double *solve(double *Ab, int solver_id, struct parms parms)
{
int i, j;
double final_res_norm;
int time_index, n_pre, n_post, num_iterations;
n_pre = 1; n_post = 1;
double *A_val, *b_val;
A_val = (double *) calloc(parms.N*parms.nsten, sizeof(double));
b_val = (double *) calloc(parms.N, sizeof(double));
for (i = 0; i < (parms.N*parms.nsten); i++){
A_val[i] = Ab[i];
}
for (i = 0; i < parms.N; i++){
b_val[i] = Ab[i+parms.N*parms.nsten];
}
// HYPRE //
HYPRE_StructGrid grid;
HYPRE_StructStencil stencil;
HYPRE_StructMatrix A;
HYPRE_StructVector b;
HYPRE_StructVector x;
HYPRE_StructSolver solver;
HYPRE_StructSolver precond;
#if Dim == 2
HYPRE_Int ilower[2] = {parms.x0, parms.y0};
HYPRE_Int iupper[2] = {parms.x1, parms.y1};
#endif
#if Dim == 3
HYPRE_Int ilower[3] = {parms.x0, parms.y0, 0};
HYPRE_Int iupper[3] = {parms.x1, parms.y1, parms.Nz-1};
#endif
{
// Create an empty 2D grid object
HYPRE_StructGridCreate(MPI_COMM_WORLD, Dim, &grid);
// Add a new box to the grid
HYPRE_StructGridSetExtents(grid, ilower, iupper);
// 1. Set up periodic boundary condition in y-direction and create the grid
int pset[3];
pset[0] = 0; pset[1] = parms.Ny; pset[2] = 0;
#if Dim == 3
pset[2] = parms.Nz;
#endif
//HYPRE_StructGridSetNumGhost(grid,pset)
HYPRE_StructGridSetPeriodic(grid, pset);
HYPRE_StructGridAssemble(grid);
}
// 2. Define the discretization stencil
{
if (Dim == 2){
// Create an empty 2D, 5-pt stencil object
HYPRE_StructStencilCreate(2, parms.nsten, &stencil);
// Define the geometry of the stencil
{
int offsets[5][2] = {{0,0}, {-1,0}, {0,-1}, {0,1}, {1,0}};
for (i = 0; i < parms.nsten; i++)
HYPRE_StructStencilSetElement(stencil, i, offsets[i]);
}
}
else
{
HYPRE_StructStencilCreate(3, parms.nsten, &stencil);
// Define the geometry of the 3D stencil
{
int offsets[7][3] = {{0,0,0}, {-1,0,0}, {0,-1,0}, {0,1,0}, {1,0,0}, {0,0,-1}, {0,0,1}};
for (i = 0; i < parms.nsten; i++)
HYPRE_StructStencilSetElement(stencil, i, offsets[i]);
}
}
}
// 3. Set up a Struct Matrix A from Aval
{
HYPRE_Int stencil_indices[parms.nsten];
// Create an empty matrix object
HYPRE_StructMatrixCreate(MPI_COMM_WORLD, grid, stencil, &A);
// Indicate that the matrix coefficients are ready to be set
HYPRE_StructMatrixInitialize(A);
for (j = 0; j < parms.nsten; j++)
stencil_indices[j] = j;
HYPRE_StructMatrixSetBoxValues(A, ilower, iupper, parms.nsten, stencil_indices, A_val);
free(A_val);
}
// 4. Set up Struct Vectors for b from b_val and set x = 0
{
double *values;
HYPRE_StructVectorCreate(MPI_COMM_WORLD, grid, &b);
HYPRE_StructVectorCreate(MPI_COMM_WORLD, grid, &x);
HYPRE_StructVectorInitialize(b);
HYPRE_StructVectorInitialize(x);
values = calloc((parms.N), sizeof(double));
for (i = 0; i < (parms.N); i++)
values[i] = 0.0;
HYPRE_StructVectorSetBoxValues(x, ilower, iupper, values);
HYPRE_StructVectorSetBoxValues(b, ilower, iupper, b_val);
free(b_val);
free(values);
}
//Finalize the vector and matrix assembly
HYPRE_StructMatrixAssemble(A);
HYPRE_StructVectorAssemble(b);
HYPRE_StructVectorAssemble(x);
#if DEBUG == 3
HYPRE_StructMatrixPrint("./poisson.matrix", A, 0);
HYPRE_StructVectorPrint("./poisson.rhs", b, 0);
/*char fname[64];
char Aname[64], bname[64];
sprintf(Aname,"data/A%d.",parms.cyc);
sprintf(bname,"data/b%d.",parms.cyc);
filename(fname, Aname, parms.wkdir, parms);
HYPRE_StructMatrixPrint(fname, A, 0);
filename(fname, bname, parms.wkdir, parms);
HYPRE_StructVectorPrint(fname, b, 0);*/
#endif
// 6. Set up and use a solver (SMG)
if (solver_id == 0)
{
time_index = hypre_InitializeTiming("SMG Setup");
hypre_BeginTiming(time_index);
HYPRE_StructSMGCreate(MPI_COMM_WORLD, &solver);
HYPRE_StructSMGSetMemoryUse(solver, 0);
HYPRE_StructSMGSetMaxIter(solver, 100);
HYPRE_StructSMGSetTol(solver, 1.0e-12);
HYPRE_StructSMGSetRelChange(solver, 0);
HYPRE_StructSMGSetNumPreRelax(solver, n_pre);
HYPRE_StructSMGSetNumPostRelax(solver, n_post);
// Logging must be on to get iterations and residual norm info below
HYPRE_StructSMGSetLogging(solver, 1);
// Setup and print setup timings
HYPRE_StructSMGSetup(solver, A, b, x);
hypre_EndTiming(time_index);
#if DEBUG == 3
hypre_PrintTiming("Setup phase times", MPI_COMM_WORLD);
#endif
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
// Solve and print solve timings
time_index = hypre_InitializeTiming("SMG Solve");
hypre_BeginTiming(time_index);
HYPRE_StructSMGSolve(solver, A, b, x);
hypre_EndTiming(time_index);
#if DEBUG == 3
hypre_PrintTiming("Solve phase times", MPI_COMM_WORLD);
#endif
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
// Get some info on the run
HYPRE_StructSMGGetNumIterations(solver, &num_iterations);
HYPRE_StructSMGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#if DEBUG == 2
if (parms.rank == 0){
fprintf(stdout, "Number of Iterations = %4d ; Final Relative Residual Norm = %e\n\n", num_iterations, final_res_norm);
}
#endif
// Clean up
HYPRE_StructSMGDestroy(solver);
}
// 6. Set up and use a solver (PCG) with SMG Preconditioner
if (solver_id == 1)
{
HYPRE_StructPCGCreate(MPI_COMM_WORLD, &solver);
//HYPRE_StructPCGSetMemoryUse(solver, 0);
HYPRE_StructPCGSetMaxIter(solver, 100);
HYPRE_StructPCGSetTol(solver, 1.0e-12);
HYPRE_StructPCGSetTwoNorm(solver, 1);
HYPRE_StructPCGSetRelChange(solver, 0);
//HYPRE_StructPCGSetPrintLevel(solver, 2 ); /* print each CG iteration */
HYPRE_StructPCGSetLogging(solver, 1);
/* Use symmetric SMG as preconditioner */
HYPRE_StructSMGCreate(MPI_COMM_WORLD, &precond);
HYPRE_StructSMGSetMemoryUse(precond, 0);
HYPRE_StructSMGSetMaxIter(precond, 32);
HYPRE_StructSMGSetTol(precond, 0.0);
HYPRE_StructSMGSetZeroGuess(precond);
HYPRE_StructSMGSetNumPreRelax(precond, 1);
HYPRE_StructSMGSetNumPostRelax(precond, 1);
/* Set the preconditioner and solve */
HYPRE_StructPCGSetPrecond(solver, HYPRE_StructSMGSolve, HYPRE_StructSMGSetup, precond);
HYPRE_StructPCGSetup(solver, A, b, x);
HYPRE_StructPCGSolve(solver, A, b, x);
/* Get some info on the run */
HYPRE_StructPCGGetNumIterations(solver, &num_iterations);
HYPRE_StructPCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
#if DEBUG == 2
if (parms.rank == 0){
fprintf(stdout, "Number of Iterations = %4d ; Final Relative Residual Norm = %e\n\n", num_iterations, final_res_norm);
}
#endif
/* Clean up */
HYPRE_StructSMGDestroy(precond);
HYPRE_StructPCGDestroy(solver);
}
// get the local solution
double *values = calloc(parms.N, sizeof(double));
HYPRE_StructVectorGetBoxValues(x, ilower, iupper, values);
// Free memory
HYPRE_StructGridDestroy(grid);
HYPRE_StructStencilDestroy(stencil);
HYPRE_StructMatrixDestroy(A);
HYPRE_StructVectorDestroy(b);
HYPRE_StructVectorDestroy(x);
free(Ab);
return(values);
}
hypre_int
main( hypre_int argc,
char *argv[] )
{
HYPRE_Int arg_index;
HYPRE_Int print_usage;
HYPRE_Int nx, ny, nz;
HYPRE_Int P, Q, R;
HYPRE_Int bx, by, bz;
HYPRE_StructGrid from_grid, to_grid;
HYPRE_StructVector from_vector, to_vector, check_vector;
HYPRE_CommPkg comm_pkg;
HYPRE_Int time_index;
HYPRE_Int num_procs, myid;
HYPRE_Int p, q, r;
HYPRE_Int dim;
HYPRE_Int nblocks ;
HYPRE_Int **ilower, **iupper, **iupper2;
HYPRE_Int istart[3];
HYPRE_Int i, ix, iy, iz, ib;
HYPRE_Int print_system = 0;
HYPRE_Real check;
/*-----------------------------------------------------------
* Initialize some stuff
*-----------------------------------------------------------*/
/* Initialize MPI */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs );
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
/*-----------------------------------------------------------
* Set defaults
*-----------------------------------------------------------*/
dim = 3;
nx = 2;
ny = 2;
nz = 2;
P = num_procs;
Q = 1;
R = 1;
bx = 1;
by = 1;
bz = 1;
istart[0] = 1;
istart[1] = 1;
istart[2] = 1;
/*-----------------------------------------------------------
* Parse command line
*-----------------------------------------------------------*/
print_usage = 0;
arg_index = 1;
while (arg_index < argc)
{
if ( strcmp(argv[arg_index], "-n") == 0 )
{
arg_index++;
nx = atoi(argv[arg_index++]);
ny = atoi(argv[arg_index++]);
nz = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-istart") == 0 )
{
arg_index++;
istart[0] = atoi(argv[arg_index++]);
istart[1] = atoi(argv[arg_index++]);
istart[2] = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-P") == 0 )
{
arg_index++;
P = atoi(argv[arg_index++]);
Q = atoi(argv[arg_index++]);
R = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-b") == 0 )
{
arg_index++;
bx = atoi(argv[arg_index++]);
by = atoi(argv[arg_index++]);
bz = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-d") == 0 )
{
arg_index++;
dim = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-print") == 0 )
{
arg_index++;
print_system = 1;
}
else if ( strcmp(argv[arg_index], "-help") == 0 )
{
print_usage = 1;
break;
}
else
{
arg_index++;
}
}
/*-----------------------------------------------------------
* Print usage info
*-----------------------------------------------------------*/
if ( (print_usage) && (myid == 0) )
{
hypre_printf("\n");
hypre_printf("Usage: %s [<options>]\n", argv[0]);
hypre_printf("\n");
hypre_printf(" -n <nx> <ny> <nz> : problem size per block\n");
hypre_printf(" -istart <ix> <iy> <iz> : start of box\n");
hypre_printf(" -P <Px> <Py> <Pz> : processor topology\n");
hypre_printf(" -b <bx> <by> <bz> : blocking per processor\n");
hypre_printf(" -d <dim> : problem dimension (2 or 3)\n");
hypre_printf(" -print : print vectors\n");
hypre_printf("\n");
}
if ( print_usage )
{
exit(1);
}
/*-----------------------------------------------------------
* Check a few things
*-----------------------------------------------------------*/
if ((P*Q*R) > num_procs)
{
if (myid == 0)
{
hypre_printf("Error: PxQxR is more than the number of processors\n");
}
exit(1);
}
else if ((P*Q*R) < num_procs)
{
if (myid == 0)
{
hypre_printf("Warning: PxQxR is less than the number of processors\n");
}
}
/*-----------------------------------------------------------
* Print driver parameters
*-----------------------------------------------------------*/
if (myid == 0)
{
hypre_printf("Running with these driver parameters:\n");
hypre_printf(" (nx, ny, nz) = (%d, %d, %d)\n", nx, ny, nz);
hypre_printf(" (ix, iy, iz) = (%d, %d, %d)\n",
istart[0],istart[1],istart[2]);
hypre_printf(" (Px, Py, Pz) = (%d, %d, %d)\n", P, Q, R);
hypre_printf(" (bx, by, bz) = (%d, %d, %d)\n", bx, by, bz);
hypre_printf(" dim = %d\n", dim);
}
/*-----------------------------------------------------------
* Set up the stencil structure (7 points) when matrix is NOT read from file
* Set up the grid structure used when NO files are read
*-----------------------------------------------------------*/
switch (dim)
{
case 1:
nblocks = bx;
p = myid % P;
break;
case 2:
nblocks = bx*by;
p = myid % P;
q = (( myid - p)/P) % Q;
break;
case 3:
nblocks = bx*by*bz;
p = myid % P;
q = (( myid - p)/P) % Q;
r = ( myid - p - P*q)/( P*Q );
break;
}
if (myid >= (P*Q*R))
{
/* My processor has no data on it */
nblocks = bx = by = bz = 0;
}
/*-----------------------------------------------------------
* prepare space for the extents
*-----------------------------------------------------------*/
ilower = hypre_CTAlloc(HYPRE_Int*, nblocks);
iupper = hypre_CTAlloc(HYPRE_Int*, nblocks);
iupper2 = hypre_CTAlloc(HYPRE_Int*, nblocks);
for (i = 0; i < nblocks; i++)
{
ilower[i] = hypre_CTAlloc(HYPRE_Int, dim);
iupper[i] = hypre_CTAlloc(HYPRE_Int, dim);
iupper2[i] = hypre_CTAlloc(HYPRE_Int, dim);
}
ib = 0;
switch (dim)
{
case 1:
for (ix = 0; ix < bx; ix++)
{
ilower[ib][0] = istart[0]+ nx*(bx*p+ix);
iupper[ib][0] = istart[0]+ nx*(bx*p+ix+1) - 1;
iupper2[ib][0] = iupper[ib][0];
if ( (ix == (bx-1)) && (p < (P-1)) )
iupper2[ib][0] = iupper[ib][0] + 1;
ib++;
}
break;
case 2:
for (iy = 0; iy < by; iy++)
for (ix = 0; ix < bx; ix++)
{
ilower[ib][0] = istart[0]+ nx*(bx*p+ix);
iupper[ib][0] = istart[0]+ nx*(bx*p+ix+1) - 1;
ilower[ib][1] = istart[1]+ ny*(by*q+iy);
iupper[ib][1] = istart[1]+ ny*(by*q+iy+1) - 1;
iupper2[ib][0] = iupper[ib][0];
iupper2[ib][1] = iupper[ib][1];
if ( (ix == (bx-1)) && (p < (P-1)) )
iupper2[ib][0] = iupper[ib][0] + 1;
if ( (iy == (by-1)) && (q < (Q-1)) )
iupper2[ib][1] = iupper[ib][1] + 1;
ib++;
}
break;
case 3:
for (iz = 0; iz < bz; iz++)
for (iy = 0; iy < by; iy++)
for (ix = 0; ix < bx; ix++)
{
ilower[ib][0] = istart[0]+ nx*(bx*p+ix);
iupper[ib][0] = istart[0]+ nx*(bx*p+ix+1) - 1;
ilower[ib][1] = istart[1]+ ny*(by*q+iy);
iupper[ib][1] = istart[1]+ ny*(by*q+iy+1) - 1;
ilower[ib][2] = istart[2]+ nz*(bz*r+iz);
iupper[ib][2] = istart[2]+ nz*(bz*r+iz+1) - 1;
iupper2[ib][0] = iupper[ib][0];
iupper2[ib][1] = iupper[ib][1];
iupper2[ib][2] = iupper[ib][2];
if ( (ix == (bx-1)) && (p < (P-1)) )
iupper2[ib][0] = iupper[ib][0] + 1;
if ( (iy == (by-1)) && (q < (Q-1)) )
iupper2[ib][1] = iupper[ib][1] + 1;
if ( (iz == (bz-1)) && (r < (R-1)) )
iupper2[ib][2] = iupper[ib][2] + 1;
ib++;
}
break;
}
HYPRE_StructGridCreate(hypre_MPI_COMM_WORLD, dim, &from_grid);
HYPRE_StructGridCreate(hypre_MPI_COMM_WORLD, dim, &to_grid);
for (ib = 0; ib < nblocks; ib++)
{
HYPRE_StructGridSetExtents(from_grid, ilower[ib], iupper[ib]);
HYPRE_StructGridSetExtents(to_grid, ilower[ib], iupper2[ib]);
}
HYPRE_StructGridAssemble(from_grid);
HYPRE_StructGridAssemble(to_grid);
/*-----------------------------------------------------------
* Set up the vectors
*-----------------------------------------------------------*/
HYPRE_StructVectorCreate(hypre_MPI_COMM_WORLD, from_grid, &from_vector);
HYPRE_StructVectorInitialize(from_vector);
AddValuesVector(from_grid, from_vector, 1.0);
HYPRE_StructVectorAssemble(from_vector);
HYPRE_StructVectorCreate(hypre_MPI_COMM_WORLD, to_grid, &to_vector);
HYPRE_StructVectorInitialize(to_vector);
AddValuesVector(to_grid, to_vector, 0.0);
HYPRE_StructVectorAssemble(to_vector);
/* Vector used to check the migration */
HYPRE_StructVectorCreate(hypre_MPI_COMM_WORLD, to_grid, &check_vector);
HYPRE_StructVectorInitialize(check_vector);
AddValuesVector(to_grid, check_vector, 1.0);
HYPRE_StructVectorAssemble(check_vector);
/*-----------------------------------------------------------
* Migrate
*-----------------------------------------------------------*/
time_index = hypre_InitializeTiming("Struct Migrate");
hypre_BeginTiming(time_index);
HYPRE_StructVectorGetMigrateCommPkg(from_vector, to_vector, &comm_pkg);
HYPRE_StructVectorMigrate(comm_pkg, from_vector, to_vector);
HYPRE_CommPkgDestroy(comm_pkg);
hypre_EndTiming(time_index);
hypre_PrintTiming("Struct Migrate", hypre_MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
/*-----------------------------------------------------------
* Check the migration and print the result
*-----------------------------------------------------------*/
hypre_StructAxpy(-1.0, to_vector, check_vector);
check = hypre_StructInnerProd (check_vector, check_vector);
if (myid == 0)
{
printf("\nCheck = %1.0f (success = 0)\n\n", check);
}
/*-----------------------------------------------------------
* Print out the vectors
*-----------------------------------------------------------*/
if (print_system)
{
HYPRE_StructVectorPrint("struct_migrate.out.xfr", from_vector, 0);
HYPRE_StructVectorPrint("struct_migrate.out.xto", to_vector, 0);
}
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
HYPRE_StructGridDestroy(from_grid);
HYPRE_StructGridDestroy(to_grid);
for (i = 0; i < nblocks; i++)
{
hypre_TFree(ilower[i]);
hypre_TFree(iupper[i]);
hypre_TFree(iupper2[i]);
}
hypre_TFree(ilower);
hypre_TFree(iupper);
hypre_TFree(iupper2);
HYPRE_StructVectorDestroy(from_vector);
HYPRE_StructVectorDestroy(to_vector);
HYPRE_StructVectorDestroy(check_vector);
/* Finalize MPI */
hypre_MPI_Finalize();
return (0);
}
HYPRE_Int main( HYPRE_Int argc, char *argv[] )
{
HYPRE_Int arg_index;
HYPRE_Int print_usage;
HYPRE_Int build_matrix_arg_index;
HYPRE_Int solver_id;
HYPRE_Int ierr,i,j;
HYPRE_Int num_iterations;
HYPRE_ParCSRMatrix parcsr_A;
HYPRE_Int num_procs, myid;
HYPRE_Int local_row;
HYPRE_Int time_index;
MPI_Comm comm;
HYPRE_Int M, N;
HYPRE_Int first_local_row, last_local_row;
HYPRE_Int first_local_col, last_local_col;
HYPRE_Int size, *col_ind;
HYPRE_Real *values;
/* parameters for BoomerAMG */
HYPRE_Real strong_threshold;
HYPRE_Int num_grid_sweeps;
HYPRE_Real relax_weight;
/* parameters for GMRES */
HYPRE_Int k_dim;
char *paramString = new char[100];
/*-----------------------------------------------------------
* Initialize some stuff
*-----------------------------------------------------------*/
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs );
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
/*-----------------------------------------------------------
* Set defaults
*-----------------------------------------------------------*/
build_matrix_arg_index = argc;
solver_id = 0;
strong_threshold = 0.25;
num_grid_sweeps = 2;
relax_weight = 0.5;
k_dim = 20;
/*-----------------------------------------------------------
* Parse command line
*-----------------------------------------------------------*/
print_usage = 0;
arg_index = 1;
while ( (arg_index < argc) && (!print_usage) )
{
if ( strcmp(argv[arg_index], "-solver") == 0 )
{
arg_index++;
solver_id = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-dbg") == 0 )
{
arg_index++;
atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-help") == 0 )
{
print_usage = 1;
}
else
{
arg_index++;
}
}
/*-----------------------------------------------------------
* Print usage info
*-----------------------------------------------------------*/
if ( (print_usage) && (myid == 0) )
{
hypre_printf("\n");
hypre_printf("Usage: %s [<options>]\n", argv[0]);
hypre_printf("\n");
hypre_printf(" -solver <ID> : solver ID\n");
hypre_printf(" 0=DS-PCG 1=ParaSails-PCG \n");
hypre_printf(" 2=AMG-PCG 3=DS-GMRES \n");
hypre_printf(" 4=PILUT-GMRES 5=AMG-GMRES \n");
hypre_printf("\n");
hypre_printf(" -rlx <val> : relaxation type\n");
hypre_printf(" 0=Weighted Jacobi \n");
hypre_printf(" 1=Gauss-Seidel (very slow!) \n");
hypre_printf(" 3=Hybrid Jacobi/Gauss-Seidel \n");
hypre_printf("\n");
exit(1);
}
/*-----------------------------------------------------------
* Print driver parameters
*-----------------------------------------------------------*/
if (myid == 0)
{
hypre_printf("Running with these driver parameters:\n");
hypre_printf(" solver ID = %d\n", solver_id);
}
/*-----------------------------------------------------------
* Set up matrix
*-----------------------------------------------------------*/
strcpy(paramString, "LS Interface");
time_index = hypre_InitializeTiming(paramString);
hypre_BeginTiming(time_index);
BuildParLaplacian27pt(argc, argv, build_matrix_arg_index, &parcsr_A);
/*-----------------------------------------------------------
* Copy the parcsr matrix into the LSI through interface calls
*-----------------------------------------------------------*/
ierr = HYPRE_ParCSRMatrixGetComm( parcsr_A, &comm );
ierr += HYPRE_ParCSRMatrixGetDims( parcsr_A, &M, &N );
ierr = HYPRE_ParCSRMatrixGetLocalRange( parcsr_A,
&first_local_row, &last_local_row ,
&first_local_col, &last_local_col );
HYPRE_LinSysCore H(hypre_MPI_COMM_WORLD);
HYPRE_Int numLocalEqns = last_local_row - first_local_row + 1;
H.createMatricesAndVectors(M,first_local_row+1,numLocalEqns);
HYPRE_Int index;
HYPRE_Int *rowLengths = new HYPRE_Int[numLocalEqns];
HYPRE_Int **colIndices = new HYPRE_Int*[numLocalEqns];
local_row = 0;
for (i=first_local_row; i<= last_local_row; i++)
{
ierr += HYPRE_ParCSRMatrixGetRow(parcsr_A,i,&size,&col_ind,&values );
rowLengths[local_row] = size;
colIndices[local_row] = new HYPRE_Int[size];
for (j=0; j<size; j++) colIndices[local_row][j] = col_ind[j] + 1;
local_row++;
HYPRE_ParCSRMatrixRestoreRow(parcsr_A,i,&size,&col_ind,&values);
}
H.allocateMatrix(colIndices, rowLengths);
delete [] rowLengths;
for (i=0; i< numLocalEqns; i++) delete [] colIndices[i];
delete [] colIndices;
HYPRE_Int *newColInd;
for (i=first_local_row; i<= last_local_row; i++)
{
ierr += HYPRE_ParCSRMatrixGetRow(parcsr_A,i,&size,&col_ind,&values );
newColInd = new HYPRE_Int[size];
for (j=0; j<size; j++) newColInd[j] = col_ind[j] + 1;
H.sumIntoSystemMatrix(i+1,size,(const HYPRE_Real*)values,
(const HYPRE_Int*)newColInd);
delete [] newColInd;
ierr += HYPRE_ParCSRMatrixRestoreRow(parcsr_A,i,&size,&col_ind,&values);
}
H.matrixLoadComplete();
HYPRE_ParCSRMatrixDestroy(parcsr_A);
/*-----------------------------------------------------------
* Set up the RHS and initial guess
*-----------------------------------------------------------*/
HYPRE_Real ddata=1.0;
HYPRE_Int status;
for (i=first_local_row; i<= last_local_row; i++)
{
index = i + 1;
H.sumIntoRHSVector(1,(const HYPRE_Real*) &ddata, (const HYPRE_Int*) &index);
}
hypre_EndTiming(time_index);
strcpy(paramString, "LS Interface");
hypre_PrintTiming(paramString, hypre_MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
/*-----------------------------------------------------------
* Solve the system using PCG
*-----------------------------------------------------------*/
if ( solver_id == 0 )
{
strcpy(paramString, "solver cg");
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: DS-PCG\n");
strcpy(paramString, "preconditioner diagonal");
H.parameters(1, ¶mString);
}
else if ( solver_id == 1 )
{
strcpy(paramString, "solver cg");
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: ParaSails-PCG\n");
strcpy(paramString, "preconditioner parasails");
H.parameters(1, ¶mString);
strcpy(paramString, "parasailsNlevels 1");
H.parameters(1, ¶mString);
strcpy(paramString, "parasailsThreshold 0.1");
H.parameters(1, ¶mString);
}
else if ( solver_id == 2 )
{
strcpy(paramString, "solver cg");
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: AMG-PCG\n");
strcpy(paramString, "preconditioner boomeramg");
H.parameters(1, ¶mString);
strcpy(paramString, "amgCoarsenType falgout");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgStrongThreshold %e", strong_threshold);
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgNumSweeps %d", num_grid_sweeps);
H.parameters(1, ¶mString);
strcpy(paramString, "amgRelaxType jacobi");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgRelaxWeight %e", relax_weight);
H.parameters(1, ¶mString);
}
else if ( solver_id == 3 )
{
strcpy(paramString, "solver cg");
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: Poly-PCG\n");
strcpy(paramString, "preconditioner poly");
H.parameters(1, ¶mString);
strcpy(paramString, "polyOrder 9");
H.parameters(1, ¶mString);
}
else if ( solver_id == 4 )
{
strcpy(paramString, "solver gmres");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "gmresDim %d", k_dim);
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: DS-GMRES\n");
strcpy(paramString, "preconditioner diagonal");
H.parameters(1, ¶mString);
}
else if ( solver_id == 5 )
{
strcpy(paramString, "solver gmres");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "gmresDim %d", k_dim);
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: PILUT-GMRES\n");
strcpy(paramString, "preconditioner pilut");
H.parameters(1, ¶mString);
strcpy(paramString, "pilutRowSize 0");
H.parameters(1, ¶mString);
strcpy(paramString, "pilutDropTol 0.0");
H.parameters(1, ¶mString);
}
else if ( solver_id == 6 )
{
strcpy(paramString, "solver gmres");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "gmresDim %d", k_dim);
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: AMG-GMRES\n");
strcpy(paramString, "preconditioner boomeramg");
H.parameters(1, ¶mString);
strcpy(paramString, "amgCoarsenType falgout");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgStrongThreshold %e", strong_threshold);
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgNumSweeps %d", num_grid_sweeps);
H.parameters(1, ¶mString);
strcpy(paramString, "amgRelaxType jacobi");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "amgRelaxWeight %e", relax_weight);
H.parameters(1, ¶mString);
}
else if ( solver_id == 7 )
{
strcpy(paramString, "solver gmres");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "gmresDim %d", k_dim);
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: DDILUT-GMRES\n");
strcpy(paramString, "preconditioner ddilut");
H.parameters(1, ¶mString);
strcpy(paramString, "ddilutFillin 5.0");
H.parameters(1, ¶mString);
strcpy(paramString, "ddilutDropTol 0.0");
H.parameters(1, ¶mString);
}
else if ( solver_id == 8 )
{
strcpy(paramString, "solver gmres");
H.parameters(1, ¶mString);
hypre_sprintf(paramString, "gmresDim %d", k_dim);
H.parameters(1, ¶mString);
if (myid == 0) hypre_printf("Solver: POLY-GMRES\n");
strcpy(paramString, "preconditioner poly");
H.parameters(1, ¶mString);
strcpy(paramString, "polyOrder 5");
H.parameters(1, ¶mString);
}
strcpy(paramString, "Krylov Solve");
time_index = hypre_InitializeTiming(paramString);
hypre_BeginTiming(time_index);
H.launchSolver(status, num_iterations);
hypre_EndTiming(time_index);
strcpy(paramString, "Solve phase times");
hypre_PrintTiming(paramString, hypre_MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
if (myid == 0)
{
hypre_printf("\n Iterations = %d\n", num_iterations);
hypre_printf("\n");
}
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
delete [] paramString;
hypre_MPI_Finalize();
return (0);
}
int main (int argc, char *argv[])
{
int myid, num_procs;
int n, N, pi, pj, pk;
double h;
double tol, theta;
int maxit, cycle_type;
int rlx_type, rlx_sweeps, rlx_weight, rlx_omega;
int amg_coarsen_type, amg_agg_levels, amg_rlx_type;
int amg_interp_type, amg_Pmax;
int singular_problem ;
HYPRE_Int time_index;
HYPRE_SStructGrid edge_grid;
HYPRE_SStructGraph A_graph;
HYPRE_SStructMatrix A;
HYPRE_SStructVector b;
HYPRE_SStructVector x;
HYPRE_SStructGrid node_grid;
HYPRE_SStructGraph G_graph;
HYPRE_SStructStencil G_stencil[3];
HYPRE_SStructMatrix G;
HYPRE_SStructVector xcoord, ycoord, zcoord;
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);
/* Set default parameters */
n = 10;
optionAlpha = 0;
optionBeta = 0;
maxit = 100;
tol = 1e-6;
cycle_type = 13;
rlx_type = 2;
rlx_sweeps = 1;
rlx_weight = 1.0;
rlx_omega = 1.0;
amg_coarsen_type = 10;
amg_agg_levels = 1;
amg_rlx_type = 6;
theta = 0.25;
amg_interp_type = 6;
amg_Pmax = 4;
singular_problem = 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], "-a") == 0 )
{
arg_index++;
optionAlpha = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-b") == 0 )
{
arg_index++;
optionBeta = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-maxit") == 0 )
{
arg_index++;
maxit = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-tol") == 0 )
{
arg_index++;
tol = atof(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-type") == 0 )
{
arg_index++;
cycle_type = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-rlx") == 0 )
{
arg_index++;
rlx_type = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-rlxn") == 0 )
{
arg_index++;
rlx_sweeps = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-rlxw") == 0 )
{
arg_index++;
rlx_weight = atof(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-rlxo") == 0 )
{
arg_index++;
rlx_omega = atof(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-ctype") == 0 )
{
arg_index++;
amg_coarsen_type = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-amgrlx") == 0 )
{
arg_index++;
amg_rlx_type = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-agg") == 0 )
{
arg_index++;
amg_agg_levels = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-itype") == 0 )
{
arg_index++;
amg_interp_type = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-pmax") == 0 )
{
arg_index++;
amg_Pmax = atoi(argv[arg_index++]);
}
else if ( strcmp(argv[arg_index], "-sing") == 0 )
{
arg_index++;
singular_problem = 1;
}
else if ( strcmp(argv[arg_index], "-theta") == 0 )
{
arg_index++;
theta = atof(argv[arg_index++]);
}
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 per processor (default: 10)\n");
printf(" -a <alpha_opt> : choice for the curl-curl coefficient (default: 1)\n");
printf(" -b <beta_opt> : choice for the mass coefficient (default: 1)\n");
printf("\n");
printf("PCG-AMS solver options: \n");
printf(" -maxit <num> : maximum number of iterations (100) \n");
printf(" -tol <num> : convergence tolerance (1e-6) \n");
printf(" -type <num> : 3-level cycle type (0-8, 11-14) \n");
printf(" -theta <num> : BoomerAMG threshold (0.25) \n");
printf(" -ctype <num> : BoomerAMG coarsening type \n");
printf(" -agg <num> : Levels of BoomerAMG agg. coarsening \n");
printf(" -amgrlx <num> : BoomerAMG relaxation type \n");
printf(" -itype <num> : BoomerAMG interpolation type \n");
printf(" -pmax <num> : BoomerAMG interpolation truncation \n");
printf(" -rlx <num> : relaxation type \n");
printf(" -rlxn <num> : number of relaxation sweeps \n");
printf(" -rlxw <num> : damping parameter (usually <=1) \n");
printf(" -rlxo <num> : SOR parameter (usually in (0,2)) \n");
printf(" -sing : curl-curl only (singular) problem \n");
printf("\n");
printf("\n");
}
if (print_usage)
{
MPI_Finalize();
return (0);
}
}
/* Figure out the processor grid (N x N x N). The local problem size is n^3,
while pi, pj and pk indicate the position in the processor grid. */
N = pow(num_procs,1.0/3.0) + 0.5;
if (num_procs != N*N*N)
{
if (myid == 0) printf("Can't run on %d processors, try %d.\n",
num_procs, N*N*N);
MPI_Finalize();
exit(1);
}
h = 1.0 / (N*n);
pk = myid / (N*N);
pj = myid/N - pk*N;
pi = myid - pj*N - pk*N*N;
/* Start timing */
time_index = hypre_InitializeTiming("SStruct Setup");
hypre_BeginTiming(time_index);
/* 1. Set up the edge and nodal grids. Note that we do this simultaneously
to make sure that they have the same extents. For simplicity we use
only one part to represent the unit cube. */
{
HYPRE_Int ndim = 3;
HYPRE_Int nparts = 1;
/* Create empty 2D grid objects */
HYPRE_SStructGridCreate(MPI_COMM_WORLD, ndim, nparts, &node_grid);
HYPRE_SStructGridCreate(MPI_COMM_WORLD, ndim, nparts, &edge_grid);
/* Set the extents of the grid - each processor sets its grid boxes. */
{
HYPRE_Int part = 0;
HYPRE_Int ilower[3] = {1 + pi*n, 1 + pj*n, 1 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructGridSetExtents(node_grid, part, ilower, iupper);
HYPRE_SStructGridSetExtents(edge_grid, part, ilower, iupper);
}
/* Set the variable type and number of variables on each grid. */
{
HYPRE_Int i;
HYPRE_Int nnodevars = 1;
HYPRE_Int nedgevars = 3;
HYPRE_SStructVariable nodevars[1] = {HYPRE_SSTRUCT_VARIABLE_NODE};
HYPRE_SStructVariable edgevars[3] = {HYPRE_SSTRUCT_VARIABLE_XEDGE,
HYPRE_SSTRUCT_VARIABLE_YEDGE,
HYPRE_SSTRUCT_VARIABLE_ZEDGE};
for (i = 0; i < nparts; i++)
{
HYPRE_SStructGridSetVariables(node_grid, i, nnodevars, nodevars);
HYPRE_SStructGridSetVariables(edge_grid, i, nedgevars, edgevars);
}
}
/* Since there is only one part, there is no need to call the
SetNeighborPart or SetSharedPart functions, which determine the spatial
relation between the parts. See Examples 12, 13 and 14 for
illustrations of these calls. */
/* Now the grids are ready to be used */
HYPRE_SStructGridAssemble(node_grid);
HYPRE_SStructGridAssemble(edge_grid);
}
/* 2. Create the finite element stiffness matrix A and load vector b. */
{
HYPRE_Int part = 0; /* this problem has only one part */
/* Set the ordering of the variables in the finite element problem. This
is done by listing the variable offset directions relative to the
element's center. See the Reference Manual for more details. */
{
HYPRE_Int ordering[48] = { 0, 0, -1, -1, /* x-edge [0]-[1] */
1, +1, 0, -1, /* y-edge [1]-[2] */
/* [7]------[6] */ 0, 0, +1, -1, /* x-edge [3]-[2] */
/* /| /| */ 1, -1, 0, -1, /* y-edge [0]-[3] */
/* / | / | */ 0, 0, -1, +1, /* x-edge [4]-[5] */
/* [4]------[5] | */ 1, +1, 0, +1, /* y-edge [5]-[6] */
/* | [3]----|-[2] */ 0, 0, +1, +1, /* x-edge [7]-[6] */
/* | / | / */ 1, -1, 0, +1, /* y-edge [4]-[7] */
/* |/ |/ */ 2, -1, -1, 0, /* z-edge [0]-[4] */
/* [0]------[1] */ 2, +1, -1, 0, /* z-edge [1]-[5] */
2, +1, +1, 0, /* z-edge [2]-[6] */
2, -1, +1, 0 }; /* z-edge [3]-[7] */
HYPRE_SStructGridSetFEMOrdering(edge_grid, part, ordering);
}
/* Set up the Graph - this determines the non-zero structure of the
matrix. */
{
HYPRE_Int part = 0;
/* Create the graph object */
HYPRE_SStructGraphCreate(MPI_COMM_WORLD, edge_grid, &A_graph);
/* See MatrixSetObjectType below */
HYPRE_SStructGraphSetObjectType(A_graph, HYPRE_PARCSR);
/* Indicate that this problem uses finite element stiffness matrices and
load vectors, instead of stencils. */
HYPRE_SStructGraphSetFEM(A_graph, part);
/* The edge finite element matrix is full, so there is no need to call the
HYPRE_SStructGraphSetFEMSparsity() function. */
/* Assemble the graph */
HYPRE_SStructGraphAssemble(A_graph);
}
/* Set up the SStruct Matrix and right-hand side vector */
{
/* Create the matrix object */
HYPRE_SStructMatrixCreate(MPI_COMM_WORLD, A_graph, &A);
/* Use a ParCSR storage */
HYPRE_SStructMatrixSetObjectType(A, HYPRE_PARCSR);
/* Indicate that the matrix coefficients are ready to be set */
HYPRE_SStructMatrixInitialize(A);
/* Create an empty vector object */
HYPRE_SStructVectorCreate(MPI_COMM_WORLD, edge_grid, &b);
/* Use a ParCSR storage */
HYPRE_SStructVectorSetObjectType(b, HYPRE_PARCSR);
/* Indicate that the vector coefficients are ready to be set */
HYPRE_SStructVectorInitialize(b);
}
/* Set the matrix and vector entries by finite element assembly */
{
/* local stiffness matrix and load vector */
double S[12][12], F[12];
int i, j, k;
HYPRE_Int index[3];
for (i = 1; i <= n; i++)
for (j = 1; j <= n; j++)
for (k = 1; k <= n; k++)
{
/* Compute the FEM matrix and r.h.s. for cell (i,j,k) with
coefficients evaluated at the cell center. */
index[0] = i + pi*n; index[1] = j + pj*n; index[2] = k + pk*n;
ComputeFEMND1(S,F,(pi*n+i)*h-h/2,(pj*n+j)*h-h/2,(pk*n+k)*h-h/2,h);
/* Eliminate boundary conditions on x = 0 */
if (index[0] == 1)
{
int ii, jj, bc_edges[4] = { 3, 11, 7, 8 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Eliminate boundary conditions on y = 0 */
if (index[1] == 1)
{
int ii, jj, bc_edges[4] = { 0, 9, 4, 8 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Eliminate boundary conditions on z = 0 */
if (index[2] == 1)
{
int ii, jj, bc_edges[4] = { 0, 1, 2, 3 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Eliminate boundary conditions on x = 1 */
if (index[0] == N*n)
{
int ii, jj, bc_edges[4] = { 1, 10, 5, 9 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Eliminate boundary conditions on y = 1 */
if (index[1] == N*n)
{
int ii, jj, bc_edges[4] = { 2, 10, 6, 11 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Eliminate boundary conditions on z = 1 */
if (index[2] == N*n)
{
int ii, jj, bc_edges[4] = { 4, 5, 6, 7 };
for (ii = 0; ii < 4; ii++)
{
for (jj = 0; jj < 12; jj++)
S[bc_edges[ii]][jj] = S[jj][bc_edges[ii]] = 0.0;
S[bc_edges[ii]][bc_edges[ii]] = 1.0;
F[bc_edges[ii]] = 0.0;
}
}
/* Assemble the matrix */
HYPRE_SStructMatrixAddFEMValues(A, part, index, &S[0][0]);
/* Assemble the vector */
HYPRE_SStructVectorAddFEMValues(b, part, index, F);
}
}
/* Collective calls finalizing the matrix and vector assembly */
HYPRE_SStructMatrixAssemble(A);
HYPRE_SStructVectorAssemble(b);
}
/* 3. Create the discrete gradient matrix G, which is needed in AMS. */
{
HYPRE_Int part = 0;
HYPRE_Int stencil_size = 2;
/* Define the discretization stencil relating the edges and nodes of the
grid. */
{
HYPRE_Int ndim = 3;
HYPRE_Int entry;
HYPRE_Int var = 0; /* the node variable */
/* The discrete gradient stencils connect edge to node variables. */
HYPRE_Int Gx_offsets[2][3] = {{-1,0,0},{0,0,0}}; /* x-edge [7]-[6] */
HYPRE_Int Gy_offsets[2][3] = {{0,-1,0},{0,0,0}}; /* y-edge [5]-[6] */
HYPRE_Int Gz_offsets[2][3] = {{0,0,-1},{0,0,0}}; /* z-edge [2]-[6] */
HYPRE_SStructStencilCreate(ndim, stencil_size, &G_stencil[0]);
HYPRE_SStructStencilCreate(ndim, stencil_size, &G_stencil[1]);
HYPRE_SStructStencilCreate(ndim, stencil_size, &G_stencil[2]);
for (entry = 0; entry < stencil_size; entry++)
{
HYPRE_SStructStencilSetEntry(G_stencil[0], entry, Gx_offsets[entry], var);
HYPRE_SStructStencilSetEntry(G_stencil[1], entry, Gy_offsets[entry], var);
HYPRE_SStructStencilSetEntry(G_stencil[2], entry, Gz_offsets[entry], var);
}
}
/* Set up the Graph - this determines the non-zero structure of the
matrix. */
{
HYPRE_Int nvars = 3;
HYPRE_Int var; /* the edge variables */
/* Create the discrete gradient graph object */
HYPRE_SStructGraphCreate(MPI_COMM_WORLD, edge_grid, &G_graph);
/* See MatrixSetObjectType below */
HYPRE_SStructGraphSetObjectType(G_graph, HYPRE_PARCSR);
/* Since the discrete gradient relates edge and nodal variables (it is a
rectangular matrix), we have to specify the domain (column) grid. */
HYPRE_SStructGraphSetDomainGrid(G_graph, node_grid);
/* Tell the graph which stencil to use for each edge variable on each
part (we only have one part). */
for (var = 0; var < nvars; var++)
HYPRE_SStructGraphSetStencil(G_graph, part, var, G_stencil[var]);
/* Assemble the graph */
HYPRE_SStructGraphAssemble(G_graph);
}
/* Set up the SStruct Matrix */
{
/* Create the matrix object */
HYPRE_SStructMatrixCreate(MPI_COMM_WORLD, G_graph, &G);
/* Use a ParCSR storage */
HYPRE_SStructMatrixSetObjectType(G, HYPRE_PARCSR);
/* Indicate that the matrix coefficients are ready to be set */
HYPRE_SStructMatrixInitialize(G);
}
/* Set the discrete gradient values, assuming a "natural" orientation of
the edges (i.e. one in agreement with the coordinate directions). */
{
int i;
int nedges = n*(n+1)*(n+1);
double *values;
HYPRE_Int stencil_indices[2] = {0,1}; /* the nodes of each edge */
values = calloc(2*nedges, sizeof(double));
/* The edge orientation is fixed: from first to second node */
for (i = 0; i < nedges; i++)
{
values[2*i] = -1.0;
values[2*i+1] = 1.0;
}
/* Set the values in the discrete gradient x-edges */
{
HYPRE_Int var = 0;
HYPRE_Int ilower[3] = {1 + pi*n, 0 + pj*n, 0 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructMatrixSetBoxValues(G, part, ilower, iupper, var,
stencil_size, stencil_indices,
values);
}
/* Set the values in the discrete gradient y-edges */
{
HYPRE_Int var = 1;
HYPRE_Int ilower[3] = {0 + pi*n, 1 + pj*n, 0 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructMatrixSetBoxValues(G, part, ilower, iupper, var,
stencil_size, stencil_indices,
values);
}
/* Set the values in the discrete gradient z-edges */
{
HYPRE_Int var = 2;
HYPRE_Int ilower[3] = {0 + pi*n, 0 + pj*n, 1 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructMatrixSetBoxValues(G, part, ilower, iupper, var,
stencil_size, stencil_indices,
values);
}
free(values);
}
/* Finalize the matrix assembly */
HYPRE_SStructMatrixAssemble(G);
}
/* 4. Create the vectors of nodal coordinates xcoord, ycoord and zcoord,
which are needed in AMS. */
{
int i, j, k;
HYPRE_Int part = 0;
HYPRE_Int var = 0; /* the node variable */
HYPRE_Int index[3];
double xval, yval, zval;
/* Create empty vector objects */
HYPRE_SStructVectorCreate(MPI_COMM_WORLD, node_grid, &xcoord);
HYPRE_SStructVectorCreate(MPI_COMM_WORLD, node_grid, &ycoord);
HYPRE_SStructVectorCreate(MPI_COMM_WORLD, node_grid, &zcoord);
/* Set the object type to ParCSR */
HYPRE_SStructVectorSetObjectType(xcoord, HYPRE_PARCSR);
HYPRE_SStructVectorSetObjectType(ycoord, HYPRE_PARCSR);
HYPRE_SStructVectorSetObjectType(zcoord, HYPRE_PARCSR);
/* Indicate that the vector coefficients are ready to be set */
HYPRE_SStructVectorInitialize(xcoord);
HYPRE_SStructVectorInitialize(ycoord);
HYPRE_SStructVectorInitialize(zcoord);
/* Compute and set the coordinates of the nodes */
for (i = 0; i <= n; i++)
for (j = 0; j <= n; j++)
for (k = 0; k <= n; k++)
{
index[0] = i + pi*n; index[1] = j + pj*n; index[2] = k + pk*n;
xval = index[0]*h;
yval = index[1]*h;
zval = index[2]*h;
HYPRE_SStructVectorSetValues(xcoord, part, index, var, &xval);
HYPRE_SStructVectorSetValues(ycoord, part, index, var, &yval);
HYPRE_SStructVectorSetValues(zcoord, part, index, var, &zval);
}
/* Finalize the vector assembly */
HYPRE_SStructVectorAssemble(xcoord);
HYPRE_SStructVectorAssemble(ycoord);
HYPRE_SStructVectorAssemble(zcoord);
}
/* 5. Set up a SStruct Vector for the solution vector x */
{
HYPRE_Int part = 0;
int nvalues = n*(n+1)*(n+1);
double *values;
values = calloc(nvalues, sizeof(double));
/* Create an empty vector object */
HYPRE_SStructVectorCreate(MPI_COMM_WORLD, edge_grid, &x);
/* Set the object type to ParCSR */
HYPRE_SStructVectorSetObjectType(x, HYPRE_PARCSR);
/* Indicate that the vector coefficients are ready to be set */
HYPRE_SStructVectorInitialize(x);
/* Set the values for the initial guess x-edge */
{
HYPRE_Int var = 0;
HYPRE_Int ilower[3] = {1 + pi*n, 0 + pj*n, 0 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructVectorSetBoxValues(x, part, ilower, iupper, var, values);
}
/* Set the values for the initial guess y-edge */
{
HYPRE_Int var = 1;
HYPRE_Int ilower[3] = {0 + pi*n, 1 + pj*n, 0 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructVectorSetBoxValues(x, part, ilower, iupper, var, values);
}
/* Set the values for the initial guess z-edge */
{
HYPRE_Int var = 2;
HYPRE_Int ilower[3] = {0 + pi*n, 0 + pj*n, 1 + pk*n};
HYPRE_Int iupper[3] = {n + pi*n, n + pj*n, n + pk*n};
HYPRE_SStructVectorSetBoxValues(x, part, ilower, iupper, var, values);
}
free(values);
/* Finalize the vector assembly */
HYPRE_SStructVectorAssemble(x);
}
/* Finalize current timing */
hypre_EndTiming(time_index);
hypre_PrintTiming("SStruct phase times", MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
/* 6. Set up and call the PCG-AMS solver (Solver options can be found in the
Reference Manual.) */
{
double final_res_norm;
HYPRE_Int its;
HYPRE_ParCSRMatrix par_A;
HYPRE_ParVector par_b;
HYPRE_ParVector par_x;
HYPRE_ParCSRMatrix par_G;
HYPRE_ParVector par_xcoord;
HYPRE_ParVector par_ycoord;
HYPRE_ParVector par_zcoord;
/* Extract the ParCSR objects needed in the solver */
HYPRE_SStructMatrixGetObject(A, (void **) &par_A);
HYPRE_SStructVectorGetObject(b, (void **) &par_b);
HYPRE_SStructVectorGetObject(x, (void **) &par_x);
HYPRE_SStructMatrixGetObject(G, (void **) &par_G);
HYPRE_SStructVectorGetObject(xcoord, (void **) &par_xcoord);
HYPRE_SStructVectorGetObject(ycoord, (void **) &par_ycoord);
HYPRE_SStructVectorGetObject(zcoord, (void **) &par_zcoord);
if (myid == 0)
printf("Problem size: %lld\n\n",
hypre_ParCSRMatrixGlobalNumRows((hypre_ParCSRMatrix*)par_A));
/* Start timing */
time_index = hypre_InitializeTiming("AMS Setup");
hypre_BeginTiming(time_index);
/* Create solver */
HYPRE_ParCSRPCGCreate(MPI_COMM_WORLD, &solver);
/* Set some parameters (See Reference Manual for more parameters) */
HYPRE_PCGSetMaxIter(solver, maxit); /* max iterations */
HYPRE_PCGSetTol(solver, tol); /* conv. tolerance */
HYPRE_PCGSetTwoNorm(solver, 0); /* 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 */
/* Create AMS preconditioner */
HYPRE_AMSCreate(&precond);
/* Set AMS parameters */
HYPRE_AMSSetMaxIter(precond, 1);
HYPRE_AMSSetTol(precond, 0.0);
HYPRE_AMSSetCycleType(precond, cycle_type);
HYPRE_AMSSetPrintLevel(precond, 1);
/* Set discrete gradient */
HYPRE_AMSSetDiscreteGradient(precond, par_G);
/* Set vertex coordinates */
HYPRE_AMSSetCoordinateVectors(precond,
par_xcoord, par_ycoord, par_zcoord);
if (singular_problem)
HYPRE_AMSSetBetaPoissonMatrix(precond, NULL);
/* Smoothing and AMG options */
HYPRE_AMSSetSmoothingOptions(precond,
rlx_type, rlx_sweeps,
rlx_weight, rlx_omega);
HYPRE_AMSSetAlphaAMGOptions(precond,
amg_coarsen_type, amg_agg_levels,
amg_rlx_type, theta, amg_interp_type,
amg_Pmax);
HYPRE_AMSSetBetaAMGOptions(precond,
amg_coarsen_type, amg_agg_levels,
amg_rlx_type, theta, amg_interp_type,
amg_Pmax);
/* Set the PCG preconditioner */
HYPRE_PCGSetPrecond(solver,
(HYPRE_PtrToSolverFcn) HYPRE_AMSSolve,
(HYPRE_PtrToSolverFcn) HYPRE_AMSSetup,
precond);
/* Call the setup */
HYPRE_ParCSRPCGSetup(solver, par_A, par_b, par_x);
/* Finalize current timing */
hypre_EndTiming(time_index);
hypre_PrintTiming("Setup phase times", MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
/* Start timing again */
time_index = hypre_InitializeTiming("AMS Solve");
hypre_BeginTiming(time_index);
/* Call the solve */
HYPRE_ParCSRPCGSolve(solver, par_A, par_b, par_x);
/* Finalize current timing */
hypre_EndTiming(time_index);
hypre_PrintTiming("Solve phase times", MPI_COMM_WORLD);
hypre_FinalizeTiming(time_index);
hypre_ClearTiming();
/* Get some info */
HYPRE_PCGGetNumIterations(solver, &its);
HYPRE_PCGGetFinalRelativeResidualNorm(solver, &final_res_norm);
/* Clean up */
HYPRE_AMSDestroy(precond);
HYPRE_ParCSRPCGDestroy(solver);
/* Gather the solution vector */
HYPRE_SStructVectorGather(x);
if (myid == 0)
{
printf("\n");
printf("Iterations = %lld\n", its);
printf("Final Relative Residual Norm = %g\n", final_res_norm);
printf("\n");
}
}
/* Free memory */
HYPRE_SStructGridDestroy(edge_grid);
HYPRE_SStructGraphDestroy(A_graph);
HYPRE_SStructMatrixDestroy(A);
HYPRE_SStructVectorDestroy(b);
HYPRE_SStructVectorDestroy(x);
HYPRE_SStructGridDestroy(node_grid);
HYPRE_SStructGraphDestroy(G_graph);
HYPRE_SStructStencilDestroy(G_stencil[0]);
HYPRE_SStructStencilDestroy(G_stencil[1]);
HYPRE_SStructStencilDestroy(G_stencil[2]);
HYPRE_SStructMatrixDestroy(G);
HYPRE_SStructVectorDestroy(xcoord);
HYPRE_SStructVectorDestroy(ycoord);
HYPRE_SStructVectorDestroy(zcoord);
/* Finalize MPI */
MPI_Finalize();
return 0;
}
