HypreSolver3D::HypreSolver3D(ptrdiff_t N0_, double Lz_,
ptrdiff_t N1_, double Ly_,
ptrdiff_t N2_, double Lx_):
N0(N0_), Lz(Lz_), N1(N1_), Ly(Ly_), N2(N2_), Lx(Lx_)
{
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
N = N0*N1*N2;
dz = Lz/(N0+1);
dy = Ly/(N1+1);
dx = Lx/(N2+1);
h2 = dy*dy;
nx = N2;
ny = N1;
int zp = (int)(N0/num_procs);
int extra = N0-zp*num_procs;
if(myid<extra)
{
z0 = myid*zp + myid;
nz = zp+1;
}
else
{
z0 = myid*zp + extra;
nz = zp;
}
local_size = ny*nx*nz;
ilower = z0*nx*ny;
iupper = ilower+local_size-1;
// Build rhs and init x
HYPRE_IJVectorCreate(MPI_COMM_WORLD, ilower, iupper, &hv_b);
HYPRE_IJVectorSetObjectType(hv_b, HYPRE_PARCSR);
HYPRE_IJVectorCreate(MPI_COMM_WORLD, ilower, iupper, &hv_x);
HYPRE_IJVectorSetObjectType(hv_x, HYPRE_PARCSR);
std::vector<double> rhs_values(local_size, 1.);
std::vector<double> x_values(local_size, 0.);
rows.resize(local_size);
for (int i=0; i<local_size; i++)
rows[i] = ilower + i;
set_x0(&x_values[0]);
set_rhs(&rhs_values[0]);
build_A();
build_solver();
}
int
HYPRE_IJVectorRead( const char *filename,
MPI_Comm comm,
int type,
HYPRE_IJVector *vector_ptr )
{
HYPRE_IJVector vector;
HYPRE_BigInt jlower, jupper, j;
double value;
int myid;
char new_filename[255];
FILE *file;
MPI_Comm_rank(comm, &myid);
sprintf(new_filename,"%s.%05d", filename, myid);
if ((file = fopen(new_filename, "r")) == NULL)
{
printf("Error: can't open input file %s\n", new_filename);
hypre_error_in_arg(1);
return hypre_error_flag;
}
#ifdef HYPRE_LONG_LONG
fscanf(file, "%lld %lld", &jlower, &jupper);
#else
fscanf(file, "%d %d", &jlower, &jupper);
#endif
HYPRE_IJVectorCreate(comm, jlower, jupper, &vector);
HYPRE_IJVectorSetObjectType(vector, type);
HYPRE_IJVectorInitialize(vector);
#ifdef HYPRE_LONG_LONG
while ( fscanf(file, "%lld %le", &j, &value) != EOF )
#else
while ( fscanf(file, "%d %le", &j, &value) != EOF )
#endif
{
HYPRE_IJVectorSetValues(vector, 1, &j, &value);
}
HYPRE_IJVectorAssemble(vector);
fclose(file);
*vector_ptr = vector;
return hypre_error_flag;
}
HYPRE_Int hypre_BlockTridiagSetup(void *data, hypre_ParCSRMatrix *A,
hypre_ParVector *b, hypre_ParVector *x)
{
HYPRE_Int i, j, *index_set1, print_level, nsweeps, relax_type;
HYPRE_Int nrows, nrows1, nrows2, start1, start2, *index_set2;
HYPRE_Int count, ierr;
double threshold;
hypre_ParCSRMatrix **submatrices;
HYPRE_Solver precon1;
HYPRE_Solver precon2;
HYPRE_IJVector ij_u1, ij_u2, ij_f1, ij_f2;
hypre_ParVector *vector;
MPI_Comm comm;
hypre_BlockTridiagData *b_data = (hypre_BlockTridiagData *) data;
HYPRE_ParCSRMatrixGetComm((HYPRE_ParCSRMatrix) A, &comm);
index_set1 = b_data->index_set1;
nrows1 = index_set1[0];
nrows = hypre_ParCSRMatrixNumRows(A);
nrows2 = nrows - nrows1;
b_data->index_set2 = hypre_CTAlloc(HYPRE_Int, nrows2+1);
index_set2 = b_data->index_set2;
index_set2[0] = nrows2;
count = 1;
for (i = 0; i < index_set1[1]; i++) index_set2[count++] = i;
for (i = 1; i < nrows1; i++)
for (j = index_set1[i]+1; j < index_set1[i+1]; j++)
index_set2[count++] = j;
for (i = index_set1[nrows1]+1; i < nrows; i++) index_set2[count++] = i;
submatrices = hypre_CTAlloc(hypre_ParCSRMatrix *, 4);
hypre_ParCSRMatrixExtractSubmatrices(A, index_set1, &submatrices);
nrows1 = hypre_ParCSRMatrixNumRows(submatrices[0]);
nrows2 = hypre_ParCSRMatrixNumRows(submatrices[3]);
start1 = hypre_ParCSRMatrixFirstRowIndex(submatrices[0]);
start2 = hypre_ParCSRMatrixFirstRowIndex(submatrices[3]);
HYPRE_IJVectorCreate(comm, start1, start1+nrows1-1, &ij_u1);
HYPRE_IJVectorSetObjectType(ij_u1, HYPRE_PARCSR);
ierr = HYPRE_IJVectorInitialize(ij_u1);
ierr += HYPRE_IJVectorAssemble(ij_u1);
hypre_assert(!ierr);
HYPRE_IJVectorCreate(comm, start1, start1+nrows1-1, &ij_f1);
HYPRE_IJVectorSetObjectType(ij_f1, HYPRE_PARCSR);
ierr = HYPRE_IJVectorInitialize(ij_f1);
ierr += HYPRE_IJVectorAssemble(ij_f1);
hypre_assert(!ierr);
HYPRE_IJVectorCreate(comm, start2, start2+nrows2-1, &ij_u2);
HYPRE_IJVectorSetObjectType(ij_u2, HYPRE_PARCSR);
ierr = HYPRE_IJVectorInitialize(ij_u2);
ierr += HYPRE_IJVectorAssemble(ij_u2);
hypre_assert(!ierr);
HYPRE_IJVectorCreate(comm, start2, start2+nrows1-1, &ij_f2);
HYPRE_IJVectorSetObjectType(ij_f2, HYPRE_PARCSR);
ierr = HYPRE_IJVectorInitialize(ij_f2);
ierr += HYPRE_IJVectorAssemble(ij_f2);
hypre_assert(!ierr);
HYPRE_IJVectorGetObject(ij_f1, (void **) &vector);
b_data->F1 = vector;
HYPRE_IJVectorGetObject(ij_u1, (void **) &vector);
b_data->U1 = vector;
HYPRE_IJVectorGetObject(ij_f2, (void **) &vector);
b_data->F2 = vector;
HYPRE_IJVectorGetObject(ij_u2, (void **) &vector);
b_data->U2 = vector;
print_level = b_data->print_level;
threshold = b_data->threshold;
nsweeps = b_data->num_sweeps;
relax_type = b_data->relax_type;
threshold = b_data->threshold;
HYPRE_BoomerAMGCreate(&precon1);
HYPRE_BoomerAMGSetMaxIter(precon1, 1);
HYPRE_BoomerAMGSetCycleType(precon1, 1);
HYPRE_BoomerAMGSetPrintLevel(precon1, print_level);
HYPRE_BoomerAMGSetMaxLevels(precon1, 25);
HYPRE_BoomerAMGSetMeasureType(precon1, 0);
HYPRE_BoomerAMGSetCoarsenType(precon1, 0);
HYPRE_BoomerAMGSetStrongThreshold(precon1, threshold);
HYPRE_BoomerAMGSetNumFunctions(precon1, 1);
HYPRE_BoomerAMGSetNumSweeps(precon1, nsweeps);
HYPRE_BoomerAMGSetRelaxType(precon1, relax_type);
hypre_BoomerAMGSetup(precon1, submatrices[0], b_data->U1, b_data->F1);
HYPRE_BoomerAMGCreate(&precon2);
HYPRE_BoomerAMGSetMaxIter(precon2, 1);
HYPRE_BoomerAMGSetCycleType(precon2, 1);
HYPRE_BoomerAMGSetPrintLevel(precon2, print_level);
HYPRE_BoomerAMGSetMaxLevels(precon2, 25);
HYPRE_BoomerAMGSetMeasureType(precon2, 0);
HYPRE_BoomerAMGSetCoarsenType(precon2, 0);
HYPRE_BoomerAMGSetMeasureType(precon2, 1);
HYPRE_BoomerAMGSetStrongThreshold(precon2, threshold);
HYPRE_BoomerAMGSetNumFunctions(precon2, 1);
HYPRE_BoomerAMGSetNumSweeps(precon2, nsweeps);
HYPRE_BoomerAMGSetRelaxType(precon2, relax_type);
hypre_BoomerAMGSetup(precon2, submatrices[3], NULL, NULL);
b_data->precon1 = precon1;
b_data->precon2 = precon2;
b_data->A11 = submatrices[0];
hypre_ParCSRMatrixDestroy(submatrices[1]);
b_data->A21 = submatrices[2];
b_data->A22 = submatrices[3];
hypre_TFree(submatrices);
return (0);
}
//=======================================================
EpetraExt_HypreIJMatrix::EpetraExt_HypreIJMatrix(HYPRE_IJMatrix matrix)
: Epetra_BasicRowMatrix(Epetra_MpiComm(hypre_IJMatrixComm(matrix))),
Matrix_(matrix),
ParMatrix_(0),
NumMyRows_(-1),
NumGlobalRows_(-1),
NumGlobalCols_(-1),
MyRowStart_(-1),
MyRowEnd_(-1),
MatType_(-1),
TransposeSolve_(false),
SolveOrPrec_(Solver)
{
IsSolverSetup_ = new bool[1];
IsPrecondSetup_ = new bool[1];
IsSolverSetup_[0] = false;
IsPrecondSetup_[0] = false;
// Initialize default values for global variables
int ierr = 0;
ierr += InitializeDefaults();
TEUCHOS_TEST_FOR_EXCEPTION(ierr != 0, std::logic_error, "Couldn't initialize default values.");
// Create array of global row ids
Teuchos::Array<int> GlobalRowIDs; GlobalRowIDs.resize(NumMyRows_);
for (int i = MyRowStart_; i <= MyRowEnd_; i++) {
GlobalRowIDs[i-MyRowStart_] = i;
}
// Create array of global column ids
int new_value = 0; int entries = 0;
std::set<int> Columns;
int num_entries;
double *values;
int *indices;
for(int i = 0; i < NumMyRows_; i++){
ierr += HYPRE_ParCSRMatrixGetRow(ParMatrix_, i+MyRowStart_, &num_entries, &indices, &values);
ierr += HYPRE_ParCSRMatrixRestoreRow(ParMatrix_, i+MyRowStart_,&num_entries,&indices,&values);
TEUCHOS_TEST_FOR_EXCEPTION(ierr != 0, std::logic_error, "Couldn't get row of matrix.");
entries = num_entries;
for(int j = 0; j < num_entries; j++){
// Insert column ids from this row into set
new_value = indices[j];
Columns.insert(new_value);
}
}
int NumMyCols = Columns.size();
Teuchos::Array<int> GlobalColIDs; GlobalColIDs.resize(NumMyCols);
std::set<int>::iterator it;
int counter = 0;
for (it = Columns.begin(); it != Columns.end(); it++) {
// Get column ids in order
GlobalColIDs[counter] = *it;
counter = counter + 1;
}
//printf("Proc[%d] Rows from %d to %d, num = %d\n", Comm().MyPID(), MyRowStart_,MyRowEnd_, NumMyRows_);
Epetra_Map RowMap(-1, NumMyRows_, &GlobalRowIDs[0], 0, Comm());
Epetra_Map ColMap(-1, NumMyCols, &GlobalColIDs[0], 0, Comm());
//Need to call SetMaps()
SetMaps(RowMap, ColMap);
// Get an MPI_Comm to create vectors.
// The vectors will be reused in Multiply(), so that they aren't recreated every time.
MPI_Comm comm;
ierr += HYPRE_ParCSRMatrixGetComm(ParMatrix_, &comm);
TEUCHOS_TEST_FOR_EXCEPTION(ierr != 0, std::logic_error, "Couldn't get communicator from Hypre Matrix.");
ierr += HYPRE_IJVectorCreate(comm, MyRowStart_, MyRowEnd_, &X_hypre);
ierr += HYPRE_IJVectorSetObjectType(X_hypre, HYPRE_PARCSR);
ierr += HYPRE_IJVectorInitialize(X_hypre);
ierr += HYPRE_IJVectorAssemble(X_hypre);
ierr += HYPRE_IJVectorGetObject(X_hypre, (void**) &par_x);
TEUCHOS_TEST_FOR_EXCEPTION(ierr != 0, std::logic_error, "Couldn't create Hypre X vector.");
ierr += HYPRE_IJVectorCreate(comm, MyRowStart_, MyRowEnd_, &Y_hypre);
ierr += HYPRE_IJVectorSetObjectType(Y_hypre, HYPRE_PARCSR);
ierr += HYPRE_IJVectorInitialize(Y_hypre);
ierr += HYPRE_IJVectorAssemble(Y_hypre);
ierr += HYPRE_IJVectorGetObject(Y_hypre, (void**) &par_y);
TEUCHOS_TEST_FOR_EXCEPTION(ierr != 0, std::logic_error, "Couldn't create Hypre Y vector.");
x_vec = (hypre_ParVector *) hypre_IJVectorObject(((hypre_IJVector *) X_hypre));
x_local = hypre_ParVectorLocalVector(x_vec);
y_vec = (hypre_ParVector *) hypre_IJVectorObject(((hypre_IJVector *) Y_hypre));
y_local = hypre_ParVectorLocalVector(y_vec);
SolverCreatePtr_ = &EpetraExt_HypreIJMatrix::Hypre_ParCSRPCGCreate;
SolverDestroyPtr_ = &HYPRE_ParCSRPCGDestroy;
SolverSetupPtr_ = &HYPRE_ParCSRPCGSetup;
SolverSolvePtr_ = &HYPRE_ParCSRPCGSolve;
SolverPrecondPtr_ = &HYPRE_ParCSRPCGSetPrecond;
CreateSolver();
PrecondCreatePtr_ = &EpetraExt_HypreIJMatrix::Hypre_EuclidCreate;
PrecondDestroyPtr_ = &HYPRE_EuclidDestroy;
PrecondSetupPtr_ = &HYPRE_EuclidSetup;
PrecondSolvePtr_ = &HYPRE_EuclidSolve;
CreatePrecond();
ComputeNumericConstants();
ComputeStructureConstants();
} //EpetraExt_HYPREIJMatrix(Hypre_IJMatrix) Constructor
inline void numfact(unsigned int ncol, int* I, int* loc2glob, int* J, K* C) {
static_assert(std::is_same<double, K>::value, "Hypre only supports double-precision floating-point real numbers");
static_assert(S == 'G', "Hypre only supports nonsymmetric matrices");
HYPRE_IJMatrixCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], loc2glob[0], loc2glob[1], &_A);
HYPRE_IJMatrixSetObjectType(_A, HYPRE_PARCSR);
HYPRE_IJMatrixSetRowSizes(_A, I + 1);
_local = ncol;
int* rows = new int[3 * _local]();
int* diag_sizes = rows + _local;
int* offdiag_sizes = diag_sizes + _local;
rows[0] = I[0];
for(unsigned int i = 0; i < _local; ++i) {
std::for_each(J + rows[0], J + rows[0] + I[i + 1], [&](int& j) { (j < loc2glob[0] || loc2glob[1] < j) ? ++offdiag_sizes[i] : ++diag_sizes[i]; });
rows[0] += I[i + 1];
}
HYPRE_IJMatrixSetDiagOffdSizes(_A, diag_sizes, offdiag_sizes);
HYPRE_IJMatrixSetMaxOffProcElmts(_A, 0);
HYPRE_IJMatrixInitialize(_A);
std::iota(rows, rows + _local, loc2glob[0]);
HYPRE_IJMatrixSetValues(_A, _local, I + 1, rows, J, C);
HYPRE_IJMatrixAssemble(_A);
HYPRE_IJVectorCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], &_b);
HYPRE_IJVectorSetObjectType(_b, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(_b);
HYPRE_IJVectorCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], &_x);
HYPRE_IJVectorSetObjectType(_x, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(_x);
delete [] rows;
delete [] I;
delete [] loc2glob;
HYPRE_BoomerAMGCreate(_strategy == 1 ? &_solver : &_precond);
HYPRE_BoomerAMGSetCoarsenType(_strategy == 1 ? _solver : _precond, 6); /* Falgout coarsening */
HYPRE_BoomerAMGSetRelaxType(_strategy == 1 ? _solver : _precond, 6); /* G-S/Jacobi hybrid relaxation */
HYPRE_BoomerAMGSetNumSweeps(_strategy == 1 ? _solver : _precond, 1); /* sweeps on each level */
HYPRE_BoomerAMGSetMaxLevels(_strategy == 1 ? _solver : _precond, 10); /* maximum number of levels */
HYPRE_ParCSRMatrix parcsr_A;
HYPRE_IJMatrixGetObject(_A, reinterpret_cast<void**>(&parcsr_A));
HYPRE_ParVector par_b;
HYPRE_IJVectorGetObject(_b, reinterpret_cast<void**>(&par_b));
HYPRE_ParVector par_x;
HYPRE_IJVectorGetObject(_x, reinterpret_cast<void**>(&par_x));
if(_strategy == 1) {
HYPRE_BoomerAMGSetTol(_solver, 1.0e-8);
HYPRE_BoomerAMGSetMaxIter(_solver, 1000);
HYPRE_BoomerAMGSetPrintLevel(_solver, 1);
HYPRE_BoomerAMGSetup(_solver, parcsr_A, nullptr, nullptr);
}
else {
HYPRE_BoomerAMGSetTol(_precond, 0.0);
HYPRE_BoomerAMGSetMaxIter(_precond, 1);
HYPRE_BoomerAMGSetPrintLevel(_precond, 1);
if(_strategy == 2) {
HYPRE_ParCSRPCGCreate(DMatrix::_communicator, &_solver);
HYPRE_PCGSetMaxIter(_solver, 500);
HYPRE_PCGSetTol(_solver, 1.0e-8);
HYPRE_PCGSetTwoNorm(_solver, 1);
HYPRE_PCGSetPrintLevel(_solver, 1);
HYPRE_PCGSetLogging(_solver, 1);
HYPRE_PCGSetPrecond(_solver, reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSolve), reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSetup), _precond);
HYPRE_ParCSRPCGSetup(_solver, parcsr_A, par_b, par_x);
}
else {
HYPRE_ParCSRFlexGMRESCreate(DMatrix::_communicator, &_solver);
HYPRE_FlexGMRESSetKDim(_solver, 50);
HYPRE_FlexGMRESSetMaxIter(_solver, 500);
HYPRE_FlexGMRESSetTol(_solver, 1.0e-8);
HYPRE_FlexGMRESSetPrintLevel(_solver, 1);
HYPRE_FlexGMRESSetLogging(_solver, 1);
HYPRE_FlexGMRESSetPrecond(_solver, reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSolve), reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSetup), _precond);
HYPRE_ParCSRFlexGMRESSetup(_solver, parcsr_A, par_b, par_x);
}
}
}
int main (int argc, char *argv[])
{
HYPRE_Int i;
int myid, num_procs;
int N, n;
HYPRE_Int ilower, iupper;
HYPRE_Int local_size, extra;
int solver_id;
int print_solution, print_system;
double h, h2;
HYPRE_IJMatrix A;
HYPRE_ParCSRMatrix parcsr_A;
HYPRE_IJVector b;
HYPRE_ParVector par_b;
HYPRE_IJVector x;
HYPRE_ParVector par_x;
HYPRE_Solver solver, precond;
/* Initialize MPI */
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
/* Default problem parameters */
n = 33;
solver_id = 0;
print_solution = 0;
print_system = 0;
/* Parse command line */
{
int arg_index = 0;
int print_usage = 0;
while (arg_index < argc)
{
if ( strcmp(argv[arg_index], "-n") == 0 )
{
arg_index++;
n = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-solver") == 0 )
{
arg_index++;
solver_id = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-print_solution") == 0 )
{
arg_index++;
print_solution = 1;
}
else if ( strcmp(argv[arg_index], "-print_system") == 0 )
{
arg_index++;
print_system = 1;
}
else if ( strcmp(argv[arg_index], "-help") == 0 )
{
print_usage = 1;
break;
}
else
{
arg_index++;
}
}
if ((print_usage) && (myid == 0))
{
printf("\n");
printf("Usage: %s [<options>]\n", argv[0]);
printf("\n");
printf(" -n <n> : problem size in each direction (default: 33)\n");
printf(" -solver <ID> : solver ID\n");
printf(" 0 - AMG (default) \n");
printf(" 1 - AMG-PCG\n");
printf(" 8 - ParaSails-PCG\n");
printf(" 50 - PCG\n");
printf(" 61 - AMG-FlexGMRES\n");
printf(" -print_solution : print the solution vector\n");
printf(" -print_system : print the matrix and rhs\n");
printf("\n");
}
if (print_usage)
{
MPI_Finalize();
return (0);
}
}
/* Preliminaries: want at least one processor per row */
if (n*n < num_procs) n = sqrt(num_procs) + 1;
N = n*n; /* global number of rows */
h = 1.0/(n+1); /* mesh size*/
h2 = h*h;
/* Each processor knows only of its own rows - the range is denoted by ilower
and upper. Here we partition the rows. We account for the fact that
N may not divide evenly by the number of processors. */
local_size = N/num_procs;
extra = N - local_size*num_procs;
ilower = local_size*myid;
ilower += hypre_min(myid, extra);
iupper = local_size*(myid+1);
iupper += hypre_min(myid+1, extra);
iupper = iupper - 1;
/* How many rows do I have? */
local_size = iupper - ilower + 1;
/* Create the matrix.
Note that this is a square matrix, so we indicate the row partition
size twice (since number of rows = number of cols) */
HYPRE_IJMatrixCreate(MPI_COMM_WORLD, ilower, iupper, ilower, iupper, &A);
/* Choose a parallel csr format storage (see the User's Manual) */
HYPRE_IJMatrixSetObjectType(A, HYPRE_PARCSR);
/* Initialize before setting coefficients */
HYPRE_IJMatrixInitialize(A);
/* Now go through my local rows and set the matrix entries.
Each row has at most 5 entries. For example, if n=3:
A = [M -I 0; -I M -I; 0 -I M]
M = [4 -1 0; -1 4 -1; 0 -1 4]
Note that here we are setting one row at a time, though
one could set all the rows together (see the User's Manual).
*/
{
HYPRE_Int nnz;
double values[5];
HYPRE_Int cols[5];
for (i = ilower; i <= iupper; i++)
{
nnz = 0;
/* The left identity block:position i-n */
if ((i-n)>=0)
{
cols[nnz] = i-n;
values[nnz] = -1.0;
nnz++;
}
/* The left -1: position i-1 */
if (i%n)
{
cols[nnz] = i-1;
values[nnz] = -1.0;
nnz++;
}
/* Set the diagonal: position i */
cols[nnz] = i;
values[nnz] = 4.0;
nnz++;
/* The right -1: position i+1 */
if ((i+1)%n)
{
cols[nnz] = i+1;
values[nnz] = -1.0;
nnz++;
}
/* The right identity block:position i+n */
if ((i+n)< N)
{
cols[nnz] = i+n;
values[nnz] = -1.0;
nnz++;
}
/* Set the values for row i */
HYPRE_IJMatrixSetValues(A, 1, &nnz, &i, cols, values);
}
}
/* Assemble after setting the coefficients */
HYPRE_IJMatrixAssemble(A);
/* Note: for the testing of small problems, one may wish to read
in a matrix in IJ format (for the format, see the output files
from the -print_system option).
In this case, one would use the following routine:
HYPRE_IJMatrixRead( <filename>, MPI_COMM_WORLD,
HYPRE_PARCSR, &A );
<filename> = IJ.A.out to read in what has been printed out
by -print_system (processor numbers are omitted).
A call to HYPRE_IJMatrixRead is an *alternative* to the
following sequence of HYPRE_IJMatrix calls:
Create, SetObjectType, Initialize, SetValues, and Assemble
*/
/* Get the parcsr matrix object to use */
HYPRE_IJMatrixGetObject(A, (void**) &parcsr_A);
/* Create the rhs and solution */
HYPRE_IJVectorCreate(MPI_COMM_WORLD, ilower, iupper,&b);
HYPRE_IJVectorSetObjectType(b, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(b);
HYPRE_IJVectorCreate(MPI_COMM_WORLD, ilower, iupper,&x);
HYPRE_IJVectorSetObjectType(x, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(x);
/* Set the rhs values to h^2 and the solution to zero */
{
double *rhs_values, *x_values;
HYPRE_Int *rows;
rhs_values = calloc(local_size, sizeof(double));
x_values = calloc(local_size, sizeof(double));
rows = calloc(local_size, sizeof(HYPRE_Int));
for (i=0; i<local_size; i++)
{
rhs_values[i] = h2;
x_values[i] = 0.0;
rows[i] = ilower + i;
}
HYPRE_IJVectorSetValues(b, local_size, rows, rhs_values);
HYPRE_IJVectorSetValues(x, local_size, rows, x_values);
free(x_values);
free(rhs_values);
free(rows);
}
HYPRE_IJVectorAssemble(b);
/* As with the matrix, for testing purposes, one may wish to read in a rhs:
HYPRE_IJVectorRead( <filename>, MPI_COMM_WORLD,
HYPRE_PARCSR, &b );
as an alternative to the
following sequence of HYPRE_IJVectors calls:
Create, SetObjectType, Initialize, SetValues, and Assemble
*/
HYPRE_IJVectorGetObject(b, (void **) &par_b);
HYPRE_IJVectorAssemble(x);
HYPRE_IJVectorGetObject(x, (void **) &par_x);
/* Print out the system - files names will be IJ.out.A.XXXXX
and IJ.out.b.XXXXX, where XXXXX = processor id */
if (print_system)
{
HYPRE_IJMatrixPrint(A, "IJ.out.A");
HYPRE_IJVectorPrint(b, "IJ.out.b");
}
/* Choose a solver and solve the system */
/* AMG */
if (solver_id == 0)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
HYPRE_BoomerAMGCreate(&solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_BoomerAMGSetPrintLevel(solver, 3); /* print solve info + parameters */
HYPRE_BoomerAMGSetCoarsenType(solver, 6); /* Falgout coarsening */
HYPRE_BoomerAMGSetRelaxType(solver, 3); /* G-S/Jacobi hybrid relaxation */
HYPRE_BoomerAMGSetNumSweeps(solver, 1); /* Sweeeps on each level */
HYPRE_BoomerAMGSetMaxLevels(solver, 20); /* maximum number of levels */
HYPRE_BoomerAMGSetTol(solver, 1e-7); /* conv. tolerance */
/* Now setup and solve! */
HYPRE_BoomerAMGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_BoomerAMGSolve(solver, parcsr_A, par_b, par_x);
/* Run info - needed logging turned on */
HYPRE_BoomerAMGGetNumIterations(solver, &num_iterations);
HYPRE_BoomerAMGGetFinalRelativeResidualNorm(solver, &final_res_norm);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", num_iterations);
printf("Final Relative Residual Norm = %e\n", final_res_norm);
printf("\n");
}
/* Destroy solver */
HYPRE_BoomerAMGDestroy(solver);
}
/* PCG */
else if (solver_id == 50)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
HYPRE_ParCSRPCGCreate(MPI_COMM_WORLD, &solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* prints out the iteration info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
/* Now setup and solve! */
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
/* Run info - needed logging turned on */
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", num_iterations);
printf("Final Relative Residual Norm = %e\n", final_res_norm);
printf("\n");
}
/* Destroy solver */
HYPRE_ParCSRPCGDestroy(solver);
}
/* PCG with AMG preconditioner */
else if (solver_id == 1)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
HYPRE_ParCSRPCGCreate(MPI_COMM_WORLD, &solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* print solve info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
/* Now set up the AMG preconditioner and specify any parameters */
HYPRE_BoomerAMGCreate(&precond);
HYPRE_BoomerAMGSetPrintLevel(precond, 1); /* print amg solution info */
HYPRE_BoomerAMGSetCoarsenType(precond, 6);
HYPRE_BoomerAMGSetRelaxType(precond, 6); /* Sym G.S./Jacobi hybrid */
HYPRE_BoomerAMGSetNumSweeps(precond, 1);
HYPRE_BoomerAMGSetTol(precond, 0.0); /* conv. tolerance zero */
HYPRE_BoomerAMGSetMaxIter(precond, 1); /* do only one iteration! */
/* Set the PCG preconditioner */
HYPRE_PCGSetPrecond(solver, (HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSolve,
(HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSetup, precond);
/* Now setup and solve! */
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
/* Run info - needed logging turned on */
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", num_iterations);
printf("Final Relative Residual Norm = %e\n", final_res_norm);
printf("\n");
}
/* Destroy solver and preconditioner */
HYPRE_ParCSRPCGDestroy(solver);
HYPRE_BoomerAMGDestroy(precond);
}
/* PCG with Parasails Preconditioner */
else if (solver_id == 8)
{
HYPRE_Int num_iterations;
double final_res_norm;
int sai_max_levels = 1;
double sai_threshold = 0.1;
double sai_filter = 0.05;
int sai_sym = 1;
/* Create solver */
HYPRE_ParCSRPCGCreate(MPI_COMM_WORLD, &solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* print solve info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
/* Now set up the ParaSails preconditioner and specify any parameters */
HYPRE_ParaSailsCreate(MPI_COMM_WORLD, &precond);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_ParaSailsSetParams(precond, sai_threshold, sai_max_levels);
HYPRE_ParaSailsSetFilter(precond, sai_filter);
HYPRE_ParaSailsSetSym(precond, sai_sym);
HYPRE_ParaSailsSetLogging(precond, 3);
/* Set the PCG preconditioner */
HYPRE_PCGSetPrecond(solver, (HYPRE_PtrToSolverFcn) HYPRE_ParaSailsSolve,
(HYPRE_PtrToSolverFcn) HYPRE_ParaSailsSetup, precond);
/* Now setup and solve! */
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
/* Run info - needed logging turned on */
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", num_iterations);
printf("Final Relative Residual Norm = %e\n", final_res_norm);
printf("\n");
}
/* Destory solver and preconditioner */
HYPRE_ParCSRPCGDestroy(solver);
HYPRE_ParaSailsDestroy(precond);
}
/* Flexible GMRES with AMG Preconditioner */
else if (solver_id == 61)
{
HYPRE_Int num_iterations;
double final_res_norm;
int restart = 30;
int modify = 1;
/* Create solver */
HYPRE_ParCSRFlexGMRESCreate(MPI_COMM_WORLD, &solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_FlexGMRESSetKDim(solver, restart);
HYPRE_FlexGMRESSetMaxIter(solver, 1000); /* max iterations */
HYPRE_FlexGMRESSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_FlexGMRESSetPrintLevel(solver, 2); /* print solve info */
HYPRE_FlexGMRESSetLogging(solver, 1); /* needed to get run info later */
/* Now set up the AMG preconditioner and specify any parameters */
HYPRE_BoomerAMGCreate(&precond);
HYPRE_BoomerAMGSetPrintLevel(precond, 1); /* print amg solution info */
HYPRE_BoomerAMGSetCoarsenType(precond, 6);
HYPRE_BoomerAMGSetRelaxType(precond, 6); /* Sym G.S./Jacobi hybrid */
HYPRE_BoomerAMGSetNumSweeps(precond, 1);
HYPRE_BoomerAMGSetTol(precond, 0.0); /* conv. tolerance zero */
HYPRE_BoomerAMGSetMaxIter(precond, 1); /* do only one iteration! */
/* Set the FlexGMRES preconditioner */
HYPRE_FlexGMRESSetPrecond(solver, (HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSolve,
(HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSetup, precond);
if (modify)
/* this is an optional call - if you don't call it, hypre_FlexGMRESModifyPCDefault
is used - which does nothing. Otherwise, you can define your own, similar to
the one used here */
HYPRE_FlexGMRESSetModifyPC( solver,
(HYPRE_PtrToModifyPCFcn) hypre_FlexGMRESModifyPCAMGExample);
/* Now setup and solve! */
HYPRE_ParCSRFlexGMRESSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRFlexGMRESSolve(solver, parcsr_A, par_b, par_x);
/* Run info - needed logging turned on */
HYPRE_FlexGMRESGetNumIterations(solver, &num_iterations);
HYPRE_FlexGMRESGetFinalRelativeResidualNorm(solver, &final_res_norm);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", num_iterations);
printf("Final Relative Residual Norm = %e\n", final_res_norm);
printf("\n");
}
/* Destory solver and preconditioner */
HYPRE_ParCSRFlexGMRESDestroy(solver);
HYPRE_BoomerAMGDestroy(precond);
}
else
{
if (myid ==0) printf("Invalid solver id specified.\n");
}
/* Print the solution */
if (print_solution)
HYPRE_IJVectorPrint(x, "ij.out.x");
/* Clean up */
HYPRE_IJMatrixDestroy(A);
HYPRE_IJVectorDestroy(b);
HYPRE_IJVectorDestroy(x);
/* Finalize MPI*/
MPI_Finalize();
return(0);
}
int main(int argc, char *argv[])
{
realtype dx, reltol, abstol, t, tout, umax;
N_Vector u;
UserData data;
void *cvode_mem;
int iout, flag, my_pe, npes;
long int nst;
HYPRE_Int local_N, nperpe, nrem, my_base;
HYPRE_ParVector Upar; /* Declare HYPRE parallel vector */
HYPRE_IJVector Uij; /* Declare "IJ" interface to HYPRE vector */
MPI_Comm comm;
u = NULL;
data = NULL;
cvode_mem = NULL;
/* Get processor number, total number of pe's, and my_pe. */
MPI_Init(&argc, &argv);
comm = MPI_COMM_WORLD;
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &my_pe);
/* Set partitioning. */
nperpe = NEQ/npes;
nrem = NEQ - npes*nperpe;
local_N = (my_pe < nrem) ? nperpe+1 : nperpe;
my_base = (my_pe < nrem) ? my_pe*local_N : my_pe*nperpe + nrem;
/* Allocate hypre vector */
HYPRE_IJVectorCreate(comm, my_base, my_base + local_N - 1, &Uij);
HYPRE_IJVectorSetObjectType(Uij, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(Uij);
/* Allocate user defined data */
data = (UserData) malloc(sizeof *data); /* Allocate data memory */
if(check_flag((void *)data, "malloc", 2, my_pe)) MPI_Abort(comm, 1);
data->comm = comm;
data->npes = npes;
data->my_pe = my_pe;
reltol = ZERO; /* Set the tolerances */
abstol = ATOL;
dx = data->dx = XMAX/((realtype)(MX+1)); /* Set grid coefficients in data */
data->hdcoef = RCONST(1.0)/(dx*dx);
data->hacoef = RCONST(0.5)/(RCONST(2.0)*dx);
/* Initialize solutin vector. */
SetIC(Uij, dx, local_N, my_base);
HYPRE_IJVectorAssemble(Uij);
HYPRE_IJVectorGetObject(Uij, (void**) &Upar);
u = N_VMake_ParHyp(Upar); /* Create wrapper u around hypre vector */
if(check_flag((void *)u, "N_VNew", 0, my_pe)) MPI_Abort(comm, 1);
/* Call CVodeCreate to create the solver memory and specify the
* Adams-Moulton LMM and the use of a functional iteration */
cvode_mem = CVodeCreate(CV_ADAMS, CV_FUNCTIONAL);
if(check_flag((void *)cvode_mem, "CVodeCreate", 0, my_pe)) MPI_Abort(comm, 1);
flag = CVodeSetUserData(cvode_mem, data);
if(check_flag(&flag, "CVodeSetUserData", 1, my_pe)) MPI_Abort(comm, 1);
/* Call CVodeInit to initialize the integrator memory and specify the
* user's right hand side function in u'=f(t,u), the inital time T0, and
* the initial dependent variable vector u. */
flag = CVodeInit(cvode_mem, f, T0, u);
if(check_flag(&flag, "CVodeInit", 1, my_pe)) return(1);
/* Call CVodeSStolerances to specify the scalar relative tolerance
* and scalar absolute tolerances */
flag = CVodeSStolerances(cvode_mem, reltol, abstol);
if (check_flag(&flag, "CVodeSStolerances", 1, my_pe)) return(1);
if (my_pe == 0) PrintIntro(npes);
umax = N_VMaxNorm(u);
if (my_pe == 0) {
t = T0;
PrintData(t, umax, 0);
}
/* In loop over output points, call CVode, print results, test for error */
for (iout=1, tout=T1; iout <= NOUT; iout++, tout += DTOUT) {
flag = CVode(cvode_mem, tout, u, &t, CV_NORMAL);
if(check_flag(&flag, "CVode", 1, my_pe)) break;
umax = N_VMaxNorm(u);
flag = CVodeGetNumSteps(cvode_mem, &nst);
check_flag(&flag, "CVodeGetNumSteps", 1, my_pe);
if (my_pe == 0) PrintData(t, umax, nst);
}
if (my_pe == 0)
PrintFinalStats(cvode_mem); /* Print some final statistics */
N_VDestroy(u); /* Free hypre vector wrapper */
HYPRE_IJVectorDestroy(Uij); /* Free the underlying hypre vector */
CVodeFree(&cvode_mem); /* Free the integrator memory */
free(data); /* Free user data */
MPI_Finalize();
return(0);
}
/***************************** Main Program ******************************/
int main(int argc, char *argv[])
{
realtype abstol, reltol, t, tout;
N_Vector u;
UserData data;
void *arkode_mem;
int iout, flag;
MPI_Comm comm;
HYPRE_Int local_N, npes, my_pe;
HYPRE_ParVector Upar; /* Declare HYPRE parallel vector */
HYPRE_IJVector Uij; /* Declare "IJ" interface to HYPRE vector */
u = NULL;
data = NULL;
arkode_mem = NULL;
/* Set problem size neq */
/* neq = NVARS*MX*MY; */
/* Get processor number and total number of pe's */
MPI_Init(&argc, &argv);
comm = MPI_COMM_WORLD;
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &my_pe);
if (npes != NPEX*NPEY) {
if (my_pe == 0)
fprintf(stderr, "\nMPI_ERROR(0): npes = %d is not equal to NPEX*NPEY = %d\n\n",
npes,NPEX*NPEY);
MPI_Finalize();
return(1);
}
/* Set local length */
local_N = NVARS*MXSUB*MYSUB;
/* Allocate hypre vector */
HYPRE_IJVectorCreate(comm, my_pe*local_N, (my_pe + 1)*local_N - 1, &Uij);
HYPRE_IJVectorSetObjectType(Uij, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(Uij);
/* Allocate and load user data block; allocate preconditioner block */
data = (UserData) malloc(sizeof *data);
if (check_flag((void *)data, "malloc", 2, my_pe)) MPI_Abort(comm, 1);
InitUserData(my_pe, comm, data);
/* Set initial values and allocate u */
SetInitialProfiles(Uij, data, local_N, my_pe*local_N);
HYPRE_IJVectorAssemble(Uij);
HYPRE_IJVectorGetObject(Uij, (void**) &Upar);
u = N_VMake_ParHyp(Upar); /* Create wrapper u around hypre vector */
if (check_flag((void *)u, "N_VNew", 0, my_pe)) MPI_Abort(comm, 1);
/* Set tolerances */
abstol = ATOL; reltol = RTOL;
/* Call ARKodeCreate to create the solver memory */
arkode_mem = ARKodeCreate();
if (check_flag((void *)arkode_mem, "ARKodeCreate", 0, my_pe)) MPI_Abort(comm, 1);
/* Set the pointer to user-defined data */
flag = ARKodeSetUserData(arkode_mem, data);
if (check_flag(&flag, "ARKodeSetUserData", 1, my_pe)) MPI_Abort(comm, 1);
/* Call ARKodeInit to initialize the integrator memory and specify the
user's right hand side functions in u'=fe(t,u)+fi(t,u) [here fe is NULL],
the inital time T0, and the initial dependent variable vector u. */
flag = ARKodeInit(arkode_mem, NULL, f, T0, u);
if(check_flag(&flag, "ARKodeInit", 1, my_pe)) return(1);
/* Call ARKodeSetMaxNumSteps to increase default */
flag = ARKodeSetMaxNumSteps(arkode_mem, 1000000);
if (check_flag(&flag, "ARKodeSetMaxNumSteps", 1, my_pe)) return(1);
/* Call ARKodeSStolerances to specify the scalar relative tolerance
and scalar absolute tolerances */
flag = ARKodeSStolerances(arkode_mem, reltol, abstol);
if (check_flag(&flag, "ARKodeSStolerances", 1, my_pe)) return(1);
/* Call ARKSpgmr to specify the linear solver ARKSPGMR
with left preconditioning and the default Krylov dimension maxl */
flag = ARKSpgmr(arkode_mem, PREC_LEFT, 0);
if (check_flag(&flag, "ARKSpgmr", 1, my_pe)) MPI_Abort(comm, 1);
/* Set preconditioner setup and solve routines Precond and PSolve,
and the pointer to the user-defined block data */
flag = ARKSpilsSetPreconditioner(arkode_mem, Precond, PSolve);
if (check_flag(&flag, "ARKSpilsSetPreconditioner", 1, my_pe)) MPI_Abort(comm, 1);
if (my_pe == 0)
printf("\n2-species diurnal advection-diffusion problem\n\n");
/* In loop over output points, call ARKode, print results, test for error */
for (iout=1, tout=TWOHR; iout<=NOUT; iout++, tout+=TWOHR) {
flag = ARKode(arkode_mem, tout, u, &t, ARK_NORMAL);
if (check_flag(&flag, "ARKode", 1, my_pe)) break;
PrintOutput(arkode_mem, my_pe, comm, u, t);
}
/* Print final statistics */
if (my_pe == 0) PrintFinalStats(arkode_mem);
/* Free memory */
N_VDestroy(u); /* Free hypre vector wrapper */
HYPRE_IJVectorDestroy(Uij); /* Free the underlying hypre vector */
FreeUserData(data);
ARKodeFree(&arkode_mem);
MPI_Finalize();
return(0);
}
HYPRE_Int main (HYPRE_Int argc, char *argv[])
{
HYPRE_Int i;
HYPRE_Int myid, num_procs;
HYPRE_Int N, n;
HYPRE_Int ilower, iupper;
HYPRE_Int local_size, extra;
HYPRE_Int solver_id;
HYPRE_Int print_solution;
double h, h2;
#ifdef HYPRE_FORTRAN
hypre_F90_Obj A;
hypre_F90_Obj parcsr_A;
hypre_F90_Obj b;
hypre_F90_Obj par_b;
hypre_F90_Obj x;
hypre_F90_Obj par_x;
hypre_F90_Obj solver, precond;
hypre_F90_Obj long_temp_COMM;
HYPRE_Int temp_COMM;
HYPRE_Int precond_id;
HYPRE_Int one = 1;
HYPRE_Int two = 2;
HYPRE_Int three = 3;
HYPRE_Int six = 6;
HYPRE_Int twenty = 20;
HYPRE_Int thousand = 1000;
HYPRE_Int hypre_type = HYPRE_PARCSR;
double oo1 = 1.e-3;
double tol = 1.e-7;
#else
HYPRE_IJMatrix A;
HYPRE_ParCSRMatrix parcsr_A;
HYPRE_IJVector b;
HYPRE_ParVector par_b;
HYPRE_IJVector x;
HYPRE_ParVector par_x;
HYPRE_Solver solver, precond;
#endif
/* Initialize MPI */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs);
/* Default problem parameters */
n = 33;
solver_id = 0;
print_solution = 0;
/* Parse command line */
{
HYPRE_Int arg_index = 0;
HYPRE_Int print_usage = 0;
while (arg_index < argc)
{
if ( strcmp(argv[arg_index], "-n") == 0 )
{
arg_index++;
n = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-solver") == 0 )
{
arg_index++;
solver_id = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-print_solution") == 0 )
{
arg_index++;
print_solution = 1;
}
else if ( strcmp(argv[arg_index], "-help") == 0 )
{
print_usage = 1;
break;
}
else
{
arg_index++;
}
}
if ((print_usage) && (myid == 0))
{
hypre_printf("\n");
hypre_printf("Usage: %s [<options>]\n", argv[0]);
hypre_printf("\n");
hypre_printf(" -n <n> : problem size in each direction (default: 33)\n");
hypre_printf(" -solver <ID> : solver ID\n");
hypre_printf(" 0 - AMG (default) \n");
hypre_printf(" 1 - AMG-PCG\n");
hypre_printf(" 8 - ParaSails-PCG\n");
hypre_printf(" 50 - PCG\n");
hypre_printf(" -print_solution : print the solution vector\n");
hypre_printf("\n");
}
if (print_usage)
{
hypre_MPI_Finalize();
return (0);
}
}
/* Preliminaries: want at least one processor per row */
if (n*n < num_procs) n = sqrt(num_procs) + 1;
N = n*n; /* global number of rows */
h = 1.0/(n+1); /* mesh size*/
h2 = h*h;
/* Each processor knows only of its own rows - the range is denoted by ilower
and upper. Here we partition the rows. We account for the fact that
N may not divide evenly by the number of processors. */
local_size = N/num_procs;
extra = N - local_size*num_procs;
ilower = local_size*myid;
ilower += hypre_min(myid, extra);
iupper = local_size*(myid+1);
iupper += hypre_min(myid+1, extra);
iupper = iupper - 1;
/* How many rows do I have? */
local_size = iupper - ilower + 1;
/* Create the matrix.
Note that this is a square matrix, so we indicate the row partition
size twice (since number of rows = number of cols) */
#ifdef HYPRE_FORTRAN
long_temp_COMM = (hypre_F90_Obj) hypre_MPI_COMM_WORLD;
temp_COMM = (HYPRE_Int) hypre_MPI_COMM_WORLD;
HYPRE_IJMatrixCreate(&long_temp_COMM, &ilower, &iupper, &ilower, &iupper, &A);
#else
HYPRE_IJMatrixCreate(hypre_MPI_COMM_WORLD, ilower, iupper, ilower, iupper, &A);
#endif
/* Choose a parallel csr format storage (see the User's Manual) */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixSetObjectType(&A, &hypre_type);
#else
HYPRE_IJMatrixSetObjectType(A, HYPRE_PARCSR);
#endif
/* Initialize before setting coefficients */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixInitialize(&A);
#else
HYPRE_IJMatrixInitialize(A);
#endif
/* Now go through my local rows and set the matrix entries.
Each row has at most 5 entries. For example, if n=3:
A = [M -I 0; -I M -I; 0 -I M]
M = [4 -1 0; -1 4 -1; 0 -1 4]
Note that here we are setting one row at a time, though
one could set all the rows together (see the User's Manual).
*/
{
HYPRE_Int nnz;
double values[5];
HYPRE_Int cols[5];
for (i = ilower; i <= iupper; i++)
{
nnz = 0;
/* The left identity block:position i-n */
if ((i-n)>=0)
{
cols[nnz] = i-n;
values[nnz] = -1.0;
nnz++;
}
/* The left -1: position i-1 */
if (i%n)
{
cols[nnz] = i-1;
values[nnz] = -1.0;
nnz++;
}
/* Set the diagonal: position i */
cols[nnz] = i;
values[nnz] = 4.0;
nnz++;
/* The right -1: position i+1 */
if ((i+1)%n)
{
cols[nnz] = i+1;
values[nnz] = -1.0;
nnz++;
}
/* The right identity block:position i+n */
if ((i+n)< N)
{
cols[nnz] = i+n;
values[nnz] = -1.0;
nnz++;
}
/* Set the values for row i */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixSetValues(&A, &one, &nnz, &i, &cols[0], &values[0]);
#else
HYPRE_IJMatrixSetValues(A, 1, &nnz, &i, cols, values);
#endif
}
}
/* Assemble after setting the coefficients */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixAssemble(&A);
#else
HYPRE_IJMatrixAssemble(A);
#endif
/* Get the parcsr matrix object to use */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixGetObject(&A, &parcsr_A);
HYPRE_IJMatrixGetObject(&A, &parcsr_A);
#else
HYPRE_IJMatrixGetObject(A, (void**) &parcsr_A);
HYPRE_IJMatrixGetObject(A, (void**) &parcsr_A);
#endif
/* Create the rhs and solution */
#ifdef HYPRE_FORTRAN
HYPRE_IJVectorCreate(&temp_COMM, &ilower, &iupper, &b);
HYPRE_IJVectorSetObjectType(&b, &hypre_type);
HYPRE_IJVectorInitialize(&b);
#else
HYPRE_IJVectorCreate(hypre_MPI_COMM_WORLD, ilower, iupper,&b);
HYPRE_IJVectorSetObjectType(b, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(b);
#endif
#ifdef HYPRE_FORTRAN
HYPRE_IJVectorCreate(&temp_COMM, &ilower, &iupper, &x);
HYPRE_IJVectorSetObjectType(&x, &hypre_type);
HYPRE_IJVectorInitialize(&x);
#else
HYPRE_IJVectorCreate(hypre_MPI_COMM_WORLD, ilower, iupper,&x);
HYPRE_IJVectorSetObjectType(x, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(x);
#endif
/* Set the rhs values to h^2 and the solution to zero */
{
double *rhs_values, *x_values;
HYPRE_Int *rows;
rhs_values = calloc(local_size, sizeof(double));
x_values = calloc(local_size, sizeof(double));
rows = calloc(local_size, sizeof(HYPRE_Int));
for (i=0; i<local_size; i++)
{
rhs_values[i] = h2;
x_values[i] = 0.0;
rows[i] = ilower + i;
}
#ifdef HYPRE_FORTRAN
HYPRE_IJVectorSetValues(&b, &local_size, &rows[0], &rhs_values[0]);
HYPRE_IJVectorSetValues(&x, &local_size, &rows[0], &x_values[0]);
#else
HYPRE_IJVectorSetValues(b, local_size, rows, rhs_values);
HYPRE_IJVectorSetValues(x, local_size, rows, x_values);
#endif
free(x_values);
free(rhs_values);
free(rows);
}
#ifdef HYPRE_FORTRAN
HYPRE_IJVectorAssemble(&b);
HYPRE_IJVectorGetObject(&b, &par_b);
#else
HYPRE_IJVectorAssemble(b);
HYPRE_IJVectorGetObject(b, (void **) &par_b);
#endif
#ifdef HYPRE_FORTRAN
HYPRE_IJVectorAssemble(&x);
HYPRE_IJVectorGetObject(&x, &par_x);
#else
HYPRE_IJVectorAssemble(x);
HYPRE_IJVectorGetObject(x, (void **) &par_x);
#endif
/* Choose a solver and solve the system */
/* AMG */
if (solver_id == 0)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGCreate(&solver);
#else
HYPRE_BoomerAMGCreate(&solver);
#endif
/* Set some parameters (See Reference Manual for more parameters) */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGSetPrintLevel(&solver, &three); /* print solve info + parameters */
HYPRE_BoomerAMGSetCoarsenType(&solver, &six); /* Falgout coarsening */
HYPRE_BoomerAMGSetRelaxType(&solver, &three); /* G-S/Jacobi hybrid relaxation */
HYPRE_BoomerAMGSetNumSweeps(&solver, &one); /* Sweeeps on each level */
HYPRE_BoomerAMGSetMaxLevels(&solver, &twenty); /* maximum number of levels */
HYPRE_BoomerAMGSetTol(&solver, &tol); /* conv. tolerance */
#else
HYPRE_BoomerAMGSetPrintLevel(solver, 3); /* print solve info + parameters */
HYPRE_BoomerAMGSetCoarsenType(solver, 6); /* Falgout coarsening */
HYPRE_BoomerAMGSetRelaxType(solver, 3); /* G-S/Jacobi hybrid relaxation */
HYPRE_BoomerAMGSetNumSweeps(solver, 1); /* Sweeeps on each level */
HYPRE_BoomerAMGSetMaxLevels(solver, 20); /* maximum number of levels */
HYPRE_BoomerAMGSetTol(solver, 1e-7); /* conv. tolerance */
#endif
/* Now setup and solve! */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGSetup(&solver, &parcsr_A, &par_b, &par_x);
HYPRE_BoomerAMGSolve(&solver, &parcsr_A, &par_b, &par_x);
#else
HYPRE_BoomerAMGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_BoomerAMGSolve(solver, parcsr_A, par_b, par_x);
#endif
/* Run info - needed logging turned on */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGGetNumIterations(&solver, &num_iterations);
HYPRE_BoomerAMGGetFinalRelativeResidualNorm(&solver, &final_res_norm);
#else
HYPRE_BoomerAMGGetNumIterations(solver, &num_iterations);
HYPRE_BoomerAMGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#endif
if (myid == 0)
{
hypre_printf("\n");
hypre_printf("Iterations = %d\n", num_iterations);
hypre_printf("Final Relative Residual Norm = %e\n", final_res_norm);
hypre_printf("\n");
}
/* Destroy solver */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGDestroy(&solver);
#else
HYPRE_BoomerAMGDestroy(solver);
#endif
}
/* PCG */
else if (solver_id == 50)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGCreate(&temp_COMM, &solver);
#else
HYPRE_ParCSRPCGCreate(hypre_MPI_COMM_WORLD, &solver);
#endif
/* Set some parameters (See Reference Manual for more parameters) */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetMaxIter(&solver, &thousand); /* max iterations */
HYPRE_ParCSRPCGSetTol(&solver, &tol); /* conv. tolerance */
HYPRE_ParCSRPCGSetTwoNorm(&solver, &one); /* use the two norm as the stopping criteria */
HYPRE_ParCSRPCGSetPrintLevel(&solver, &two); /* prints out the iteration info */
#else
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* prints out the iteration info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
#endif
/* Now setup and solve! */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetup(&solver, &parcsr_A, &par_b, &par_x);
HYPRE_ParCSRPCGSolve(&solver, &parcsr_A, &par_b, &par_x);
#else
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
#endif
/* Run info - needed logging turned on */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGGetNumIterations(&solver, &num_iterations);
HYPRE_ParCSRPCGGetFinalRelativeResidualNorm(&solver, &final_res_norm);
#else
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#endif
if (myid == 0)
{
hypre_printf("\n");
hypre_printf("Iterations = %d\n", num_iterations);
hypre_printf("Final Relative Residual Norm = %e\n", final_res_norm);
hypre_printf("\n");
}
/* Destroy solver */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGDestroy(&solver);
#else
HYPRE_ParCSRPCGDestroy(solver);
#endif
}
/* PCG with AMG preconditioner */
else if (solver_id == 1)
{
HYPRE_Int num_iterations;
double final_res_norm;
/* Create solver */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGCreate(&temp_COMM, &solver);
#else
HYPRE_ParCSRPCGCreate(hypre_MPI_COMM_WORLD, &solver);
#endif
/* Set some parameters (See Reference Manual for more parameters) */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetMaxIter(&solver, &thousand); /* max iterations */
HYPRE_ParCSRPCGSetTol(&solver, &tol); /* conv. tolerance */
HYPRE_ParCSRPCGSetTwoNorm(&solver, &one); /* use the two norm as the stopping criteria */
HYPRE_ParCSRPCGSetPrintLevel(&solver, &two); /* print solve info */
#else
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* print solve info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
#endif
/* Now set up the AMG preconditioner and specify any parameters */
#ifdef HYPRE_FORTRAN
HYPRE_BoomerAMGCreate(&precond);
HYPRE_BoomerAMGSetPrintLevel(&precond, &one); /* print amg solution info*/
HYPRE_BoomerAMGSetCoarsenType(&precond, &six);
HYPRE_BoomerAMGSetRelaxType(&precond, &three);
HYPRE_BoomerAMGSetNumSweeps(&precond, &one);
HYPRE_BoomerAMGSetTol(&precond, &oo1);
#else
HYPRE_BoomerAMGCreate(&precond);
HYPRE_BoomerAMGSetPrintLevel(precond, 1); /* print amg solution info*/
HYPRE_BoomerAMGSetCoarsenType(precond, 6);
HYPRE_BoomerAMGSetRelaxType(precond, 3);
HYPRE_BoomerAMGSetNumSweeps(precond, 1);
HYPRE_BoomerAMGSetTol(precond, 1e-3);
#endif
/* Set the PCG preconditioner */
#ifdef HYPRE_FORTRAN
precond_id = 2;
HYPRE_ParCSRPCGSetPrecond(&solver, &precond_id, &precond);
#else
HYPRE_PCGSetPrecond(solver, (HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSolve,
(HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSetup, precond);
#endif
/* Now setup and solve! */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetup(&solver, &parcsr_A, &par_b, &par_x);
HYPRE_ParCSRPCGSolve(&solver, &parcsr_A, &par_b, &par_x);
#else
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
#endif
/* Run info - needed logging turned on */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGGetNumIterations(&solver, &num_iterations);
HYPRE_ParCSRPCGGetFinalRelativeResidualNorm(&solver, &final_res_norm);
#else
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#endif
if (myid == 0)
{
hypre_printf("\n");
hypre_printf("Iterations = %d\n", num_iterations);
hypre_printf("Final Relative Residual Norm = %e\n", final_res_norm);
hypre_printf("\n");
}
/* Destroy solver and preconditioner */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGDestroy(&solver);
HYPRE_BoomerAMGDestroy(&precond);
#else
HYPRE_ParCSRPCGDestroy(solver);
HYPRE_BoomerAMGDestroy(precond);
#endif
}
/* PCG with Parasails Preconditioner */
else if (solver_id == 8)
{
HYPRE_Int num_iterations;
double final_res_norm;
HYPRE_Int sai_max_levels = 1;
double sai_threshold = 0.1;
double sai_filter = 0.05;
HYPRE_Int sai_sym = 1;
/* Create solver */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGCreate(&temp_COMM, &solver);
#else
HYPRE_ParCSRPCGCreate(hypre_MPI_COMM_WORLD, &solver);
#endif
/* Set some parameters (See Reference Manual for more parameters) */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetMaxIter(&solver, &thousand); /* max iterations */
HYPRE_ParCSRPCGSetTol(&solver, &tol); /* conv. tolerance */
HYPRE_ParCSRPCGSetTwoNorm(&solver, &one); /* use the two norm as the stopping criteria */
HYPRE_ParCSRPCGSetPrintLevel(&solver, &two); /* print solve info */
#else
HYPRE_PCGSetMaxIter(solver, 1000); /* max iterations */
HYPRE_PCGSetTol(solver, 1e-7); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 1); /* use the two norm as the stopping criteria */
HYPRE_PCGSetPrintLevel(solver, 2); /* print solve info */
HYPRE_PCGSetLogging(solver, 1); /* needed to get run info later */
#endif
/* Now set up the ParaSails preconditioner and specify any parameters */
#ifdef HYPRE_FORTRAN
HYPRE_ParaSailsCreate(&temp_COMM, &precond);
#else
HYPRE_ParaSailsCreate(hypre_MPI_COMM_WORLD, &precond);
#endif
/* Set some parameters (See Reference Manual for more parameters) */
#ifdef HYPRE_FORTRAN
HYPRE_ParaSailsSetParams(&precond, &sai_threshold, &sai_max_levels);
HYPRE_ParaSailsSetFilter(&precond, &sai_filter);
HYPRE_ParaSailsSetSym(&precond, &sai_sym);
HYPRE_ParaSailsSetLogging(&precond, &three);
#else
HYPRE_ParaSailsSetParams(precond, sai_threshold, sai_max_levels);
HYPRE_ParaSailsSetFilter(precond, sai_filter);
HYPRE_ParaSailsSetSym(precond, sai_sym);
HYPRE_ParaSailsSetLogging(precond, 3);
#endif
/* Set the PCG preconditioner */
#ifdef HYPRE_FORTRAN
precond_id = 4;
HYPRE_ParCSRPCGSetPrecond(&solver, &precond_id, &precond);
#else
HYPRE_PCGSetPrecond(solver, (HYPRE_PtrToSolverFcn) HYPRE_ParaSailsSolve,
(HYPRE_PtrToSolverFcn) HYPRE_ParaSailsSetup, precond);
#endif
/* Now setup and solve! */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGSetup(&solver, &parcsr_A, &par_b, &par_x);
HYPRE_ParCSRPCGSolve(&solver, &parcsr_A, &par_b, &par_x);
#else
HYPRE_ParCSRPCGSetup(solver, parcsr_A, par_b, par_x);
HYPRE_ParCSRPCGSolve(solver, parcsr_A, par_b, par_x);
#endif
/* Run info - needed logging turned on */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGGetNumIterations(&solver, &num_iterations);
HYPRE_ParCSRPCGGetFinalRelativeResidualNorm(&solver, &final_res_norm);
#else
HYPRE_PCGGetNumIterations(solver, &num_iterations);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#endif
if (myid == 0)
{
hypre_printf("\n");
hypre_printf("Iterations = %d\n", num_iterations);
hypre_printf("Final Relative Residual Norm = %e\n", final_res_norm);
hypre_printf("\n");
}
/* Destory solver and preconditioner */
#ifdef HYPRE_FORTRAN
HYPRE_ParCSRPCGDestroy(&solver);
HYPRE_ParaSailsDestroy(&precond);
#else
HYPRE_ParCSRPCGDestroy(solver);
HYPRE_ParaSailsDestroy(precond);
#endif
}
else
{
if (myid ==0) hypre_printf("Invalid solver id specified.\n");
}
/* Print the solution */
#ifdef HYPRE_FORTRAN
if (print_solution)
HYPRE_IJVectorPrint(&x, "ij.out.x");
#else
if (print_solution)
HYPRE_IJVectorPrint(x, "ij.out.x");
#endif
/* Clean up */
#ifdef HYPRE_FORTRAN
HYPRE_IJMatrixDestroy(&A);
HYPRE_IJVectorDestroy(&b);
HYPRE_IJVectorDestroy(&x);
#else
HYPRE_IJMatrixDestroy(A);
HYPRE_IJVectorDestroy(b);
HYPRE_IJVectorDestroy(x);
#endif
/* Finalize MPI*/
hypre_MPI_Finalize();
return(0);
}
int
main(int argc, char *argv[])
{
GRID *g;
DOF *u_h;
MAT *A, *A0, *B;
MAP *map;
INT i;
size_t nnz, mem, mem_peak;
VEC *x, *y0, *y1, *y2;
double t0, t1, dnz, dnz1, mflops, mop;
char *fn = "../test/cube.dat";
FLOAT mem_max = 300;
INT refine = 0;
phgOptionsRegisterFilename("-mesh_file", "Mesh file", (char **)&fn);
phgOptionsRegisterInt("-loop_count", "Loop count", &loop_count);
phgOptionsRegisterInt("-refine", "Refinement level", &refine);
phgOptionsRegisterFloat("-mem_max", "Maximum memory", &mem_max);
phgInit(&argc, &argv);
g = phgNewGrid(-1);
if (!phgImport(g, fn, FALSE))
phgError(1, "can't read file \"%s\".\n", fn);
phgRefineAllElements(g, refine);
u_h = phgDofNew(g, DOF_DEFAULT, 1, "u_h", DofNoAction);
while (TRUE) {
phgPrintf("\n");
if (phgBalanceGrid(g, 1.2, 1, NULL, 0.))
phgPrintf("Repartition mesh, %d submeshes, load imbalance: %lg\n",
g->nprocs, (double)g->lif);
map = phgMapCreate(u_h, NULL);
A = phgMapCreateMat(map, map);
A->handle_bdry_eqns = TRUE;
build_matrix(A, u_h);
phgMatAssemble(A);
/* Note: A is unsymmetric (A' != A) if boundary entries not removed */
phgMatRemoveBoundaryEntries(A);
#if 0
/* test block matrix operation */
A0 = phgMatCreateBlockMatrix(g->comm, 1, 1, &A, NULL);
#else
A0 = A;
#endif
phgPrintf("%d DOF, %d elems, %d submeshes, matrix size: %d, LIF: %lg\n",
DofGetDataCountGlobal(u_h), g->nleaf_global,
g->nprocs, A->rmap->nglobal, (double)g->lif);
/* test PHG mat-vec multiply */
x = phgMapCreateVec(A->cmap, 1);
y1 = phgMapCreateVec(A->rmap, 1);
phgVecRandomize(x, 123);
phgMatVec(MAT_OP_N, 1.0, A0, x, 0.0, &y1);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
phgMatVec(MAT_OP_N, 1.0, A0, x, 0.0, &y1);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
y0 = phgVecCopy(y1, NULL);
nnz = A->nnz_d + A->nnz_o;
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
mop = loop_count * (dnz + dnz - A->rmap->nlocal) * 1e-6;
phgPrintf("\n");
t1 -= t0;
phgPrintf(" PHG: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF)\n",
t1, dnz, mop / (t1 == 0 ? 1. : t1), mflops);
/* test trans(A)*x */
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
phgMatVec(MAT_OP_T, 1.0, A0, x, 0.0, &y1);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1 == 0 ? 1. : t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
/* time A * trans(A) */
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
B = phgMatMat(MAT_OP_N, MAT_OP_N, 1.0, A, A, 0.0, NULL);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
nnz = B->nnz_d + B->nnz_o;
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
/* compare B*x <--> A*A*x */
y2 = phgMatVec(MAT_OP_N, 1.0, B, x, 0.0, NULL);
phgMatVec(MAT_OP_N, 1.0, A0, y0, 0.0, &y1);
phgMatDestroy(&B);
t1 -= t0;
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
#if USE_PETSC
{
Mat ma, mb;
MatInfo info;
Vec va, vb, vc;
PetscScalar *vec;
ma = phgPetscCreateMatAIJ(A);
MatGetVecs(ma, PETSC_NULL, &va);
VecDuplicate(va, &vb);
VecGetArray(va, &vec);
memcpy(vec, x->data, x->map->nlocal * sizeof(*vec));
VecRestoreArray(va, &vec);
MatMult(ma, va, vb);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
MatMult(ma, va, vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
MatGetInfo(ma, MAT_GLOBAL_SUM, &info);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
t1 -= t0;
dnz = info.nz_used;
phgPrintf(" PETSc: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
MatMultTranspose(ma, va, vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
MatMatMult(ma, ma, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &mb);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
t1 -= t0;
MatGetInfo(mb, MAT_GLOBAL_SUM, &info);
dnz = info.nz_used;
VecDuplicate(va, &vc);
/* compare B*x <--> A*A*x */
MatMult(ma, vb, vc);
MatMult(mb, va, vb);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
VecGetArray(vc, &vec);
memcpy(y2->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vc, &vec);
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
phgPetscMatDestroy(&mb);
phgPetscMatDestroy(&ma);
phgPetscVecDestroy(&va);
phgPetscVecDestroy(&vb);
phgPetscVecDestroy(&vc);
}
#endif /* USE_PETSC */
#if USE_HYPRE
{
HYPRE_IJMatrix ma;
HYPRE_IJVector va, vb, vc;
HYPRE_ParCSRMatrix par_ma;
hypre_ParCSRMatrix *par_mb;
HYPRE_ParVector par_va, par_vb, par_vc;
HYPRE_Int offset, *ni, start, end;
assert(sizeof(INT)==sizeof(int) && sizeof(FLOAT)==sizeof(double));
setup_hypre_mat(A, &ma);
ni = phgAlloc(2 * A->rmap->nlocal * sizeof(*ni));
offset = A->cmap->partition[A->cmap->rank];
for (i = 0; i < A->rmap->nlocal; i++)
ni[i] = i + offset;
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&va);
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&vb);
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&vc);
HYPRE_IJVectorSetObjectType(va, HYPRE_PARCSR);
HYPRE_IJVectorSetObjectType(vb, HYPRE_PARCSR);
HYPRE_IJVectorSetObjectType(vc, HYPRE_PARCSR);
HYPRE_IJVectorSetMaxOffProcElmts(va, 0);
HYPRE_IJVectorSetMaxOffProcElmts(vb, 0);
HYPRE_IJVectorSetMaxOffProcElmts(vc, 0);
HYPRE_IJVectorInitialize(va);
HYPRE_IJVectorInitialize(vb);
HYPRE_IJVectorInitialize(vc);
HYPRE_IJMatrixGetObject(ma, (void **)(void *)&par_ma);
HYPRE_IJVectorGetObject(va, (void **)(void *)&par_va);
HYPRE_IJVectorGetObject(vb, (void **)(void *)&par_vb);
HYPRE_IJVectorGetObject(vc, (void **)(void *)&par_vc);
HYPRE_IJVectorSetValues(va, A->cmap->nlocal, ni, (double *)x->data);
HYPRE_IJVectorAssemble(va);
HYPRE_IJVectorAssemble(vb);
HYPRE_IJVectorAssemble(vc);
HYPRE_IJMatrixGetRowCounts(ma, A->cmap->nlocal,
ni, ni + A->rmap->nlocal);
for (i = 0, nnz = 0; i < A->rmap->nlocal; i++)
nnz += ni[A->rmap->nlocal + i];
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_va, 0.0, par_vb);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_va, 0.0, par_vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
t1 -= t0;
phgPrintf(" HYPRE: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
HYPRE_ParCSRMatrixMatvecT(1.0, par_ma, par_va, 0.0, par_vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
/* Note: 'HYPRE_ParCSRMatrix' is currently typedef'ed to
* 'hypre_ParCSRMatrix *' */
par_mb = hypre_ParMatmul((hypre_ParCSRMatrix *)par_ma,
(hypre_ParCSRMatrix *)par_ma);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
start = hypre_ParCSRMatrixFirstRowIndex(par_mb);
end = hypre_ParCSRMatrixLastRowIndex(par_mb) + 1;
for (i = start, nnz = 0; i < end; i++) {
HYPRE_Int ncols;
hypre_ParCSRMatrixGetRow(par_mb, i, &ncols, NULL, NULL);
hypre_ParCSRMatrixRestoreRow(par_mb, i, &ncols, NULL, NULL);
nnz += ncols;
}
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
/* compare B*x <--> A*A*x */
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_vb, 0.0, par_vc);
HYPRE_ParCSRMatrixMatvec(1.0, (void *)par_mb, par_va, 0.0, par_vb);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
HYPRE_IJVectorGetValues(vc, A->rmap->nlocal, ni, (double*)y2->data);
hypre_ParCSRMatrixDestroy((par_mb));
t1 -= t0;
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
phgFree(ni);
HYPRE_IJMatrixDestroy(ma);
HYPRE_IJVectorDestroy(va);
HYPRE_IJVectorDestroy(vb);
HYPRE_IJVectorDestroy(vc);
}
#endif /* USE_HYPRE */
if (A0 != A)
phgMatDestroy(&A0);
#if 0
if (A->rmap->nglobal > 1000) {
VEC *v = phgMapCreateVec(A->rmap, 3);
for (i = 0; i < v->map->nlocal; i++) {
v->data[i + 0 * v->map->nlocal] = 1 * (i + v->map->partition[g->rank]);
v->data[i + 1 * v->map->nlocal] = 2 * (i + v->map->partition[g->rank]);
v->data[i + 2 * v->map->nlocal] = 3 * (i + v->map->partition[g->rank]);
}
phgMatDumpMATLAB(A, "A", "A.m");
phgVecDumpMATLAB(v, "v", "v.m");
phgFinalize();
exit(0);
}
#endif
phgMatDestroy(&A);
phgVecDestroy(&x);
phgVecDestroy(&y0);
phgVecDestroy(&y1);
phgVecDestroy(&y2);
phgMapDestroy(&map);
mem = phgMemoryUsage(g, &mem_peak);
dnz = mem / (1024.0 * 1024.0);
dnz1 = mem_peak / (1024.0 * 1024.0);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
phgPrintf(" Memory: current %0.4lgMB, peak %0.4lgMB\n", dnz, dnz1);
#if 0
{
static int loop_count = 0;
if (++loop_count == 4)
break;
}
#endif
if (mem_peak > 1024 * (size_t)1024 * mem_max)
break;
phgRefineAllElements(g, 1);
}
phgDofFree(&u_h);
phgFreeGrid(&g);
phgFinalize();
return 0;
}
