int
HYPRE_IJVectorPrint( HYPRE_IJVector vector,
const char *filename )
{
MPI_Comm comm = hypre_IJVectorComm(vector);
HYPRE_BigInt *partitioning;
HYPRE_BigInt jlower, jupper, j;
double value;
int myid;
char new_filename[255];
FILE *file;
if (!vector)
{
printf("Variable vec is NULL -- HYPRE_IJVectorPrint\n");
hypre_error_in_arg(1);
return hypre_error_flag;
}
MPI_Comm_rank(comm, &myid);
sprintf(new_filename,"%s.%05d", filename, myid);
if ((file = fopen(new_filename, "w")) == NULL)
{
printf("Error: can't open output file %s\n", new_filename);
hypre_error_in_arg(2);
return hypre_error_flag;
}
partitioning = hypre_IJVectorPartitioning(vector);
#ifdef HYPRE_NO_GLOBAL_PARTITION
jlower = partitioning[0];
jupper = partitioning[1] - 1;
#else
jlower = partitioning[myid];
jupper = partitioning[myid+1] - 1;
#endif
#ifdef HYPRE_LONG_LONG
fprintf(file, "%lld %lld\n", jlower, jupper);
#else
fprintf(file, "%d %d\n", jlower, jupper);
#endif
for (j = jlower; j <= jupper; j++)
{
HYPRE_IJVectorGetValues(vector, 1, &j, &value);
#ifdef HYPRE_LONG_LONG
fprintf(file, "%lld %.14e\n", j, value);
#else
fprintf(file, "%d %.14e\n", j, value);
#endif
}
fclose(file);
return hypre_error_flag;
}
void HypreSolver2D::solve(double *x)
{
set_rhs(x);
HYPRE_BoomerAMGSolve(solver, parcsr_A, par_b, par_x);
HYPRE_IJVectorGetValues(hv_x, local_size, &rows[0], x);
HYPRE_BoomerAMGGetNumIterations(solver, &num_iterations);
HYPRE_BoomerAMGGetFinalRelativeResidualNorm(solver, &final_res_norm);
}
int main (int argc, char *argv[])
{
int i;
int myid, num_procs;
int N, n;
int ilower, iupper;
int local_size, extra;
int solver_id;
int vis, 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;
vis = 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], "-vis") == 0 )
{
arg_index++;
vis = 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(" -vis : save the solution for GLVis visualization\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).
*/
{
int nnz;
double values[5];
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;
int *rows;
rhs_values = calloc(local_size, sizeof(double));
x_values = calloc(local_size, sizeof(double));
rows = calloc(local_size, sizeof(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)
{
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 = %d\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)
{
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 = %d\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)
{
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 = %d\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)
{
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 = %d\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)
{
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 = %d\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");
}
/* Save the solution for GLVis visualization, see vis/glvis-ex5.sh */
if (vis)
{
FILE *file;
char filename[255];
int nvalues = local_size;
int *rows = calloc(nvalues, sizeof(int));
double *values = calloc(nvalues, sizeof(double));
for (i = 0; i < nvalues; i++)
rows[i] = ilower + i;
/* get the local solution */
HYPRE_IJVectorGetValues(x, nvalues, rows, values);
sprintf(filename, "%s.%06d", "vis/ex5.sol", myid);
if ((file = fopen(filename, "w")) == NULL)
{
printf("Error: can't open output file %s\n", filename);
MPI_Finalize();
exit(1);
}
/* save solution */
for (i = 0; i < nvalues; i++)
fprintf(file, "%.14e\n", values[i]);
fflush(file);
fclose(file);
free(rows);
free(values);
/* save global finite element mesh */
if (myid == 0)
GLVis_PrintGlobalSquareMesh("vis/ex5.mesh", n-1);
}
/* Clean up */
HYPRE_IJMatrixDestroy(A);
HYPRE_IJVectorDestroy(b);
HYPRE_IJVectorDestroy(x);
/* Finalize MPI*/
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;
}
