Teuchos::RCP<Epetra_LinearProblem> build_problem_mm(Teuchos::ParameterList& test_params, Epetra_CrsMatrix* A, Epetra_MultiVector* b)
{
const Epetra_Map& rowmap = A->RowMap();
Epetra_MultiVector* x = new Epetra_MultiVector(rowmap, 1);
if (b == NULL) {
std::cout << "creating b = A*random" << std::endl;
b = new Epetra_MultiVector(rowmap, 1);
x->Random();
A->Apply(*x, *b);
}
x->PutScalar(0);
Teuchos::RCP<Epetra_LinearProblem> problem = Teuchos::rcp(new Epetra_LinearProblem(A,x,b));
return problem;
}
//
// Amesos_TestMultiSolver.cpp reads in a matrix in Harwell-Boeing format,
// calls one of the sparse direct solvers, using blocked right hand sides
// and computes the error and residual.
//
// TestSolver ignores the Harwell-Boeing right hand sides, creating
// random right hand sides instead.
//
// Amesos_TestMultiSolver can test either A x = b or A^T x = b.
// This can be a bit confusing because sparse direct solvers
// use compressed column storage - the transpose of Trilinos'
// sparse row storage.
//
// Matrices:
// readA - Serial. As read from the file.
// transposeA - Serial. The transpose of readA.
// serialA - if (transpose) then transposeA else readA
// distributedA - readA distributed to all processes
// passA - if ( distributed ) then distributedA else serialA
//
//
int Amesos_TestMultiSolver( Epetra_Comm &Comm, char *matrix_file, int numsolves,
SparseSolverType SparseSolver, bool transpose,
int special, AMESOS_MatrixType matrix_type ) {
int iam = Comm.MyPID() ;
// int hatever;
// if ( iam == 0 ) std::cin >> hatever ;
Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
std::string FileName = matrix_file ;
int FN_Size = FileName.size() ;
std::string LastFiveBytes = FileName.substr( EPETRA_MAX(0,FN_Size-5), FN_Size );
std::string LastFourBytes = FileName.substr( EPETRA_MAX(0,FN_Size-4), FN_Size );
bool NonContiguousMap = false;
if ( LastFiveBytes == ".triU" ) {
NonContiguousMap = true;
// Call routine to read in unsymmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, false, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFiveBytes == ".triS" ) {
NonContiguousMap = true;
// Call routine to read in symmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, true, Comm,
readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFourBytes == ".mtx" ) {
EPETRA_CHK_ERR( Trilinos_Util_ReadMatrixMarket2Epetra( matrix_file, Comm, readMap,
readA, readx, readb, readxexact) );
} else {
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra( matrix_file, Comm, readMap, readA, readx,
readb, readxexact) ;
}
}
}
Epetra_CrsMatrix transposeA(Copy, *readMap, 0);
Epetra_CrsMatrix *serialA ;
if ( transpose ) {
assert( CrsMatrixTranspose( readA, &transposeA ) == 0 );
serialA = &transposeA ;
} else {
serialA = readA ;
}
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
Epetra_Map* map_;
if( NonContiguousMap ) {
//
// map gives us NumMyElements and MyFirstElement;
//
int NumGlobalElements = readMap->NumGlobalElements();
int NumMyElements = map.NumMyElements();
int MyFirstElement = map.MinMyGID();
std::vector<int> MapMap_( NumGlobalElements );
readMap->MyGlobalElements( &MapMap_[0] ) ;
Comm.Broadcast( &MapMap_[0], NumGlobalElements, 0 ) ;
map_ = new Epetra_Map( NumGlobalElements, NumMyElements, &MapMap_[MyFirstElement], 0, Comm);
} else {
map_ = new Epetra_Map( map ) ;
}
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, *map_);
Epetra_CrsMatrix A(Copy, *map_, 0);
Epetra_RowMatrix * passA = 0;
Epetra_MultiVector * passx = 0;
Epetra_MultiVector * passb = 0;
Epetra_MultiVector * passxexact = 0;
Epetra_MultiVector * passresid = 0;
Epetra_MultiVector * passtmp = 0;
Epetra_MultiVector x(*map_,numsolves);
Epetra_MultiVector b(*map_,numsolves);
Epetra_MultiVector xexact(*map_,numsolves);
Epetra_MultiVector resid(*map_,numsolves);
Epetra_MultiVector tmp(*map_,numsolves);
Epetra_MultiVector serialx(*readMap,numsolves);
Epetra_MultiVector serialb(*readMap,numsolves);
Epetra_MultiVector serialxexact(*readMap,numsolves);
Epetra_MultiVector serialresid(*readMap,numsolves);
Epetra_MultiVector serialtmp(*readMap,numsolves);
bool distribute_matrix = ( matrix_type == AMESOS_Distributed ) ;
if ( distribute_matrix ) {
//
// Initialize x, b and xexact to the values read in from the file
//
A.Export(*serialA, exporter, Add);
Comm.Barrier();
assert(A.FillComplete()==0);
Comm.Barrier();
passA = &A;
passx = &x;
passb = &b;
passxexact = &xexact;
passresid = &resid;
passtmp = &tmp;
} else {
passA = serialA;
passx = &serialx;
passb = &serialb;
passxexact = &serialxexact;
passresid = &serialresid;
passtmp = &serialtmp;
}
passxexact->SetSeed(131) ;
passxexact->Random();
passx->SetSeed(11231) ;
passx->Random();
passb->PutScalar( 0.0 );
passA->Multiply( transpose, *passxexact, *passb ) ;
Epetra_MultiVector CopyB( *passb ) ;
double Anorm = passA->NormInf() ;
SparseDirectTimingVars::SS_Result.Set_Anorm(Anorm) ;
Epetra_LinearProblem Problem( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb );
double max_resid = 0.0;
for ( int j = 0 ; j < special+1 ; j++ ) {
Epetra_Time TotalTime( Comm ) ;
if ( false ) {
#ifdef TEST_UMFPACK
unused code
} else if ( SparseSolver == UMFPACK ) {
UmfpackOO umfpack( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
umfpack.SetTrans( transpose ) ;
umfpack.Solve() ;
#endif
#ifdef TEST_SUPERLU
} else if ( SparseSolver == SuperLU ) {
SuperluserialOO superluserial( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
superluserial.SetPermc( SuperLU_permc ) ;
superluserial.SetTrans( transpose ) ;
superluserial.SetUseDGSSV( special == 0 ) ;
superluserial.Solve() ;
#endif
#ifdef HAVE_AMESOS_SLUD
} else if ( SparseSolver == SuperLUdist ) {
SuperludistOO superludist( Problem ) ;
superludist.SetTrans( transpose ) ;
EPETRA_CHK_ERR( superludist.Solve( true ) ) ;
#endif
#ifdef HAVE_AMESOS_SLUD2
} else if ( SparseSolver == SuperLUdist2 ) {
Superludist2_OO superludist2( Problem ) ;
superludist2.SetTrans( transpose ) ;
EPETRA_CHK_ERR( superludist2.Solve( true ) ) ;
#endif
#ifdef TEST_SPOOLES
} else if ( SparseSolver == SPOOLES ) {
SpoolesOO spooles( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spooles.SetTrans( transpose ) ;
spooles.Solve() ;
#endif
#ifdef HAVE_AMESOS_DSCPACK
} else if ( SparseSolver == DSCPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Dscpack dscpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( dscpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( dscpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_UMFPACK
} else if ( SparseSolver == UMFPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Umfpack umfpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( umfpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( umfpack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( umfpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_KLU
} else if ( SparseSolver == KLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Klu klu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( klu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( klu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( klu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( klu.NumericFactorization( ) );
EPETRA_CHK_ERR( klu.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARAKLETE
} else if ( SparseSolver == PARAKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Paraklete paraklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( paraklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( paraklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( paraklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( paraklete.NumericFactorization( ) );
EPETRA_CHK_ERR( paraklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SLUS
} else if ( SparseSolver == SuperLU ) {
Epetra_SLU superluserial( &Problem ) ;
EPETRA_CHK_ERR( superluserial.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superluserial.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superluserial.NumericFactorization( ) );
EPETRA_CHK_ERR( superluserial.Solve( ) );
#endif
#ifdef HAVE_AMESOS_LAPACK
} else if ( SparseSolver == LAPACK ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Lapack lapack( Problem ) ;
EPETRA_CHK_ERR( lapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( lapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( lapack.NumericFactorization( ) );
EPETRA_CHK_ERR( lapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_TAUCS
} else if ( SparseSolver == TAUCS ) {
Teuchos::ParameterList ParamList ;
Amesos_Taucs taucs( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( taucs.SetParameters( ParamList ) );
EPETRA_CHK_ERR( taucs.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( taucs.SymbolicFactorization( ) );
EPETRA_CHK_ERR( taucs.NumericFactorization( ) );
EPETRA_CHK_ERR( taucs.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARDISO
} else if ( SparseSolver == PARDISO ) {
Teuchos::ParameterList ParamList ;
Amesos_Pardiso pardiso( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( pardiso.SetParameters( ParamList ) );
EPETRA_CHK_ERR( pardiso.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( pardiso.SymbolicFactorization( ) );
EPETRA_CHK_ERR( pardiso.NumericFactorization( ) );
EPETRA_CHK_ERR( pardiso.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARKLETE
} else if ( SparseSolver == PARKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Parklete parklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( parklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( parklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( parklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( parklete.NumericFactorization( ) );
EPETRA_CHK_ERR( parklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_MUMPS
} else if ( SparseSolver == MUMPS ) {
Teuchos::ParameterList ParamList ;
Amesos_Mumps mumps( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( mumps.SetParameters( ParamList ) );
EPETRA_CHK_ERR( mumps.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( mumps.SymbolicFactorization( ) );
EPETRA_CHK_ERR( mumps.NumericFactorization( ) );
EPETRA_CHK_ERR( mumps.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SCALAPACK
} else if ( SparseSolver == SCALAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Scalapack scalapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( scalapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( scalapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( scalapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( scalapack.NumericFactorization( ) );
EPETRA_CHK_ERR( scalapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLUDIST
} else if ( SparseSolver == SUPERLUDIST ) {
Teuchos::ParameterList ParamList ;
Amesos_Superludist superludist( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( superludist.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superludist.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superludist.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superludist.NumericFactorization( ) );
EPETRA_CHK_ERR( superludist.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLU
} else if ( SparseSolver == SUPERLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Superlu superlu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( superlu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superlu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superlu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superlu.NumericFactorization( ) );
EPETRA_CHK_ERR( superlu.Solve( ) );
#endif
#ifdef TEST_SPOOLESSERIAL
} else if ( SparseSolver == SPOOLESSERIAL ) {
SpoolesserialOO spoolesserial( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spoolesserial.Solve() ;
#endif
} else {
SparseDirectTimingVars::log_file << "Solver not implemented yet" << std::endl ;
std::cerr << "\n\n#################### Requested solver not available (Or not tested with blocked RHS) on this platform #####################\n" << std::endl ;
}
SparseDirectTimingVars::SS_Result.Set_Total_Time( TotalTime.ElapsedTime() );
// SparseDirectTimingVars::SS_Result.Set_First_Time( 0.0 );
// SparseDirectTimingVars::SS_Result.Set_Middle_Time( 0.0 );
// SparseDirectTimingVars::SS_Result.Set_Last_Time( 0.0 );
//
// Compute the error = norm(xcomp - xexact )
//
std::vector <double> error(numsolves) ;
double max_error = 0.0;
passresid->Update(1.0, *passx, -1.0, *passxexact, 0.0);
passresid->Norm2(&error[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( error[i] > max_error ) max_error = error[i] ;
SparseDirectTimingVars::SS_Result.Set_Error(max_error) ;
// passxexact->Norm2(&error[0] ) ;
// passx->Norm2(&error ) ;
//
// Compute the residual = norm(Ax - b)
//
std::vector <double> residual(numsolves) ;
passtmp->PutScalar(0.0);
passA->Multiply( transpose, *passx, *passtmp);
passresid->Update(1.0, *passtmp, -1.0, *passb, 0.0);
// passresid->Update(1.0, *passtmp, -1.0, CopyB, 0.0);
passresid->Norm2(&residual[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( residual[i] > max_resid ) max_resid = residual[i] ;
SparseDirectTimingVars::SS_Result.Set_Residual(max_resid) ;
std::vector <double> bnorm(numsolves);
passb->Norm2( &bnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Bnorm(bnorm[0]) ;
std::vector <double> xnorm(numsolves);
passx->Norm2( &xnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Xnorm(xnorm[0]) ;
if ( false && iam == 0 ) {
std::cout << " Amesos_TestMutliSolver.cpp " << std::endl ;
for ( int i = 0 ; i< numsolves && i < 10 ; i++ ) {
std::cout << "i=" << i
<< " error = " << error[i]
<< " xnorm = " << xnorm[i]
<< " residual = " << residual[i]
<< " bnorm = " << bnorm[i]
<< std::endl ;
}
std::cout << std::endl << " max_resid = " << max_resid ;
std::cout << " max_error = " << max_error << std::endl ;
std::cout << " Get_residual() again = " << SparseDirectTimingVars::SS_Result.Get_Residual() << std::endl ;
}
}
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
delete map_;
Comm.Barrier();
return 0 ;
}
int main(int argc, char *argv[])
{
int ierr = 0;
double elapsed_time;
double total_flops;
double MFLOPs;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm comm;
#endif
bool verbose = false;
bool summary = false;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='v') verbose = true;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='s') summary = true;
if(argc < 6) {
cerr << "Usage: " << argv[0]
<< " NumNodesX NumNodesY NumProcX NumProcY NumPoints [-v|-s]" << endl
<< "where:" << endl
<< "NumNodesX - Number of mesh nodes in X direction per processor" << endl
<< "NumNodesY - Number of mesh nodes in Y direction per processor" << endl
<< "NumProcX - Number of processors to use in X direction" << endl
<< "NumProcY - Number of processors to use in Y direction" << endl
<< "NumPoints - Number of points to use in stencil (5, 9 or 25 only)" << endl
<< "-v|-s - (Optional) Run in verbose mode if -v present or summary mode if -s present" << endl
<< " NOTES: NumProcX*NumProcY must equal the number of processors used to run the problem." << endl << endl
<< " Serial example:" << endl
<< argv[0] << " 16 12 1 1 25 -v" << endl
<< " Run this program in verbose mode on 1 processor using a 16 X 12 grid with a 25 point stencil."<< endl <<endl
<< " MPI example:" << endl
<< "mpirun -np 32 " << argv[0] << " 10 12 4 8 9 -v" << endl
<< " Run this program in verbose mode on 32 processors putting a 10 X 12 subgrid on each processor using 4 processors "<< endl
<< " in the X direction and 8 in the Y direction. Total grid size is 40 points in X and 96 in Y with a 9 point stencil."<< endl
<< endl;
return(1);
}
//char tmp;
//if (comm.MyPID()==0) cout << "Press any key to continue..."<< endl;
//if (comm.MyPID()==0) cin >> tmp;
//comm.Barrier();
comm.SetTracebackMode(0); // This should shut down any error traceback reporting
if (verbose && comm.MyPID()==0)
cout << Epetra_Version() << endl << endl;
if (summary && comm.MyPID()==0) {
if (comm.NumProc()==1)
cout << Epetra_Version() << endl << endl;
else
cout << endl << endl; // Print two blank line to keep output columns lined up
}
if (verbose) cout << comm <<endl;
// Redefine verbose to only print on PE 0
if (verbose && comm.MyPID()!=0) verbose = false;
if (summary && comm.MyPID()!=0) summary = false;
int numNodesX = atoi(argv[1]);
int numNodesY = atoi(argv[2]);
int numProcsX = atoi(argv[3]);
int numProcsY = atoi(argv[4]);
int numPoints = atoi(argv[5]);
if (verbose || (summary && comm.NumProc()==1)) {
cout << " Number of local nodes in X direction = " << numNodesX << endl
<< " Number of local nodes in Y direction = " << numNodesY << endl
<< " Number of global nodes in X direction = " << numNodesX*numProcsX << endl
<< " Number of global nodes in Y direction = " << numNodesY*numProcsY << endl
<< " Number of local nonzero entries = " << numNodesX*numNodesY*numPoints << endl
<< " Number of global nonzero entries = " << numNodesX*numNodesY*numPoints*numProcsX*numProcsY << endl
<< " Number of Processors in X direction = " << numProcsX << endl
<< " Number of Processors in Y direction = " << numProcsY << endl
<< " Number of Points in stencil = " << numPoints << endl << endl;
}
// Print blank line to keep output columns lined up
if (summary && comm.NumProc()>1)
cout << endl << endl << endl << endl << endl << endl << endl << endl<< endl << endl;
if (numProcsX*numProcsY!=comm.NumProc()) {
cerr << "Number of processors = " << comm.NumProc() << endl
<< " is not the product of " << numProcsX << " and " << numProcsY << endl << endl;
return(1);
}
if (numPoints!=5 && numPoints!=9 && numPoints!=25) {
cerr << "Number of points specified = " << numPoints << endl
<< " is not 5, 9, 25" << endl << endl;
return(1);
}
if (numNodesX*numNodesY<=0) {
cerr << "Product of number of nodes is <= zero" << endl << endl;
return(1);
}
Epetra_IntSerialDenseVector Xoff, XLoff, XUoff;
Epetra_IntSerialDenseVector Yoff, YLoff, YUoff;
if (numPoints==5) {
// Generate a 5-point 2D Finite Difference matrix
Xoff.Size(5);
Yoff.Size(5);
Xoff[0] = -1; Xoff[1] = 1; Xoff[2] = 0; Xoff[3] = 0; Xoff[4] = 0;
Yoff[0] = 0; Yoff[1] = 0; Yoff[2] = 0; Yoff[3] = -1; Yoff[4] = 1;
// Generate a 2-point 2D Lower triangular Finite Difference matrix
XLoff.Size(2);
YLoff.Size(2);
XLoff[0] = -1; XLoff[1] = 0;
YLoff[0] = 0; YLoff[1] = -1;
// Generate a 3-point 2D upper triangular Finite Difference matrix
XUoff.Size(3);
YUoff.Size(3);
XUoff[0] = 0; XUoff[1] = 1; XUoff[2] = 0;
YUoff[0] = 0; YUoff[1] = 0; YUoff[2] = 1;
}
else if (numPoints==9) {
// Generate a 9-point 2D Finite Difference matrix
Xoff.Size(9);
Yoff.Size(9);
Xoff[0] = -1; Xoff[1] = 0; Xoff[2] = 1;
Yoff[0] = -1; Yoff[1] = -1; Yoff[2] = -1;
Xoff[3] = -1; Xoff[4] = 0; Xoff[5] = 1;
Yoff[3] = 0; Yoff[4] = 0; Yoff[5] = 0;
Xoff[6] = -1; Xoff[7] = 0; Xoff[8] = 1;
Yoff[6] = 1; Yoff[7] = 1; Yoff[8] = 1;
// Generate a 5-point lower triangular 2D Finite Difference matrix
XLoff.Size(5);
YLoff.Size(5);
XLoff[0] = -1; XLoff[1] = 0; Xoff[2] = 1;
YLoff[0] = -1; YLoff[1] = -1; Yoff[2] = -1;
XLoff[3] = -1; XLoff[4] = 0;
YLoff[3] = 0; YLoff[4] = 0;
// Generate a 4-point upper triangular 2D Finite Difference matrix
XUoff.Size(4);
YUoff.Size(4);
XUoff[0] = 1;
YUoff[0] = 0;
XUoff[1] = -1; XUoff[2] = 0; XUoff[3] = 1;
YUoff[1] = 1; YUoff[2] = 1; YUoff[3] = 1;
}
else {
// Generate a 25-point 2D Finite Difference matrix
Xoff.Size(25);
Yoff.Size(25);
int xi = 0, yi = 0;
int xo = -2, yo = -2;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
// Generate a 13-point lower triangular 2D Finite Difference matrix
XLoff.Size(13);
YLoff.Size(13);
xi = 0, yi = 0;
xo = -2, yo = -2;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
// Generate a 13-point upper triangular 2D Finite Difference matrix
XUoff.Size(13);
YUoff.Size(13);
xi = 0, yi = 0;
xo = 0, yo = 0;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
}
Epetra_Map * map;
Epetra_Map * mapL;
Epetra_Map * mapU;
Epetra_CrsMatrix * A;
Epetra_CrsMatrix * L;
Epetra_CrsMatrix * U;
Epetra_MultiVector * b;
Epetra_MultiVector * bt;
Epetra_MultiVector * xexact;
Epetra_MultiVector * bL;
Epetra_MultiVector * btL;
Epetra_MultiVector * xexactL;
Epetra_MultiVector * bU;
Epetra_MultiVector * btU;
Epetra_MultiVector * xexactU;
Epetra_SerialDenseVector resvec(0);
//Timings
Epetra_Flops flopcounter;
Epetra_Time timer(comm);
#ifdef EPETRA_VERY_SHORT_PERFTEST
int jstop = 1;
#elif EPETRA_SHORT_PERFTEST
int jstop = 1;
#else
int jstop = 2;
#endif
for (int j=0; j<jstop; j++) {
for (int k=1; k<17; k++) {
#ifdef EPETRA_VERY_SHORT_PERFTEST
if (k<3 || (k%4==0 && k<9)) {
#elif EPETRA_SHORT_PERFTEST
if (k<6 || k%4==0) {
#else
if (k<7 || k%2==0) {
#endif
int nrhs=k;
if (verbose) cout << "\n*************** Results for " << nrhs << " RHS with ";
bool StaticProfile = (j!=0);
if (verbose) {
if (StaticProfile) cout << " static profile\n";
else cout << " dynamic profile\n";
}
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
Xoff.Values(), Yoff.Values(), nrhs, comm, verbose, summary,
map, A, b, bt, xexact, StaticProfile, false);
#ifdef EPETRA_HAVE_JADMATRIX
timer.ResetStartTime();
Epetra_JadMatrix JA(*A);
elapsed_time = timer.ElapsedTime();
if (verbose) cout << "Time to create Jagged diagonal matrix = " << elapsed_time << endl;
//cout << "A = " << *A << endl;
//cout << "JA = " << JA << endl;
runJadMatrixTests(&JA, b, bt, xexact, StaticProfile, verbose, summary);
#endif
runMatrixTests(A, b, bt, xexact, StaticProfile, verbose, summary);
delete A;
delete b;
delete bt;
delete xexact;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XLoff.Length(),
XLoff.Values(), YLoff.Values(), nrhs, comm, verbose, summary,
mapL, L, bL, btL, xexactL, StaticProfile, true);
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XUoff.Length(),
XUoff.Values(), YUoff.Values(), nrhs, comm, verbose, summary,
mapU, U, bU, btU, xexactU, StaticProfile, true);
runLUMatrixTests(L, bL, btL, xexactL, U, bU, btU, xexactU, StaticProfile, verbose, summary);
delete L;
delete bL;
delete btL;
delete xexactL;
delete mapL;
delete U;
delete bU;
delete btU;
delete xexactU;
delete mapU;
Epetra_MultiVector q(*map, nrhs);
Epetra_MultiVector z(q);
Epetra_MultiVector r(q);
delete map;
q.SetFlopCounter(flopcounter);
z.SetFlopCounter(q);
r.SetFlopCounter(q);
resvec.Resize(nrhs);
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 norms
for( int i = 0; i < 10; ++i )
q.Norm2( resvec.Values() );
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "\nTotal MFLOPs for 10 Norm2's= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Norm2" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Dot(z, resvec.Values());
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Dot's = " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "DotProd" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Update(1.0, z, 1.0, r, 0.0);
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Updates= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Update" << '\t';
cout << MFLOPs << endl;
}
}
}
}
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return ierr ;
}
// Constructs a 2D PDE finite difference matrix using the list of x and y offsets.
//
// nx (In) - number of grid points in x direction
// ny (In) - number of grid points in y direction
// The total number of equations will be nx*ny ordered such that the x direction changes
// most rapidly:
// First equation is at point (0,0)
// Second at (1,0)
// ...
// nx equation at (nx-1,0)
// nx+1st equation at (0,1)
// numPoints (In) - number of points in finite difference stencil
// xoff (In) - stencil offsets in x direction (of length numPoints)
// yoff (In) - stencil offsets in y direction (of length numPoints)
// A standard 5-point finite difference stencil would be described as:
// numPoints = 5
// xoff = [-1, 1, 0, 0, 0]
// yoff = [ 0, 0, 0, -1, 1]
// nrhs - Number of rhs to generate. (First interface produces vectors, so nrhs is not needed
// comm (In) - an Epetra_Comm object describing the parallel machine (numProcs and my proc ID)
// map (Out) - Epetra_Map describing distribution of matrix and vectors/multivectors
// A (Out) - Epetra_CrsMatrix constructed for nx by ny grid using prescribed stencil
// Off-diagonal values are random between 0 and 1. If diagonal is part of stencil,
// diagonal will be slightly diag dominant.
// b (Out) - Generated RHS. Values satisfy b = A*xexact
// bt (Out) - Generated RHS. Values satisfy b = A'*xexact
// xexact (Out) - Generated exact solution to Ax = b and b' = A'xexact
// Note: Caller of this function is responsible for deleting all output objects.
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_Vector *& b,
Epetra_Vector *& bt,
Epetra_Vector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_MultiVector * b1, * bt1, * xexact1;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
xoff, yoff, 1, comm, verbose, summary,
map, A, b1, bt1, xexact1, StaticProfile, MakeLocalOnly);
b = dynamic_cast<Epetra_Vector *>(b1);
bt = dynamic_cast<Epetra_Vector *>(bt1);
xexact = dynamic_cast<Epetra_Vector *>(xexact1);
return;
}
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff, int nrhs,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_MultiVector *& b,
Epetra_MultiVector *& bt,
Epetra_MultiVector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_Time timer(comm);
// Determine my global IDs
long long * myGlobalElements;
GenerateMyGlobalElements(numNodesX, numNodesY, numProcsX, numProcsY, comm.MyPID(), myGlobalElements);
int numMyEquations = numNodesX*numNodesY;
map = new Epetra_Map((long long)-1, numMyEquations, myGlobalElements, 0, comm); // Create map with 2D block partitioning.
delete [] myGlobalElements;
long long numGlobalEquations = map->NumGlobalElements64();
int profile = 0; if (StaticProfile) profile = numPoints;
#ifdef EPETRA_HAVE_STATICPROFILE
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile, StaticProfile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile, StaticProfile); // Construct matrix
#else
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile); // Construct matrix
#endif
long long * indices = new long long[numPoints];
double * values = new double[numPoints];
double dnumPoints = (double) numPoints;
int nx = numNodesX*numProcsX;
for (int i=0; i<numMyEquations; i++) {
long long rowID = map->GID64(i);
int numIndices = 0;
for (int j=0; j<numPoints; j++) {
long long colID = rowID + xoff[j] + nx*yoff[j]; // Compute column ID based on stencil offsets
if (colID>-1 && colID<numGlobalEquations) {
indices[numIndices] = colID;
double value = - ((double) rand())/ ((double) RAND_MAX);
if (colID==rowID)
values[numIndices++] = dnumPoints - value; // Make diagonal dominant
else
values[numIndices++] = value;
}
}
//cout << "Building row " << rowID << endl;
A->InsertGlobalValues(rowID, numIndices, values, indices);
}
delete [] indices;
delete [] values;
double insertTime = timer.ElapsedTime();
timer.ResetStartTime();
A->FillComplete(false);
double fillCompleteTime = timer.ElapsedTime();
if (verbose)
cout << "Time to insert matrix values = " << insertTime << endl
<< "Time to complete fill = " << fillCompleteTime << endl;
if (summary) {
if (comm.NumProc()==1) cout << "InsertTime" << '\t';
cout << insertTime << endl;
if (comm.NumProc()==1) cout << "FillCompleteTime" << '\t';
cout << fillCompleteTime << endl;
}
if (nrhs<=1) {
b = new Epetra_Vector(*map);
bt = new Epetra_Vector(*map);
xexact = new Epetra_Vector(*map);
}
else {
b = new Epetra_MultiVector(*map, nrhs);
bt = new Epetra_MultiVector(*map, nrhs);
xexact = new Epetra_MultiVector(*map, nrhs);
}
xexact->Random(); // Fill xexact with random values
A->Multiply(false, *xexact, *b);
A->Multiply(true, *xexact, *bt);
return;
}
int main(int argc, char** argv)
{
#ifdef HAVE_MPI
Teuchos::GlobalMPISession mpiSession(&argc, &argv, 0);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
bool success = true;
string pass = "******";
string fail = "End Result: TEST FAILED";
int myPID = Comm.MyPID();
if(myPID == 0)
{
cout << "Starting Epetra interface test" << endl;
}
/*----------------Load a test matrix---------------*/
string matrixFileName = "wathenSmall.mtx";
Epetra_CrsMatrix *A;
Epetra_CrsMatrix *AHat;
Epetra_MultiVector *b;
Epetra_MultiVector *bHat;
Epetra_MultiVector *x;
int n = 0;
//Get Matrix
int err = EpetraExt::MatrixMarketFileToCrsMatrix(matrixFileName.c_str(), Comm, A);
if(err!=0 && myPID ==0)
{
cout << "Error reading matrix file, info = " << err << endl;
cout << fail << endl;
exit(1);
}
n = A->NumGlobalRows();
//Make b vecotor
Epetra_Map vecMap(n,0,Comm);
b = new Epetra_MultiVector(vecMap,1,false);
b->Random();
x = new Epetra_MultiVector(vecMap,1,false);
x->Random();
cout << "Epetra matrices loaded" << endl;
/*-----------------have_interface-----------------*/
/*---The have_interface checks is all the parameter list makes sense---*/
Teuchos::RCP <Teuchos::ParameterList> pLUList;
string pListFileName = "ShyLU_epetra_interface.xml";
pLUList = Teuchos::getParametersFromXmlFile(pListFileName);
/*----------------partitioning_interface--------------*/
/*-----------Will use check the epetra matrix on partition_interface------*/
//Isorropia Test - graph/Parmetis
pLUList->set("Partitioning Package","Isorropia");
Teuchos::ParameterList ptemp;
ptemp = pLUList->sublist("Isorropia Input");
Teuchos::ParameterList pptemp;
pptemp = ptemp.sublist("Zoltan");
pptemp.set("GRAPH_PACKAGE", "Parmetis");
pptemp.set("DEBUG_LEVEL", "1");
ptemp.set("partitioning method", "graph");
ptemp.set("Zoltan", pptemp);
pLUList->set("Isorropia Input", ptemp);
cout << " \n\n--------------------BIG BREAK --------------\n\n";
Teuchos::writeParameterListToXmlOStream(*pLUList, std::cout);
ShyLU::PartitionInterface<Epetra_CrsMatrix, Epetra_MultiVector> partI(A, pLUList.get());
partI.partition();
AHat = partI.reorderMatrix();
bHat = partI.reorderVector(b);
EpetraExt::RowMatrixToMatlabFile("Epetra_Isorropia_Parmetis.mat", *AHat);
cout << "Done with graph - parmetis" << endl;
/*
//Isorropia Test - Graph/PT-Scotch
pLUList->set("Partitioning Package","Isorropia");
ptemp = pLUList->sublist("Isorropia Input");
//Teuchos::ParameterList pptemp;
pptemp = ptemp.sublist("Zoltan");
pptemp.set("GRAPH_PACKAGE", "scotch");
pptemp.set("DEBUG_LEVEL", "1");
ptemp.set("partitioning method", "graph");
ptemp.set("Zoltan", pptemp);
pLUList->set("Isorropia Input", ptemp);
cout << " \n\n--------------------BIG BREAK --------------\n\n";
Teuchos::writeParameterListToXmlOStream(*pLUList, std::cout);
PartitionInterface<Epetra_CrsMatrix, Epetra_MultiVector> partI2(A, pLUList.get());
partI2.partition();
AHat = partI2.reorderMatrix();
bHat = partI2.reorderVector(b);
cout << "Done with graph - pt-scotch" << endl;
*/
//Zoltan2 Test
#if defined(HAVE_SHYLU_ZOLTAN2) || defined(HAVE_SHYLU_ZOLTAN2)
//Isorropia Test - Graph/ParMetis
pLUList->set("Partitioning Package","Zoltan2");
ptemp = pLUList->sublist("Zoltan2 Input");
ptemp.set("algorithm", "parmetis");
ptemp.set("debug_level", "detailed_status");
pLUList->set("Zoltan2 Input", ptemp);
cout << " \n\n--------------------BIG BREAK --------------\n\n";
Teuchos::writeParameterListToXmlOStream(*pLUList, std::cout);
ShyLU::PartitionInterface<Epetra_CrsMatrix, Epetra_MultiVector> partI3(A, pLUList.get());
partI3.partition();
AHat = partI3.reorderMatrix();
bHat = partI3.reorderVector(b);
cout << "Done with graph - parmetis" << endl;
EpetraExt::RowMatrixToMatlabFile("Epetra_Zoltan2_Parmetis.mat", *AHat);
#endif
/*----------------------Direct Solver Interfaces----------------*/
//#ifdef HAVE_SHYLU_AMESOS
//Amesos - klu
pLUList->set("Direct Solver Package", "Amesos");
ptemp = pLUList->sublist("Amesos Input");
pptemp = ptemp.sublist("Amesos_Klu Input");
pptemp.set("PrintTiming", true);
pptemp.set("PrintStatus", true);
ptemp.set("Solver", "Amesos_Klu");
ptemp.set("Amesos_Klu Input", pptemp);
pLUList->set("Amesos Input", ptemp);
cout << " \n\n--------------------BIG BREAK --------------\n\n";
Teuchos::writeParameterListToXmlOStream(*pLUList, std::cout);
ShyLU::DirectSolverInterface<Epetra_CrsMatrix, Epetra_MultiVector> directsolver(A, pLUList.get());
directsolver.factor();
directsolver.solve(b,x);
cout << "Done with Amesos-KLU" << endl;
//#endif
//Amesos2 -klu2
#ifdef HAVE_SHYLU_AMESOS2
pLUList->set("Direct Solver Package", "Amesos2");
ptemp = pLUList->sublist("Amesos2 Input");
//pptemp = ptemp.sublist("Amesos_Klu Input");
pptemp.set("PrintTiming", true);
pptemp.set("PrintStatus", true);
ptemp.set("Solver", "KLU2");
//ptemp.set("Amesos_Klu Input", pptemp);
pLUList->set("Amesos2 Input", ptemp);
cout << " \n\n--------------------BIG BREAK --------------\n\n";
Teuchos::writeParameterListToXmlOStream(*pLUList, std::cout);
ShyLU::DirectSolverInterface<Epetra_CrsMatrix, Epetra_MultiVector> directsolver2(A, pLUList.get());
directsolver2.factor();
directsolver2.solve(b,x);
cout << "Done with Amesos-KLU2" << endl;
#endif
if(success)
{
cout << pass << endl;
}
}
