int
main (int argc, char *argv[])
{
// These "using" statements make the code a bit more concise.
using std::cout;
using std::endl;
int ierr = 0, i;
// If Trilinos was built with MPI, initialize MPI, otherwise
// initialize the serial "communicator" that stands in for MPI.
#ifdef EPETRA_MPI
MPI_Init (&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
const int MyPID = Comm.MyPID();
const int NumProc = Comm.NumProc();
// We only allow (MPI) Process 0 to write to stdout.
const bool verbose = (MyPID == 0);
const int NumGlobalElements = 100;
if (verbose)
cout << Epetra_Version() << endl << endl;
// Asking the Epetra_Comm to print itself is a good test for whether
// you are running in an MPI environment. However, it will print
// something on all MPI processes, so you should remove it for a
// large-scale parallel run.
cout << Comm << endl;
if (NumGlobalElements < NumProc)
{
if (verbose)
cout << "numGlobalBlocks = " << NumGlobalElements
<< " cannot be < number of processors = " << NumProc << endl;
std::exit (EXIT_FAILURE);
}
// Construct a Map that puts approximately the same number of rows
// of the matrix A 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]);
// NumNz[i] is the number of nonzero elements in row i of the sparse
// matrix on this MPI process. Epetra_CrsMatrix uses this to figure
// out how much space to allocate.
std::vector<int> NumNz (NumMyElements);
// We are building a tridiagonal matrix where each row contains the
// nonzero elements (-1 2 -1). Thus, we need 2 off-diagonal terms,
// except for the first and last row of the matrix.
for (int i = 0; i < NumMyElements; ++i)
if (MyGlobalElements[i] == 0 || MyGlobalElements[i] == NumGlobalElements-1)
NumNz[i] = 2; // First or last row
else
NumNz[i] = 3; // Not the (first or last row)
// Create the Epetra_CrsMatrix.
Epetra_CrsMatrix A (Copy, Map, &NumNz[0]);
//
// Add rows to the sparse matrix one at a time.
//
std::vector<double> Values(2);
Values[0] = -1.0; Values[1] = -1.0;
std::vector<int> Indices(2);
const double two = 2.0;
int NumEntries;
for (int i = 0; i < NumMyElements; ++i)
{
if (MyGlobalElements[i] == 0)
{ // The first row of the matrix.
Indices[0] = 1;
NumEntries = 1;
}
else if (MyGlobalElements[i] == NumGlobalElements - 1)
{ // The last row of the matrix.
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
}
else
{ // Any row of the matrix other than the first or last.
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);
// Insert the diagonal entry.
ierr = A.InsertGlobalValues(MyGlobalElements[i], 1, &two, &MyGlobalElements[i]);
assert(ierr==0);
}
// Finish up. We can call FillComplete() with no arguments, because
// the matrix is square.
ierr = A.FillComplete ();
assert (ierr==0);
// Parameters for the power method.
const int niters = NumGlobalElements*10;
const double tolerance = 1.0e-2;
//
// Run the power method. Keep track of the flop count and the total
// elapsed time.
//
Epetra_Flops counter;
A.SetFlopCounter(counter);
Epetra_Time timer(Comm);
double lambda = 0.0;
ierr += powerMethod (lambda, A, niters, tolerance, verbose);
double elapsedTime = timer.ElapsedTime ();
double totalFlops =counter.Flops ();
// Mflop/s: Million floating-point arithmetic operations per second.
double Mflop_per_s = totalFlops / elapsedTime / 1000000.0;
if (verbose)
cout << endl << endl << "Total Mflop/s for first solve = "
<< Mflop_per_s << endl<< endl;
// Increase the first (0,0) diagonal entry of the matrix.
if (verbose)
cout << endl << "Increasing magnitude of first diagonal term, solving again"
<< endl << endl << endl;
if (A.MyGlobalRow (0)) {
int numvals = A.NumGlobalEntries (0);
std::vector<double> Rowvals (numvals);
std::vector<int> Rowinds (numvals);
A.ExtractGlobalRowCopy (0, numvals, numvals, &Rowvals[0], &Rowinds[0]); // Get A(0,0)
for (int i = 0; i < numvals; ++i)
if (Rowinds[i] == 0)
Rowvals[i] *= 10.0;
A.ReplaceGlobalValues (0, numvals, &Rowvals[0], &Rowinds[0]);
}
//
// Run the power method again. Keep track of the flop count and the
// total elapsed time.
//
lambda = 0.0;
timer.ResetStartTime();
counter.ResetFlops();
ierr += powerMethod (lambda, A, niters, tolerance, verbose);
elapsedTime = timer.ElapsedTime();
totalFlops = counter.Flops();
Mflop_per_s = totalFlops / elapsedTime / 1000000.0;
if (verbose)
cout << endl << endl << "Total Mflop/s for second solve = "
<< Mflop_per_s << endl << endl;
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return ierr;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
cout << Comm << endl;
int MyPID = Comm.MyPID();
bool verbose = false;
bool verbose1 = true;
if (MyPID==0) verbose = true;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename [level_fill [level_overlap [absolute_threshold [ relative_threshold]]]]" << endl
<< "where:" << endl
<< "HB_filename - filename and path of a Harwell-Boeing data set" << endl
<< "level_fill - The amount of fill to use for ILU(k) preconditioner (default 0)" << endl
<< "level_overlap - The amount of overlap used for overlapping Schwarz subdomains (default 0)" << endl
<< "absolute_threshold - The minimum value to place on the diagonal prior to factorization (default 0.0)" << endl
<< "relative_threshold - The relative amount to perturb the diagonal prior to factorization (default 1.0)" << endl << endl
<< "To specify a non-default value for one of these parameters, you must specify all" << endl
<< " preceding values but not any subsequent parameters. Example:" << endl
<< "ifpackHpcSerialMsr.exe mymatrix.hpc 1 - loads mymatrix.hpc, uses level fill of one, all other values are defaults" << endl
<< endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
bool smallProblem = false;
if (A.RowMap().NumGlobalElements()<100) smallProblem = true;
if (smallProblem)
cout << "Original Matrix = " << endl << A << endl;
x.PutScalar(0.0);
Epetra_LinearProblem FullProblem(&A, &x, &b);
double normb, norma;
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
Epetra_Time ReductionTimer(Comm);
Epetra_CrsSingletonFilter SingletonFilter;
Comm.Barrier();
double reduceInitTime = ReductionTimer.ElapsedTime();
SingletonFilter.Analyze(&A);
Comm.Barrier();
double reduceAnalyzeTime = ReductionTimer.ElapsedTime() - reduceInitTime;
if (SingletonFilter.SingletonsDetected())
cout << "Singletons found" << endl;
else {
cout << "Singletons not found" << endl;
exit(1);
}
SingletonFilter.ConstructReducedProblem(&FullProblem);
Comm.Barrier();
double reduceConstructTime = ReductionTimer.ElapsedTime() - reduceInitTime;
double totalReduceTime = ReductionTimer.ElapsedTime();
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Reduction init time (sec) = " << reduceInitTime<< endl
<< " Reduction Analyze time (sec) = " << reduceAnalyzeTime << endl
<< " Construct Reduced Problem time (sec) = " << reduceConstructTime << endl
<< " Reduction Total time (sec) = " << totalReduceTime << endl<< endl;
Statistics(SingletonFilter);
Epetra_LinearProblem * ReducedProblem = SingletonFilter.ReducedProblem();
Epetra_CrsMatrix * Ap = dynamic_cast<Epetra_CrsMatrix *>(ReducedProblem->GetMatrix());
Epetra_Vector * bp = (*ReducedProblem->GetRHS())(0);
Epetra_Vector * xp = (*ReducedProblem->GetLHS())(0);
if (smallProblem)
cout << " Reduced Matrix = " << endl << *Ap << endl
<< " LHS before sol = " << endl << *xp << endl
<< " RHS = " << endl << *bp << endl;
// Construct ILU preconditioner
double elapsed_time, total_flops, MFLOPs;
Epetra_Time timer(Comm);
int LevelFill = 0;
if (argc > 2) LevelFill = atoi(argv[2]);
if (verbose) cout << "Using Level Fill = " << LevelFill << endl;
int Overlap = 0;
if (argc > 3) Overlap = atoi(argv[3]);
if (verbose) cout << "Using Level Overlap = " << Overlap << endl;
double Athresh = 0.0;
if (argc > 4) Athresh = atof(argv[4]);
if (verbose) cout << "Using Absolute Threshold Value of = " << Athresh << endl;
double Rthresh = 1.0;
if (argc > 5) Rthresh = atof(argv[5]);
if (verbose) cout << "Using Relative Threshold Value of = " << Rthresh << endl;
Ifpack_IlukGraph * IlukGraph = 0;
Ifpack_CrsRiluk * ILUK = 0;
if (LevelFill>-1) {
elapsed_time = timer.ElapsedTime();
IlukGraph = new Ifpack_IlukGraph(Ap->Graph(), LevelFill, Overlap);
assert(IlukGraph->ConstructFilledGraph()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
if (verbose) cout << "Time to construct ILUK graph = " << elapsed_time << endl;
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
ILUK = new Ifpack_CrsRiluk(*IlukGraph);
ILUK->SetFlopCounter(fact_counter);
ILUK->SetAbsoluteThreshold(Athresh);
ILUK->SetRelativeThreshold(Rthresh);
//assert(ILUK->InitValues()==0);
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
//cout << *ILUK << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
int Maxiter = 100;
double Tolerance = 1.0E-8;
Epetra_Flops counter;
Ap->SetFlopCounter(counter);
xp->SetFlopCounter(*Ap);
bp->SetFlopCounter(*Ap);
if (ILUK!=0) ILUK->SetFlopCounter(*Ap);
elapsed_time = timer.ElapsedTime();
double normreducedb, normreduceda;
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
Epetra_Vector resid(x);
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
double normx, normxexact;
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (verbose)
cout << "********************************************************" << endl
<< " Solving again with 2*Ax=2*b" << endl
<< "********************************************************" << endl;
A.Scale(1.0); // A = 2*A
b.Scale(1.0); // b = 2*b
x.PutScalar(0.0);
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
double updateReducedProblemTime = ReductionTimer.ElapsedTime();
SingletonFilter.UpdateReducedProblem(&FullProblem);
Comm.Barrier();
updateReducedProblemTime = ReductionTimer.ElapsedTime() - updateReducedProblemTime;
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Update Reduced Problem time (sec) = " << updateReducedProblemTime<< endl
<< "****************************************************" << endl;
Statistics(SingletonFilter);
if (LevelFill>-1) {
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (ILUK!=0) delete ILUK;
if (IlukGraph!=0) delete IlukGraph;
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
