************
//
if (proc_verbose) {
cout << endl << endl;
cout << "Dimension of matrix: " << NumGlobalElements << endl;
cout << "Number of right-hand sides: " << numrhs << endl;
cout << "Relative residual tolerance: " << tol << endl;
cout << endl;
}
//
// Perform solve
//
Belos::ReturnType ret = Belos::Unconverged;
try {
ret = newSolver->solve();
} catch (std::exception& e) {
if (MyPID == 0)
cout << "MINRES solver threw an exception: " << e.what() << endl;
throw e;
}
//
// Compute actual residuals.
//
bool badRes = false;
std::vector<double> actual_resids( numrhs );
std::vector<double> rhs_norm( numrhs );
Epetra_MultiVector resid(A->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 > tol) badRes = true;
}
}
if (ret!=Belos::Converged || badRes) {
if (proc_verbose)
std::cout << std::endl << "End Result: TEST FAILED" << std::endl;
}
//
// Default return value
//
if (proc_verbose)
std::cout << std::endl << "End Result: TEST PASSED" << std::endl;
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
if (ret!=Belos::Converged || badRes)
return -1;
else
return 0;
}
int main(int argc, char *argv[])
{
int i;
bool ierr, gerr;
gerr = true;
#ifdef HAVE_MPI
// Initialize MPI and setup an Epetra communicator
MPI_Init(&argc,&argv);
Teuchos::RCP<Epetra_MpiComm> Comm = Teuchos::rcp( new Epetra_MpiComm(MPI_COMM_WORLD) );
#else
// If we aren't using MPI, then setup a serial communicator.
Teuchos::RCP<Epetra_SerialComm> Comm = Teuchos::rcp( new Epetra_SerialComm() );
#endif
// number of global elements
int dim = 100;
int blockSize = 5;
bool verbose = false;
if (argc>1) {
if (argv[1][0]=='-' && argv[1][1]=='v') {
verbose = true;
}
}
// Construct a Map that puts approximately the same number of
// equations on each processor.
Teuchos::RCP<Epetra_Map> Map = Teuchos::rcp( new Epetra_Map(dim, 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 std::vector NumNz that is used to build the Petra Matrix.
// NumNz[i] is the Number of OFF-DIAGONAL term for the ith global equation
// on this processor
std::vector<int> NumNz(NumMyElements);
// We are building a tridiagonal matrix where each row has (-1 2 -1)
// So we need 2 off-diagonal terms (except for the first and last equation)
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0 || MyGlobalElements[i] == dim-1) {
NumNz[i] = 2;
}
else {
NumNz[i] = 3;
}
}
// Create an Epetra_Matrix
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(Copy, *Map, &NumNz[0]) );
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
std::vector<double> Values(2);
Values[0] = -1.0; Values[1] = -1.0;
std::vector<int> Indices(2);
double two = 2.0;
int NumEntries;
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
NumEntries = 1;
}
else if (MyGlobalElements[i] == dim-1) {
Indices[0] = dim-2;
NumEntries = 1;
}
else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
NumEntries = 2;
}
ierr = A->InsertGlobalValues(MyGlobalElements[i],NumEntries,&Values[0],&Indices[0]);
assert(ierr==0);
// Put in the diagonal entry
ierr = A->InsertGlobalValues(MyGlobalElements[i],1,&two,&MyGlobalElements[i]);
assert(ierr==0);
}
// Finish building the epetra matrix A
ierr = A->FillComplete();
assert(ierr==0);
// Issue several useful typedefs;
typedef Belos::MultiVec<double> EMV;
typedef Belos::Operator<double> EOP;
// Create an Epetra_MultiVector for an initial std::vector to start the solver.
// Note that this needs to have the same number of columns as the blocksize.
Teuchos::RCP<Belos::EpetraMultiVec> ivec = Teuchos::rcp( new Belos::EpetraMultiVec(*Map, blockSize) );
ivec->Random();
// Create an output manager to handle the I/O from the solver
Teuchos::RCP<Belos::OutputManager<double> > MyOM = Teuchos::rcp( new Belos::OutputManager<double>() );
if (verbose) {
MyOM->setVerbosity( Belos::Warnings );
}
// test the Epetra adapter multivector
ierr = Belos::TestMultiVecTraits<double,EMV>(MyOM,ivec);
gerr &= ierr;
if (ierr) {
MyOM->print(Belos::Warnings,"*** EpetraAdapter PASSED TestMultiVecTraits()\n");
}
else {
MyOM->print(Belos::Warnings,"*** EpetraAdapter FAILED TestMultiVecTraits() ***\n\n");
}
#ifdef HAVE_MPI
MPI_Finalize();
#endif
if (gerr == false) {
MyOM->print(Belos::Warnings,"End Result: TEST FAILED\n");
return -1;
}
//
// Default return value
//
MyOM->print(Belos::Warnings,"End Result: TEST PASSED\n");
return 0;
}
// ======================================================================
void GetPtent(const Epetra_RowMatrix& A, Teuchos::ParameterList& List,
double* thisns, Teuchos::RCP<Epetra_CrsMatrix>& Ptent,
Teuchos::RCP<Epetra_MultiVector>& NextNS, const int domainoffset)
{
const int nsdim = List.get<int>("null space: dimension",-1);
if (nsdim<=0) ML_THROW("null space dimension not given",-1);
const Epetra_Map& rowmap = A.RowMatrixRowMap();
const int mylength = rowmap.NumMyElements();
// wrap nullspace into something more handy
Epetra_MultiVector ns(View,rowmap,thisns,mylength,nsdim);
// vector to hold aggregation information
Epetra_IntVector aggs(rowmap,false);
// aggregation with global aggregate numbering
int naggregates = GetGlobalAggregates(const_cast<Epetra_RowMatrix&>(A),List,thisns,aggs);
// build a domain map for Ptent
// find first aggregate on proc
int firstagg = -1;
int offset = -1;
for (int i=0; i<mylength; ++i)
if (aggs[i]>=0)
{
offset = firstagg = aggs[i];
break;
}
offset *= nsdim;
if (offset<0) ML_THROW("could not find any aggregate on proc",-2);
std::vector<int> coarsegids(naggregates*nsdim);
for (int i=0; i<naggregates; ++i)
for (int j=0; j<nsdim; ++j)
{
coarsegids[i*nsdim+j] = offset + domainoffset;
++offset;
}
Epetra_Map pdomainmap(-1,naggregates*nsdim,&coarsegids[0],0,A.Comm());
// loop aggregates and build ids for dofs
std::map<int,std::vector<int> > aggdofs;
std::map<int,std::vector<int> >::iterator fool;
for (int i=0; i<naggregates; ++i)
{
std::vector<int> gids(0);
aggdofs.insert(std::pair<int,std::vector<int> >(firstagg+i,gids));
}
for (int i=0; i<mylength; ++i)
{
if (aggs[i]<0) continue;
std::vector<int>& gids = aggdofs[aggs[i]];
gids.push_back(aggs.Map().GID(i));
}
#if 0 // debugging output
for (int proc=0; proc<A.Comm().NumProc(); ++proc)
{
if (proc==A.Comm().MyPID())
{
for (fool=aggdofs.begin(); fool!=aggdofs.end(); ++fool)
{
std::cout << "Proc " << proc << " Aggregate " << fool->first << " Dofs ";
std::vector<int>& gids = fool->second;
for (int i=0; i<(int)gids.size(); ++i) std::cout << gids[i] << " ";
std::cout << std::endl;
}
}
fflush(stdout);
A.Comm().Barrier();
}
#endif
// coarse level nullspace to be filled
NextNS = Teuchos::rcp(new Epetra_MultiVector(pdomainmap,nsdim,true));
Epetra_MultiVector& nextns = *NextNS;
// matrix Ptent
Ptent = Teuchos::rcp(new Epetra_CrsMatrix(Copy,rowmap,nsdim));
// loop aggregates and extract the appropriate slices of the nullspace.
// do QR and assemble Q and R to Ptent and NextNS
for (fool=aggdofs.begin(); fool!=aggdofs.end(); ++fool)
{
// extract aggregate-local junk of nullspace
const int aggsize = (int)fool->second.size();
Epetra_SerialDenseMatrix Bagg(aggsize,nsdim);
for (int i=0; i<aggsize; ++i)
for (int j=0; j<nsdim; ++j)
Bagg(i,j) = (*ns(j))[ns.Map().LID(fool->second[i])];
// Bagg = Q*R
int m = Bagg.M();
int n = Bagg.N();
int lwork = n*10;
int info = 0;
int k = std::min(m,n);
if (k!=n) ML_THROW("Aggregate too small, fatal!",-1);
std::vector<double> work(lwork);
std::vector<double> tau(k);
Epetra_LAPACK lapack;
lapack.GEQRF(m,n,Bagg.A(),m,&tau[0],&work[0],lwork,&info);
if (info) ML_THROW("Lapack dgeqrf returned nonzero",info);
if (work[0]>lwork)
{
lwork = (int)work[0];
work.resize(lwork);
}
// get R (stored on Bagg) and assemble it into nextns
int agg_cgid = fool->first*nsdim;
if (!nextns.Map().MyGID(agg_cgid+domainoffset))
ML_THROW("Missing coarse column id on this proc",-1);
for (int i=0; i<n; ++i)
for (int j=i; j<n; ++j)
(*nextns(j))[nextns.Map().LID(domainoffset+agg_cgid+i)] = Bagg(i,j);
// get Q and assemble it into Ptent
lapack.ORGQR(m,n,k,Bagg.A(),m,&tau[0],&work[0],lwork,&info);
if (info) ML_THROW("Lapack dorgqr returned nonzero",info);
for (int i=0; i<aggsize; ++i)
{
const int actgrow = fool->second[i];
for (int j=0; j<nsdim; ++j)
{
int actgcol = fool->first*nsdim+j+domainoffset;
int errone = Ptent->SumIntoGlobalValues(actgrow,1,&Bagg(i,j),&actgcol);
if (errone>0)
{
int errtwo = Ptent->InsertGlobalValues(actgrow,1,&Bagg(i,j),&actgcol);
if (errtwo<0) ML_THROW("Epetra_CrsMatrix::InsertGlobalValues returned negative nonzero",errtwo);
}
else if (errone) ML_THROW("Epetra_CrsMatrix::SumIntoGlobalValues returned negative nonzero",errone);
}
} // for (int i=0; i<aggsize; ++i)
} // for (fool=aggdofs.begin(); fool!=aggdofs.end(); ++fool)
int err = Ptent->FillComplete(pdomainmap,rowmap);
if (err) ML_THROW("Epetra_CrsMatrix::FillComplete returned nonzero",err);
err = Ptent->OptimizeStorage();
if (err) ML_THROW("Epetra_CrsMatrix::OptimizeStorage returned nonzero",err);
return;
}
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 success = false;
bool verbose = false;
try {
bool debug = false, proc_verbose = false;
int frequency = -1; // frequency of status test output.
int numrhs = 1; // number of right-hand sides to solve for
int maxiters = -1; // maximum number of iterations allowed per linear system
MT tol = 1.0e-10; // relative residual tolerance
Teuchos::CommandLineProcessor cmdp(false,true);
cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
cmdp.setOption("debug","nodebug",&debug,"Print debugging information from the solver.");
cmdp.setOption("frequency",&frequency,"Solvers frequency for printing residuals (#iters).");
cmdp.setOption("tol",&tol,"Relative residual tolerance used by GMRES solver.");
cmdp.setOption("num-rhs",&numrhs,"Number of right-hand sides to be solved for.");
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
// **********************************************************************
// ******************Set up the problem to be solved*********************
// construct diagonal matrix
const int NumGlobalElements = 100;
const int m = 4; // number of negative eigenvalues
// Create diagonal matrix with n-m positive and m negative eigenvalues.
Epetra_Map epetraMap( NumGlobalElements, 0, Comm );
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix( Copy, epetraMap, 1 ) );
for ( int k=0; k<epetraMap.NumMyElements(); k++ )
{
int GIDk = epetraMap.GID(k);
double val = 2*(GIDk-m) + 1;
TEUCHOS_ASSERT_EQUALITY( 0, A->InsertGlobalValues( GIDk, 1, &val, &GIDk ) );
}
TEUCHOS_ASSERT_EQUALITY( 0, A->FillComplete() );
TEUCHOS_ASSERT_EQUALITY( 0, A->OptimizeStorage() );
// create initial guess and right-hand side
Teuchos::RCP<Epetra_MultiVector> vecX = Teuchos::rcp( new Epetra_MultiVector( epetraMap, numrhs ) );
Teuchos::RCP<Epetra_MultiVector> vecB = Teuchos::rcp( new Epetra_MultiVector( epetraMap, numrhs ) );
// **********************************************************************
proc_verbose = verbose && (MyPID==0); /* Only print on the zero processor */
Teuchos::RCP<Epetra_MultiVector> X;
Teuchos::RCP<Epetra_MultiVector> B;
// Check to see if the number of right-hand sides is the same as requested.
if (numrhs>1) {
X = rcp( new Epetra_MultiVector( epetraMap, numrhs ) );
B = rcp( new Epetra_MultiVector( epetraMap, numrhs ) );
X->Random();
OPT::Apply( *A, *X, *B );
X->PutScalar( 0.0 );
}
else {
X = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecX);
B = Teuchos::rcp_implicit_cast<Epetra_MultiVector>(vecB);
B->PutScalar( 1.0 );
}
//
// ********Other information used by block solver***********
// *****************(can be user specified)******************
//
if (maxiters == -1)
maxiters = NumGlobalElements - 1; // maximum number of iterations to run
//
ParameterList belosList;
belosList.set( "Maximum Iterations", maxiters ); // Maximum number of iterations allowed
belosList.set( "Convergence Tolerance", tol ); // Relative convergence tolerance requested
belosList.set( "Assert Positive Definiteness", false ); // Explicitly don't enforce positive definiteness
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;
return -1;
}
//
// *******************************************************************
// ****************Start the CG iteration*************************
// *******************************************************************
//
// Create an iterative solver manager.
RCP<Belos::SolverManager<double,MV,OP> > newSolver
= rcp( new Belos::PseudoBlockCGSolMgr<double,MV,OP>(rcp(&problem,false), rcp(&belosList,false)));
//
// **********Print out information about problem*******************
//
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 << "Relative residual tolerance: " << tol << std::endl;
std::cout << std::endl;
}
//
// Perform solve
//
Belos::ReturnType ret = newSolver->solve();
//
// Get the number of iterations for this solve.
//
int numIters = newSolver->getNumIters();
if (proc_verbose)
std::cout << "Number of iterations performed for this solve: " << numIters << std::endl;
//
// Compute actual residuals.
//
bool badRes = false;
std::vector<double> actual_resids( numrhs );
std::vector<double> rhs_norm( numrhs );
Epetra_MultiVector resid(epetraMap, 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 > tol) badRes = true;
}
}
success = ret==Belos::Converged && !badRes;
if (success) {
if (proc_verbose)
std::cout << "End Result: TEST PASSED" << std::endl;
} else {
if (proc_verbose)
std::cout << "End Result: TEST FAILED" << std::endl;
}
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return ( success ? EXIT_SUCCESS : EXIT_FAILURE );
}
int main(int argc, char *argv[])
{
int i;
bool ierr, gerr;
gerr = true;
#ifdef HAVE_MPI
// Initialize MPI and setup an Epetra communicator
MPI_Init(&argc,&argv);
Teuchos::RCP<Epetra_MpiComm> Comm = Teuchos::rcp( new Epetra_MpiComm(MPI_COMM_WORLD) );
#else
// If we aren't using MPI, then setup a serial communicator.
Teuchos::RCP<Epetra_SerialComm> Comm = Teuchos::rcp( new Epetra_SerialComm() );
#endif
// number of global elements
const int dim = 100;
const int blockSize = 5;
bool verbose = false;
if (argc>1) {
if (argv[1][0]=='-' && argv[1][1]=='v') {
verbose = true;
}
}
// Create an output manager to handle the I/O from the solver
Teuchos::RCP<Anasazi::OutputManager<double> > MyOM = Teuchos::rcp( new Anasazi::BasicOutputManager<double>() );
if (verbose) {
MyOM->setVerbosity( Anasazi::Warnings );
}
#ifndef HAVE_EPETRA_THYRA
MyOM->stream(Anasazi::Warnings)
<< "Please configure Anasazi with:" << std::endl
<< "--enable-epetra-thyra" << std::endl
<< "--enable-anasazi-thyra" << std::endl;
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return -1;
#endif
// Construct a Map that puts approximately the same number of
// equations on each processor.
Teuchos::RCP<Epetra_Map> Map = Teuchos::rcp( new Epetra_Map(dim, 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 term for the ith global equation
// on this processor
std::vector<int> NumNz(NumMyElements);
// We are building a tridiagonal matrix where each row has (-1 2 -1)
// So we need 2 off-diagonal terms (except for the first and last equation)
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0 || MyGlobalElements[i] == dim-1) {
NumNz[i] = 2;
}
else {
NumNz[i] = 3;
}
}
// Create an Epetra_Matrix
Teuchos::RCP<Epetra_CrsMatrix> A = Teuchos::rcp( new Epetra_CrsMatrix(Copy, *Map, &NumNz[0]) );
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
std::vector<double> Values(2);
Values[0] = -1.0; Values[1] = -1.0;
std::vector<int> Indices(2);
double two = 2.0;
int NumEntries;
for (i=0; i<NumMyElements; i++) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
NumEntries = 1;
}
else if (MyGlobalElements[i] == dim-1) {
Indices[0] = dim-2;
NumEntries = 1;
}
else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
NumEntries = 2;
}
ierr = A->InsertGlobalValues(MyGlobalElements[i],NumEntries,&Values[0],&Indices[0]);
assert(ierr==0);
// Put in the diagonal entry
ierr = A->InsertGlobalValues(MyGlobalElements[i],1,&two,&MyGlobalElements[i]);
assert(ierr==0);
}
// Finish building the epetra matrix A
ierr = A->FillComplete();
assert(ierr==0);
#ifdef HAVE_EPETRA_THYRA
typedef Thyra::MultiVectorBase<double> TMVB;
typedef Thyra::LinearOpBase<double> TLOB;
// first, create a Thyra::VectorSpaceBase from an Epetra_Map using the Epetra-Thyra wrappers
Teuchos::RCP<const Thyra::VectorSpaceBase<double> > space = Thyra::create_VectorSpace(Map);
// then, create a Thyra::MultiVectorBase from the Thyra::VectorSpaceBase using Thyra creational functions
Teuchos::RCP<Thyra::MultiVectorBase<double> > thyra_ivec = Thyra::createMembers(space,blockSize);
// then, create a Thyra::LinearOpBase from the Epetra_CrsMatrix using the Epetra-Thyra wrappers
Teuchos::RCP<const Thyra::LinearOpBase<double> > thyra_op = Thyra::epetraLinearOp(A);
// test the Thyra multivector adapter
ierr = Anasazi::TestMultiVecTraits<double,TMVB>(MyOM,thyra_ivec);
gerr |= ierr;
if (ierr) {
MyOM->stream(Anasazi::Warnings) << "*** ThyraAdapter PASSED TestMultiVecTraits()" << std::endl;
}
else {
MyOM->stream(Anasazi::Warnings) << "*** ThyraAdapter FAILED TestMultiVecTraits() ***" << std::endl << std::endl;
}
// test the Thyra operator adapter
ierr = Anasazi::TestOperatorTraits<double,TMVB,TLOB>(MyOM,thyra_ivec,thyra_op);
gerr |= ierr;
if (ierr) {
MyOM->stream(Anasazi::Warnings) << "*** ThyraAdapter PASSED TestOperatorTraits()" << std::endl;
}
else {
MyOM->stream(Anasazi::Warnings) << "*** ThyraAdapter FAILED TestOperatorTraits() ***" << std::endl << std::endl;
}
#endif
#ifdef HAVE_MPI
MPI_Finalize();
#endif
if (gerr == false) {
MyOM->print(Anasazi::Warnings,"End Result: TEST FAILED\n");
return -1;
}
//
// Default return value
//
MyOM->print(Anasazi::Warnings,"End Result: TEST PASSED\n");
return 0;
}