HYPRE_Real
hypre_ParKrylovInnerProd( void *x,
void *y )
{
return ( hypre_ParVectorInnerProd( (hypre_ParVector *) x,
(hypre_ParVector *) y ) );
}
HYPRE_Int
hypre_SStructInnerProd( hypre_SStructVector *x,
hypre_SStructVector *y,
double *result_ptr )
{
HYPRE_Int nparts = hypre_SStructVectorNParts(x);
double result;
double presult;
HYPRE_Int part;
HYPRE_Int x_object_type= hypre_SStructVectorObjectType(x);
HYPRE_Int y_object_type= hypre_SStructVectorObjectType(y);
if (x_object_type != y_object_type)
{
hypre_error_in_arg(2);
hypre_error_in_arg(3);
return hypre_error_flag;
}
result = 0.0;
if ( (x_object_type == HYPRE_SSTRUCT) || (x_object_type == HYPRE_STRUCT) )
{
for (part = 0; part < nparts; part++)
{
hypre_SStructPInnerProd(hypre_SStructVectorPVector(x, part),
hypre_SStructVectorPVector(y, part), &presult);
result += presult;
}
}
else if (x_object_type == HYPRE_PARCSR)
{
hypre_ParVector *x_par;
hypre_ParVector *y_par;
hypre_SStructVectorConvert(x, &x_par);
hypre_SStructVectorConvert(y, &y_par);
result= hypre_ParVectorInnerProd(x_par, y_par);
}
*result_ptr = result;
return hypre_error_flag;
}
HYPRE_Int
HYPRE_ParVectorInnerProd( HYPRE_ParVector x,
HYPRE_ParVector y,
HYPRE_Real *prod)
{
if (!x)
{
hypre_error_in_arg(1);
return hypre_error_flag;
}
if (!y)
{
hypre_error_in_arg(2);
return hypre_error_flag;
}
*prod = hypre_ParVectorInnerProd( (hypre_ParVector *) x,
(hypre_ParVector *) y) ;
return hypre_error_flag;
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
hypre_ParVector *vector1;
hypre_ParVector *vector2;
hypre_ParVector *tmp_vector;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_size = 20;
HYPRE_Int local_size;
HYPRE_Int first_index;
HYPRE_Int num_vectors, vecstride, idxstride;
HYPRE_Int i, j;
HYPRE_Int *partitioning;
double prod;
double *data, *data2;
hypre_Vector *vector;
hypre_Vector *local_vector;
hypre_Vector *local_vector2;
/* 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, &my_id );
hypre_printf(" my_id: %d num_procs: %d\n", my_id, num_procs);
partitioning = NULL;
num_vectors = 3;
vector1 = hypre_ParMultiVectorCreate
( hypre_MPI_COMM_WORLD, global_size, partitioning, num_vectors );
partitioning = hypre_ParVectorPartitioning(vector1);
hypre_ParVectorInitialize(vector1);
local_vector = hypre_ParVectorLocalVector(vector1);
data = hypre_VectorData(local_vector);
local_size = hypre_VectorSize(local_vector);
vecstride = hypre_VectorVectorStride(local_vector);
idxstride = hypre_VectorIndexStride(local_vector);
first_index = partitioning[my_id];
hypre_printf("vecstride=%i idxstride=%i local_size=%i num_vectors=%i",
vecstride, idxstride, local_size, num_vectors );
for (j=0; j<num_vectors; ++j )
for (i=0; i < local_size; i++)
data[ j*vecstride + i*idxstride ] = first_index+i + 100*j;
hypre_ParVectorPrint(vector1, "Vector");
local_vector2 = hypre_SeqMultiVectorCreate( global_size, num_vectors );
hypre_SeqVectorInitialize(local_vector2);
data2 = hypre_VectorData(local_vector2);
vecstride = hypre_VectorVectorStride(local_vector2);
idxstride = hypre_VectorIndexStride(local_vector2);
for (j=0; j<num_vectors; ++j )
for (i=0; i < global_size; i++)
data2[ j*vecstride + i*idxstride ] = i + 100*j;
/* partitioning = hypre_CTAlloc(HYPRE_Int,4);
partitioning[0] = 0;
partitioning[1] = 10;
partitioning[2] = 10;
partitioning[3] = 20;
*/
partitioning = hypre_CTAlloc(HYPRE_Int,1+num_procs);
hypre_GeneratePartitioning( global_size, num_procs, &partitioning );
vector2 = hypre_VectorToParVector(hypre_MPI_COMM_WORLD,local_vector2,partitioning);
hypre_ParVectorSetPartitioningOwner(vector2,0);
hypre_ParVectorPrint(vector2, "Convert");
vector = hypre_ParVectorToVectorAll(vector2);
/*-----------------------------------------------------------
* Copy the vector into tmp_vector
*-----------------------------------------------------------*/
/* Read doesn't work for multivectors yet...
tmp_vector = hypre_ParVectorRead(hypre_MPI_COMM_WORLD, "Convert");*/
tmp_vector = hypre_ParMultiVectorCreate
( hypre_MPI_COMM_WORLD, global_size, partitioning, num_vectors );
hypre_ParVectorInitialize( tmp_vector );
hypre_ParVectorCopy( vector2, tmp_vector );
/*
tmp_vector = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
hypre_ParVectorSetPartitioningOwner(tmp_vector,0);
hypre_ParVectorInitialize(tmp_vector);
hypre_ParVectorCopy(vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Copy");
*/
/*-----------------------------------------------------------
* Scale tmp_vector
*-----------------------------------------------------------*/
hypre_ParVectorScale(2.0, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Scale");
/*-----------------------------------------------------------
* Do an Axpy (2*vector - vector) = vector
*-----------------------------------------------------------*/
hypre_ParVectorAxpy(-1.0, vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Axpy");
/*-----------------------------------------------------------
* Do an inner product vector* tmp_vector
*-----------------------------------------------------------*/
prod = hypre_ParVectorInnerProd(vector1, tmp_vector);
hypre_printf (" prod: %8.2f \n", prod);
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
hypre_ParVectorDestroy(vector1);
hypre_ParVectorDestroy(vector2);
hypre_ParVectorDestroy(tmp_vector);
hypre_SeqVectorDestroy(local_vector2);
if (vector) hypre_SeqVectorDestroy(vector);
/* Finalize MPI */
hypre_MPI_Finalize();
return 0;
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
HYPRE_Int num_procs, myid;
HYPRE_Int verbose = 0, build_matrix_type = 1;
HYPRE_Int index, matrix_arg_index, commpkg_flag=3;
HYPRE_Int i, k, ierr=0;
HYPRE_Int row_start, row_end;
HYPRE_Int col_start, col_end, global_num_rows;
HYPRE_Int *row_part, *col_part;
char *csrfilename;
HYPRE_Int preload = 0, loop = 0, loop2 = LOOP2;
HYPRE_Int bcast_rows[2], *info;
hypre_ParCSRMatrix *parcsr_A, *small_A;
HYPRE_ParCSRMatrix A_temp, A_temp_small;
hypre_CSRMatrix *A_CSR;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int nx, ny, nz;
HYPRE_Int P, Q, R;
HYPRE_Int p, q, r;
HYPRE_Real values[4];
hypre_ParVector *x_new;
hypre_ParVector *y_new, *y;
HYPRE_Int *row_starts;
HYPRE_Real ans;
HYPRE_Real start_time, end_time, total_time, *loop_times;
HYPRE_Real T_avg, T_std;
HYPRE_Int noparmprint = 0;
#if mydebug
HYPRE_Int j, tmp_int;
#endif
/*-----------------------------------------------------------
* 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 );
/*-----------------------------------------------------------
* default - is 27pt laplace
*-----------------------------------------------------------*/
build_matrix_type = 2;
matrix_arg_index = argc;
/*-----------------------------------------------------------
* Parse command line
*-----------------------------------------------------------*/
index = 1;
while ( index < argc)
{
if ( strcmp(argv[index], "-verbose") == 0 )
{
index++;
verbose = 1;
}
else if ( strcmp(argv[index], "-fromonecsrfile") == 0 )
{
index++;
build_matrix_type = 1;
matrix_arg_index = index; /*this tells where the name is*/
}
else if ( strcmp(argv[index], "-commpkg") == 0 )
{
index++;
commpkg_flag = atoi(argv[index++]);
}
else if ( strcmp(argv[index], "-laplacian") == 0 )
{
index++;
build_matrix_type = 2;
matrix_arg_index = index;
}
else if ( strcmp(argv[index], "-27pt") == 0 )
{
index++;
build_matrix_type = 4;
matrix_arg_index = index;
}
/*
else if ( strcmp(argv[index], "-nopreload") == 0 )
{
index++;
preload = 0;
}
*/
else if ( strcmp(argv[index], "-loop") == 0 )
{
index++;
loop = atoi(argv[index++]);
}
else if ( strcmp(argv[index], "-noparmprint") == 0 )
{
index++;
noparmprint = 1;
}
else
{
index++;
/*hypre_printf("Warning: Unrecogized option '%s'\n",argv[index++] );*/
}
}
/*-----------------------------------------------------------
* Setup the Matrix problem
*-----------------------------------------------------------*/
/*-----------------------------------------------------------
* Get actual partitioning-
* read in an actual csr matrix.
*-----------------------------------------------------------*/
if (build_matrix_type ==1) /*read in a csr matrix from one file */
{
if (matrix_arg_index < argc)
{
csrfilename = argv[matrix_arg_index];
}
else
{
hypre_printf("Error: No filename specified \n");
exit(1);
}
if (myid == 0)
{
/*hypre_printf(" FromFile: %s\n", csrfilename);*/
A_CSR = hypre_CSRMatrixRead(csrfilename);
}
row_part = NULL;
col_part = NULL;
parcsr_A = hypre_CSRMatrixToParCSRMatrix(hypre_MPI_COMM_WORLD, A_CSR,
row_part, col_part);
if (myid == 0) hypre_CSRMatrixDestroy(A_CSR);
}
else if (build_matrix_type ==2)
{
myBuildParLaplacian(argc, argv, matrix_arg_index, &A_temp, !noparmprint);
parcsr_A = (hypre_ParCSRMatrix *) A_temp;
}
else if (build_matrix_type ==4)
{
myBuildParLaplacian27pt(argc, argv, matrix_arg_index, &A_temp, !noparmprint);
parcsr_A = (hypre_ParCSRMatrix *) A_temp;
}
/*-----------------------------------------------------------
* create a small problem so that timings are more accurate -
* code gets run twice (small laplace)
*-----------------------------------------------------------*/
/*this is no longer being used - preload = 0 is set at the beginning */
if (preload == 1)
{
/*hypre_printf("preload!\n");*/
values[1] = -1;
values[2] = -1;
values[3] = -1;
values[0] = - 6.0 ;
nx = 2;
ny = num_procs;
nz = 2;
P = 1;
Q = num_procs;
R = 1;
p = myid % P;
q = (( myid - p)/P) % Q;
r = ( myid - p - P*q)/( P*Q );
A_temp_small = (HYPRE_ParCSRMatrix) GenerateLaplacian(hypre_MPI_COMM_WORLD, nx, ny, nz,
P, Q, R, p, q, r, values);
small_A = (hypre_ParCSRMatrix *) A_temp_small;
/*do comm packages*/
hypre_NewCommPkgCreate(small_A);
hypre_NewCommPkgDestroy(small_A);
hypre_MatvecCommPkgCreate(small_A);
hypre_ParCSRMatrixDestroy(small_A);
}
/*-----------------------------------------------------------
* Prepare for timing
*-----------------------------------------------------------*/
/* instead of preloading, let's not time the first one if more than one*/
if (!loop)
{
loop = 1;
/* and don't do any timings */
}
else
{
loop +=1;
if (loop < 2) loop = 2;
}
loop_times = hypre_CTAlloc(HYPRE_Real, loop);
/******************************************************************************************/
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
if (commpkg_flag == 1 || commpkg_flag ==3 )
{
/*-----------------------------------------------------------
* Create new comm package
*-----------------------------------------------------------*/
if (!myid) hypre_printf("********************************************************\n" );
/*do loop times*/
for (i=0; i< loop; i++)
{
loop_times[i] = 0.0;
for (k=0; k< loop2; k++)
{
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
start_time = hypre_MPI_Wtime();
#if mpip_on
if (i==(loop-1)) hypre_MPI_Pcontrol(1);
#endif
hypre_NewCommPkgCreate(parcsr_A);
#if mpip_on
if (i==(loop-1)) hypre_MPI_Pcontrol(0);
#endif
end_time = hypre_MPI_Wtime();
end_time = end_time - start_time;
hypre_MPI_Allreduce(&end_time, &total_time, 1,
HYPRE_MPI_REAL, hypre_MPI_MAX, hypre_MPI_COMM_WORLD);
loop_times[i] += total_time;
if ( !((i+1)== loop && (k+1) == loop2)) hypre_NewCommPkgDestroy(parcsr_A);
}/*end of loop2 */
} /*end of loop*/
/* calculate the avg and std. */
if (loop > 1)
{
/* calculate the avg and std. */
stats_mo(loop_times, loop, &T_avg, &T_std);
if (!myid) hypre_printf(" NewCommPkgCreate: AVG. wall clock time = %f seconds\n", T_avg);
if (!myid) hypre_printf(" STD. for %d runs = %f\n", loop-1, T_std);
if (!myid) hypre_printf(" (Note: avg./std. timings exclude run 0.)\n");
if (!myid) hypre_printf("********************************************************\n" );
for (i=0; i< loop; i++)
{
if (!myid) hypre_printf(" run %d = %f sec.\n", i, loop_times[i]);
}
if (!myid) hypre_printf("********************************************************\n" );
}
else
{
if (!myid) hypre_printf("********************************************************\n" );
if (!myid) hypre_printf(" NewCommPkgCreate:\n");
if (!myid) hypre_printf(" run time = %f sec.\n", loop_times[0]);
if (!myid) hypre_printf("********************************************************\n" );
}
/*-----------------------------------------------------------
* Verbose printing
*-----------------------------------------------------------*/
/*some verification*/
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(parcsr_A);
if (verbose)
{
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
comm_pkg = hypre_ParCSRMatrixCommPkg(parcsr_A);
hypre_printf("myid = %i, my ACTUAL local range: [%i, %i]\n", myid,
row_start, row_end);
ierr = hypre_GetAssumedPartitionRowRange( myid, global_num_rows, &row_start,
&row_end);
hypre_printf("myid = %i, my assumed local range: [%i, %i]\n", myid,
row_start, row_end);
hypre_printf("myid = %d, num_recvs = %d\n", myid,
hypre_ParCSRCommPkgNumRecvs(comm_pkg) );
#if mydebug
for (i=0; i < hypre_ParCSRCommPkgNumRecvs(comm_pkg); i++)
{
hypre_printf("myid = %d, recv proc = %d, vec_starts = [%d : %d]\n",
myid, hypre_ParCSRCommPkgRecvProcs(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i+1]-1);
}
#endif
hypre_printf("myid = %d, num_sends = %d\n", myid,
hypre_ParCSRCommPkgNumSends(comm_pkg) );
#if mydebug
for (i=0; i <hypre_ParCSRCommPkgNumSends(comm_pkg) ; i++)
{
tmp_int = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i+1] -
hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
index = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
for (j=0; j< tmp_int; j++)
{
hypre_printf("myid = %d, send proc = %d, send element = %d\n",myid,
hypre_ParCSRCommPkgSendProcs(comm_pkg)[i],
hypre_ParCSRCommPkgSendMapElmts(comm_pkg)[index+j]);
}
}
#endif
}
/*-----------------------------------------------------------
* To verify correctness (if commpkg_flag = 3)
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
/*do a matvec - we are assuming a square matrix */
row_starts = hypre_ParCSRMatrixRowStarts(parcsr_A);
x_new = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows, row_starts);
hypre_ParVectorSetPartitioningOwner(x_new, 0);
hypre_ParVectorInitialize(x_new);
hypre_ParVectorSetRandomValues(x_new, 1);
y_new = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows, row_starts);
hypre_ParVectorSetPartitioningOwner(y_new, 0);
hypre_ParVectorInitialize(y_new);
hypre_ParVectorSetConstantValues(y_new, 0.0);
/*y = 1.0*A*x+1.0*y */
hypre_ParCSRMatrixMatvec (1.0, parcsr_A, x_new, 1.0, y_new);
}
/*-----------------------------------------------------------
* Clean up after MyComm
*-----------------------------------------------------------*/
hypre_NewCommPkgDestroy(parcsr_A);
}
/******************************************************************************************/
/******************************************************************************************/
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
if (commpkg_flag > 1 )
{
/*-----------------------------------------------------------
* Set up standard comm package
*-----------------------------------------------------------*/
bcast_rows[0] = 23;
bcast_rows[1] = 1789;
if (!myid) hypre_printf("********************************************************\n" );
/*do loop times*/
for (i=0; i< loop; i++)
{
loop_times[i] = 0.0;
for (k=0; k< loop2; k++)
{
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
start_time = hypre_MPI_Wtime();
#if time_gather
info = hypre_CTAlloc(HYPRE_Int, num_procs);
hypre_MPI_Allgather(bcast_rows, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, hypre_MPI_COMM_WORLD);
#endif
hypre_MatvecCommPkgCreate(parcsr_A);
end_time = hypre_MPI_Wtime();
end_time = end_time - start_time;
hypre_MPI_Allreduce(&end_time, &total_time, 1,
HYPRE_MPI_REAL, hypre_MPI_MAX, hypre_MPI_COMM_WORLD);
loop_times[i] += total_time;
if ( !((i+1)== loop && (k+1) == loop2)) hypre_MatvecCommPkgDestroy(hypre_ParCSRMatrixCommPkg(parcsr_A));
}/* end of loop 2*/
} /*end of loop*/
/* calculate the avg and std. */
if (loop > 1)
{
stats_mo(loop_times, loop, &T_avg, &T_std);
if (!myid) hypre_printf("Current CommPkgCreate: AVG. wall clock time = %f seconds\n", T_avg);
if (!myid) hypre_printf(" STD. for %d runs = %f\n", loop-1, T_std);
if (!myid) hypre_printf(" (Note: avg./std. timings exclude run 0.)\n");
if (!myid) hypre_printf("********************************************************\n" );
for (i=0; i< loop; i++)
{
if (!myid) hypre_printf(" run %d = %f sec.\n", i, loop_times[i]);
}
if (!myid) hypre_printf("********************************************************\n" );
}
else
{
if (!myid) hypre_printf("********************************************************\n" );
if (!myid) hypre_printf(" Current CommPkgCreate:\n");
if (!myid) hypre_printf(" run time = %f sec.\n", loop_times[0]);
if (!myid) hypre_printf("********************************************************\n" );
}
/*-----------------------------------------------------------
* Verbose printing
*-----------------------------------------------------------*/
/*some verification*/
if (verbose)
{
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
comm_pkg = hypre_ParCSRMatrixCommPkg(parcsr_A);
hypre_printf("myid = %i, std - my local range: [%i, %i]\n", myid,
row_start, row_end);
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
hypre_printf("myid = %d, std - num_recvs = %d\n", myid,
hypre_ParCSRCommPkgNumRecvs(comm_pkg) );
#if mydebug
for (i=0; i < hypre_ParCSRCommPkgNumRecvs(comm_pkg); i++)
{
hypre_printf("myid = %d, std - recv proc = %d, vec_starts = [%d : %d]\n",
myid, hypre_ParCSRCommPkgRecvProcs(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i+1]-1);
}
#endif
hypre_printf("myid = %d, std - num_sends = %d\n", myid,
hypre_ParCSRCommPkgNumSends(comm_pkg));
#if mydebug
for (i=0; i <hypre_ParCSRCommPkgNumSends(comm_pkg) ; i++)
{
tmp_int = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i+1] -
hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
index = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
for (j=0; j< tmp_int; j++)
{
hypre_printf("myid = %d, std - send proc = %d, send element = %d\n",myid,
hypre_ParCSRCommPkgSendProcs(comm_pkg)[i],
hypre_ParCSRCommPkgSendMapElmts(comm_pkg)[index+j]);
}
}
#endif
}
/*-----------------------------------------------------------
* Verify correctness
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(parcsr_A);
row_starts = hypre_ParCSRMatrixRowStarts(parcsr_A);
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y, 0);
hypre_ParVectorInitialize(y);
hypre_ParVectorSetConstantValues(y, 0.0);
hypre_ParCSRMatrixMatvec (1.0, parcsr_A, x_new, 1.0, y);
}
}
/*-----------------------------------------------------------
* Compare matvecs for both comm packages (3)
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
/*make sure that y and y_new are the same - now y_new should=0*/
hypre_ParVectorAxpy( -1.0, y, y_new );
hypre_ParVectorSetRandomValues(y, 1);
ans = hypre_ParVectorInnerProd( y, y_new );
if (!myid)
{
if ( fabs(ans) > 1e-8 )
{
hypre_printf("!!!!! WARNING !!!!! should be zero if correct = %6.10f\n",
ans);
}
else
{
hypre_printf("Matvecs match ( should be zero = %6.10f )\n",
ans);
}
}
}
/*-----------------------------------------------------------
* Clean up
*-----------------------------------------------------------*/
hypre_ParCSRMatrixDestroy(parcsr_A); /*this calls the standard comm
package destroy - but we'll destroy
ours separately until it is
incorporated */
if (commpkg_flag == 3 )
{
hypre_ParVectorDestroy(x_new);
hypre_ParVectorDestroy(y);
hypre_ParVectorDestroy(y_new);
}
hypre_MPI_Finalize();
return(ierr);
}
HYPRE_Int main( HYPRE_Int argc, char *argv[] )
{
hypre_ParCSRMatrix *par_matrix, *g_matrix, **submatrices;
hypre_CSRMatrix *A_diag, *A_offd;
hypre_CSRBlockMatrix *diag;
hypre_CSRBlockMatrix *offd;
hypre_ParCSRBlockMatrix *par_blk_matrix, *par_blk_matrixT, *rap_matrix;
hypre_Vector *x_local;
hypre_Vector *y_local;
hypre_ParVector *x;
hypre_ParVector *y;
HYPRE_Solver gmres_solver, precon;
HYPRE_Int *diag_i, *diag_j, *offd_i, *offd_j;
HYPRE_Int *diag_i2, *diag_j2, *offd_i2, *offd_j2;
double *diag_d, *diag_d2, *offd_d, *offd_d2;
HYPRE_Int mypid, local_size, nprocs;
HYPRE_Int global_num_rows, global_num_cols, num_cols_offd;
HYPRE_Int num_nonzeros_diag, num_nonzeros_offd, *colMap;
HYPRE_Int ii, jj, kk, row, col, nnz, *indices, *colMap2;
double *data, ddata, *y_data;
HYPRE_Int *row_starts, *col_starts, *rstarts, *cstarts;
HYPRE_Int *row_starts2, *col_starts2;
HYPRE_Int block_size=2, bnnz=4, *index_set;
FILE *fp;
/* --------------------------------------------- */
/* Initialize MPI */
/* --------------------------------------------- */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &mypid);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &nprocs);
/* build and fetch matrix */
MyBuildParLaplacian9pt((HYPRE_ParCSRMatrix *) &par_matrix);
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(par_matrix);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(par_matrix);
row_starts = hypre_ParCSRMatrixRowStarts(par_matrix);
col_starts = hypre_ParCSRMatrixColStarts(par_matrix);
A_diag = hypre_ParCSRMatrixDiag(par_matrix);
A_offd = hypre_ParCSRMatrixOffd(par_matrix);
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(A_diag);
num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(A_offd);
/* --------------------------------------------- */
/* build vector and apply matvec */
/* --------------------------------------------- */
x = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_cols,col_starts);
hypre_ParVectorSetPartitioningOwner(x,0);
hypre_ParVectorInitialize(x);
x_local = hypre_ParVectorLocalVector(x);
data = hypre_VectorData(x_local);
local_size = col_starts[mypid+1] - col_starts[mypid];
for (ii = 0; ii < local_size; ii++) data[ii] = 1.0;
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y,0);
hypre_ParVectorInitialize(y);
hypre_ParCSRMatrixMatvec (1.0, par_matrix, x, 0.0, y);
ddata = hypre_ParVectorInnerProd(y, y);
if (mypid == 0) hypre_printf("y inner product = %e\n", ddata);
hypre_ParVectorDestroy(x);
hypre_ParVectorDestroy(y);
/* --------------------------------------------- */
/* build block matrix */
/* --------------------------------------------- */
rstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) rstarts[ii] = row_starts[ii];
cstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) cstarts[ii] = col_starts[ii];
par_blk_matrix = hypre_ParCSRBlockMatrixCreate(hypre_MPI_COMM_WORLD,block_size,
global_num_rows, global_num_cols, rstarts,
cstarts, num_cols_offd, num_nonzeros_diag,
num_nonzeros_offd);
colMap = hypre_ParCSRMatrixColMapOffd(par_matrix);
if (num_cols_offd > 0) colMap2 = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
else colMap2 = NULL;
for (ii = 0; ii < num_cols_offd; ii++) colMap2[ii] = colMap[ii];
hypre_ParCSRBlockMatrixColMapOffd(par_blk_matrix) = colMap2;
diag_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(par_matrix));
diag_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixDiag(par_matrix));
diag_d = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(par_matrix));
diag = hypre_ParCSRBlockMatrixDiag(par_blk_matrix);
diag_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
diag_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_diag);
diag_d2 = hypre_CTAlloc(double, num_nonzeros_diag*bnnz);
for (ii = 0; ii <= local_size; ii++) diag_i2[ii] = diag_i[ii];
for (ii = 0; ii < num_nonzeros_diag; ii++) diag_j2[ii] = diag_j[ii];
hypre_CSRBlockMatrixI(diag) = diag_i2;
hypre_CSRBlockMatrixJ(diag) = diag_j2;
for (ii = 0; ii < num_nonzeros_diag; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
diag_d2[ii*bnnz+jj*block_size+kk] = diag_d[ii];
else
diag_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(diag) = diag_d2;
offd_i = hypre_CSRMatrixI(hypre_ParCSRMatrixOffd(par_matrix));
offd_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixOffd(par_matrix));
offd_d = hypre_CSRMatrixData(hypre_ParCSRMatrixOffd(par_matrix));
offd = hypre_ParCSRBlockMatrixOffd(par_blk_matrix);
offd_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
for (ii = 0; ii <= local_size; ii++) offd_i2[ii] = offd_i[ii];
hypre_CSRBlockMatrixI(offd) = offd_i2;
if (num_cols_offd)
{
offd_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_offd);
for (ii = 0; ii < num_nonzeros_offd; ii++) offd_j2[ii] = offd_j[ii];
hypre_CSRBlockMatrixJ(offd) = offd_j2;
offd_d2 = hypre_CTAlloc(double, num_nonzeros_offd*bnnz);
for (ii = 0; ii < num_nonzeros_offd; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
offd_d2[ii*bnnz+jj*block_size+kk] = offd_d[ii];
else
offd_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(offd) = offd_d2;
}
else
{
HYPRE_Int
hypre_BoomerAMGSolveT( void *amg_vdata,
hypre_ParCSRMatrix *A,
hypre_ParVector *f,
hypre_ParVector *u )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
HYPRE_Int amg_print_level;
HYPRE_Int amg_logging;
HYPRE_Real *num_coeffs;
HYPRE_Int *num_variables;
HYPRE_Real cycle_op_count;
HYPRE_Int num_levels;
/* HYPRE_Int num_unknowns; */
HYPRE_Real tol;
char *file_name;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
/* Local variables */
/*FILE *fp;*/
HYPRE_Int j;
HYPRE_Int Solve_err_flag;
HYPRE_Int min_iter;
HYPRE_Int max_iter;
HYPRE_Int cycle_count;
HYPRE_Real total_coeffs;
HYPRE_Int total_variables;
HYPRE_Int num_procs, my_id;
HYPRE_Real alpha = 1.0;
HYPRE_Real beta = -1.0;
HYPRE_Real cycle_cmplxty = 0.0;
HYPRE_Real operat_cmplxty;
HYPRE_Real grid_cmplxty;
HYPRE_Real conv_factor;
HYPRE_Real resid_nrm;
HYPRE_Real resid_nrm_init;
HYPRE_Real relative_resid;
HYPRE_Real rhs_norm;
HYPRE_Real old_resid;
hypre_ParVector *Vtemp;
hypre_ParVector *Residual;
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
amg_print_level = hypre_ParAMGDataPrintLevel(amg_data);
amg_logging = hypre_ParAMGDataLogging(amg_data);
if ( amg_logging>1 )
Residual = hypre_ParAMGDataResidual(amg_data);
file_name = hypre_ParAMGDataLogFileName(amg_data);
/* num_unknowns = hypre_ParAMGDataNumUnknowns(amg_data); */
num_levels = hypre_ParAMGDataNumLevels(amg_data);
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
min_iter = hypre_ParAMGDataMinIter(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_variables = hypre_CTAlloc(HYPRE_Int, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]);
num_variables[0] = hypre_ParCSRMatrixGlobalNumRows(A_array[0]);
A_array[0] = A;
F_array[0] = f;
U_array[0] = u;
/* Vtemp = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(A_array[0]),
hypre_ParCSRMatrixGlobalNumRows(A_array[0]),
hypre_ParCSRMatrixRowStarts(A_array[0]));
hypre_ParVectorInitialize(Vtemp);
hypre_ParVectorSetPartitioningOwner(Vtemp,0);
hypre_ParAMGDataVtemp(amg_data) = Vtemp;
*/
Vtemp = hypre_ParAMGDataVtemp(amg_data);
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]);
num_variables[j] = hypre_ParCSRMatrixGlobalNumRows(A_array[j]);
}
/*-----------------------------------------------------------------------
* Write the solver parameters
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
hypre_BoomerAMGWriteSolverParams(amg_data);
/*-----------------------------------------------------------------------
* Initialize the solver error flag and assorted bookkeeping variables
*-----------------------------------------------------------------------*/
Solve_err_flag = 0;
total_coeffs = 0;
total_variables = 0;
cycle_count = 0;
operat_cmplxty = 0;
grid_cmplxty = 0;
/*-----------------------------------------------------------------------
* open the log file and write some initial info
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
{
/*fp = fopen(file_name, "a");*/
hypre_printf("\n\nAMG SOLUTION INFO:\n");
}
/*-----------------------------------------------------------------------
* Compute initial fine-grid residual and print to logfile
*-----------------------------------------------------------------------*/
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
resid_nrm_init = resid_nrm;
rhs_norm = sqrt(hypre_ParVectorInnerProd(f, f));
relative_resid = 9999;
if (rhs_norm)
{
relative_resid = resid_nrm_init / rhs_norm;
}
if (my_id ==0 && (amg_print_level > 1))
{
hypre_printf(" relative\n");
hypre_printf(" residual factor residual\n");
hypre_printf(" -------- ------ --------\n");
hypre_printf(" Initial %e %e\n",resid_nrm_init,
relative_resid);
}
/*-----------------------------------------------------------------------
* Main V-cycle loop
*-----------------------------------------------------------------------*/
while ((relative_resid >= tol || cycle_count < min_iter)
&& cycle_count < max_iter
&& Solve_err_flag == 0)
{
hypre_ParAMGDataCycleOpCount(amg_data) = 0;
/* Op count only needed for one cycle */
Solve_err_flag = hypre_BoomerAMGCycleT(amg_data, F_array, U_array);
old_resid = resid_nrm;
/*---------------------------------------------------------------
* Compute fine-grid residual and residual norm
*----------------------------------------------------------------*/
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
conv_factor = resid_nrm / old_resid;
relative_resid = 9999;
if (rhs_norm)
{
relative_resid = resid_nrm / rhs_norm;
}
++cycle_count;
hypre_ParAMGDataRelativeResidualNorm(amg_data) = relative_resid;
hypre_ParAMGDataNumIterations(amg_data) = cycle_count;
if (my_id == 0 && (amg_print_level > 1))
{
hypre_printf(" Cycle %2d %e %f %e \n", cycle_count,
resid_nrm, conv_factor, relative_resid);
}
}
if (cycle_count == max_iter) Solve_err_flag = 1;
/*-----------------------------------------------------------------------
* Compute closing statistics
*-----------------------------------------------------------------------*/
conv_factor = pow((resid_nrm/resid_nrm_init),(1.0/((HYPRE_Real) cycle_count)));
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
total_coeffs += num_coeffs[j];
total_variables += num_variables[j];
}
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
if (num_variables[0])
grid_cmplxty = ((HYPRE_Real) total_variables) / ((HYPRE_Real) num_variables[0]);
if (num_coeffs[0])
{
operat_cmplxty = total_coeffs / num_coeffs[0];
cycle_cmplxty = cycle_op_count / num_coeffs[0];
}
if (my_id == 0 && amg_print_level > 1)
{
if (Solve_err_flag == 1)
{
hypre_printf("\n\n==============================================");
hypre_printf("\n NOTE: Convergence tolerance was not achieved\n");
hypre_printf(" within the allowed %d V-cycles\n",max_iter);
hypre_printf("==============================================");
}
hypre_printf("\n\n Average Convergence Factor = %f",conv_factor);
hypre_printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
hypre_printf(" operator = %f\n",operat_cmplxty);
hypre_printf(" cycle = %f\n\n",cycle_cmplxty);
}
/*----------------------------------------------------------
* Close the output file (if open)
*----------------------------------------------------------*/
/*if (my_id == 0 && amg_print_level >= 1)
{
fclose(fp);
}*/
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
return(Solve_err_flag);
}
HYPRE_Int
hypre_BoomerAMGSolve( void *amg_vdata,
hypre_ParCSRMatrix *A,
hypre_ParVector *f,
hypre_ParVector *u )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
HYPRE_Int amg_print_level;
HYPRE_Int amg_logging;
HYPRE_Int cycle_count;
HYPRE_Int num_levels;
/* HYPRE_Int num_unknowns; */
HYPRE_Real tol;
HYPRE_Int block_mode;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParCSRBlockMatrix **A_block_array;
/* Local variables */
HYPRE_Int j;
HYPRE_Int Solve_err_flag;
HYPRE_Int min_iter;
HYPRE_Int max_iter;
HYPRE_Int num_procs, my_id;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int simple;
HYPRE_Real alpha = 1.0;
HYPRE_Real beta = -1.0;
HYPRE_Real cycle_op_count;
HYPRE_Real total_coeffs;
HYPRE_Real total_variables;
HYPRE_Real *num_coeffs;
HYPRE_Real *num_variables;
HYPRE_Real cycle_cmplxty = 0.0;
HYPRE_Real operat_cmplxty;
HYPRE_Real grid_cmplxty;
HYPRE_Real conv_factor = 0.0;
HYPRE_Real resid_nrm = 1.0;
HYPRE_Real resid_nrm_init = 0.0;
HYPRE_Real relative_resid;
HYPRE_Real rhs_norm = 0.0;
HYPRE_Real old_resid;
HYPRE_Real ieee_check = 0.;
hypre_ParVector *Vtemp;
hypre_ParVector *Residual;
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
amg_print_level = hypre_ParAMGDataPrintLevel(amg_data);
amg_logging = hypre_ParAMGDataLogging(amg_data);
if ( amg_logging > 1 )
Residual = hypre_ParAMGDataResidual(amg_data);
/* num_unknowns = hypre_ParAMGDataNumUnknowns(amg_data); */
num_levels = hypre_ParAMGDataNumLevels(amg_data);
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
min_iter = hypre_ParAMGDataMinIter(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
simple = hypre_ParAMGDataSimple(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
A_array[0] = A;
F_array[0] = f;
U_array[0] = u;
block_mode = hypre_ParAMGDataBlockMode(amg_data);
A_block_array = hypre_ParAMGDataABlockArray(amg_data);
/* Vtemp = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(A_array[0]),
hypre_ParCSRMatrixGlobalNumRows(A_array[0]),
hypre_ParCSRMatrixRowStarts(A_array[0]));
hypre_ParVectorInitialize(Vtemp);
hypre_ParVectorSetPartitioningOwner(Vtemp,0);
hypre_ParAMGDataVtemp(amg_data) = Vtemp;
*/
Vtemp = hypre_ParAMGDataVtemp(amg_data);
/*-----------------------------------------------------------------------
* Write the solver parameters
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
hypre_BoomerAMGWriteSolverParams(amg_data);
/*-----------------------------------------------------------------------
* Initialize the solver error flag and assorted bookkeeping variables
*-----------------------------------------------------------------------*/
Solve_err_flag = 0;
total_coeffs = 0;
total_variables = 0;
cycle_count = 0;
operat_cmplxty = 0;
grid_cmplxty = 0;
/*-----------------------------------------------------------------------
* write some initial info
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1 && tol > 0.)
hypre_printf("\n\nAMG SOLUTION INFO:\n");
/*-----------------------------------------------------------------------
* Compute initial fine-grid residual and print
*-----------------------------------------------------------------------*/
if (amg_print_level > 1 || amg_logging > 1)
{
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
if (tol > 0)
hypre_ParCSRMatrixMatvec(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
if (tol > 0)
hypre_ParCSRMatrixMatvec(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
/* Since it is does not diminish performance, attempt to return an error flag
and notify users when they supply bad input. */
if (resid_nrm != 0.) ieee_check = resid_nrm/resid_nrm; /* INF -> NaN conversion */
if (ieee_check != ieee_check)
{
/* ...INFs or NaNs in input can make ieee_check a NaN. This test
for ieee_check self-equality works on all IEEE-compliant compilers/
machines, c.f. page 8 of "Lecture Notes on the Status of IEEE 754"
by W. Kahan, May 31, 1996. Currently (July 2002) this paper may be
found at http://HTTP.CS.Berkeley.EDU/~wkahan/ieee754status/IEEE754.PDF */
if (amg_print_level > 0)
{
hypre_printf("\n\nERROR detected by Hypre ... BEGIN\n");
hypre_printf("ERROR -- hypre_BoomerAMGSolve: INFs and/or NaNs detected in input.\n");
hypre_printf("User probably placed non-numerics in supplied A, x_0, or b.\n");
hypre_printf("ERROR detected by Hypre ... END\n\n\n");
}
hypre_error(HYPRE_ERROR_GENERIC);
return hypre_error_flag;
}
resid_nrm_init = resid_nrm;
rhs_norm = sqrt(hypre_ParVectorInnerProd(f, f));
if (rhs_norm)
{
relative_resid = resid_nrm_init / rhs_norm;
}
else
{
relative_resid = resid_nrm_init;
}
}
else
{
relative_resid = 1.;
}
if (my_id == 0 && amg_print_level > 1)
{
hypre_printf(" relative\n");
hypre_printf(" residual factor residual\n");
hypre_printf(" -------- ------ --------\n");
hypre_printf(" Initial %e %e\n",resid_nrm_init,
relative_resid);
}
/*-----------------------------------------------------------------------
* Main V-cycle loop
*-----------------------------------------------------------------------*/
while ((relative_resid >= tol || cycle_count < min_iter)
&& cycle_count < max_iter)
{
hypre_ParAMGDataCycleOpCount(amg_data) = 0;
/* Op count only needed for one cycle */
if ((additive < 0 || additive >= num_levels)
&& (mult_additive < 0 || mult_additive >= num_levels)
&& (simple < 0 || simple >= num_levels) )
hypre_BoomerAMGCycle(amg_data, F_array, U_array);
else
hypre_BoomerAMGAdditiveCycle(amg_data);
/*---------------------------------------------------------------
* Compute fine-grid residual and residual norm
*----------------------------------------------------------------*/
if (amg_print_level > 1 || amg_logging > 1 || tol > 0.)
{
old_resid = resid_nrm;
if ( amg_logging > 1 ) {
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[0], U_array[0], beta, F_array[0], Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[0], U_array[0], beta, F_array[0], Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
if (old_resid) conv_factor = resid_nrm / old_resid;
else conv_factor = resid_nrm;
if (rhs_norm)
{
relative_resid = resid_nrm / rhs_norm;
}
else
{
relative_resid = resid_nrm;
}
hypre_ParAMGDataRelativeResidualNorm(amg_data) = relative_resid;
}
++cycle_count;
hypre_ParAMGDataNumIterations(amg_data) = cycle_count;
#ifdef CUMNUMIT
++hypre_ParAMGDataCumNumIterations(amg_data);
#endif
if (my_id == 0 && amg_print_level > 1)
{
hypre_printf(" Cycle %2d %e %f %e \n", cycle_count,
resid_nrm, conv_factor, relative_resid);
}
}
if (cycle_count == max_iter && tol > 0.)
{
Solve_err_flag = 1;
hypre_error(HYPRE_ERROR_CONV);
}
/*-----------------------------------------------------------------------
* Compute closing statistics
*-----------------------------------------------------------------------*/
if (cycle_count > 0 && resid_nrm_init)
conv_factor = pow((resid_nrm/resid_nrm_init),(1.0/(HYPRE_Real) cycle_count));
else
conv_factor = 1.;
if (amg_print_level > 1)
{
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_variables = hypre_CTAlloc(HYPRE_Real, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A);
num_variables[0] = hypre_ParCSRMatrixGlobalNumRows(A);
if (block_mode)
{
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = (HYPRE_Real) hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]);
num_variables[j] = (HYPRE_Real) hypre_ParCSRBlockMatrixGlobalNumRows(A_block_array[j]);
}
num_coeffs[0] = hypre_ParCSRBlockMatrixDNumNonzeros(A_block_array[0]);
num_variables[0] = hypre_ParCSRBlockMatrixGlobalNumRows(A_block_array[0]);
}
else
{
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = (HYPRE_Real) hypre_ParCSRMatrixNumNonzeros(A_array[j]);
num_variables[j] = (HYPRE_Real) hypre_ParCSRMatrixGlobalNumRows(A_array[j]);
}
}
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
total_coeffs += num_coeffs[j];
total_variables += num_variables[j];
}
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
if (num_variables[0])
grid_cmplxty = total_variables / num_variables[0];
if (num_coeffs[0])
{
operat_cmplxty = total_coeffs / num_coeffs[0];
cycle_cmplxty = cycle_op_count / num_coeffs[0];
}
if (my_id == 0)
{
if (Solve_err_flag == 1)
{
hypre_printf("\n\n==============================================");
hypre_printf("\n NOTE: Convergence tolerance was not achieved\n");
hypre_printf(" within the allowed %d V-cycles\n",max_iter);
hypre_printf("==============================================");
}
hypre_printf("\n\n Average Convergence Factor = %f",conv_factor);
hypre_printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
hypre_printf(" operator = %f\n",operat_cmplxty);
hypre_printf(" cycle = %f\n\n\n\n",cycle_cmplxty);
}
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
}
return hypre_error_flag;
}
HYPRE_Int
hypre_MaxwellSolve2( void * maxwell_vdata,
hypre_SStructMatrix * A_in,
hypre_SStructVector * f,
hypre_SStructVector * u )
{
hypre_MaxwellData *maxwell_data = maxwell_vdata;
hypre_ParVector *f_edge;
hypre_ParVector *u_edge;
HYPRE_Int max_iter = maxwell_data-> max_iter;
double tol = maxwell_data-> tol;
HYPRE_Int rel_change = maxwell_data-> rel_change;
HYPRE_Int zero_guess = maxwell_data-> zero_guess;
HYPRE_Int npre_relax = maxwell_data-> num_pre_relax;
HYPRE_Int npost_relax = maxwell_data-> num_post_relax;
hypre_ParCSRMatrix **Ann_l = maxwell_data-> Ann_l;
hypre_ParCSRMatrix **Pn_l = maxwell_data-> Pn_l;
hypre_ParCSRMatrix **RnT_l = maxwell_data-> RnT_l;
hypre_ParVector **bn_l = maxwell_data-> bn_l;
hypre_ParVector **xn_l = maxwell_data-> xn_l;
hypre_ParVector **resn_l = maxwell_data-> resn_l;
hypre_ParVector **en_l = maxwell_data-> en_l;
hypre_ParVector **nVtemp2_l = maxwell_data-> nVtemp2_l;
HYPRE_Int **nCF_marker_l = maxwell_data-> nCF_marker_l;
double *nrelax_weight= maxwell_data-> nrelax_weight;
double *nomega = maxwell_data-> nomega;
HYPRE_Int nrelax_type = maxwell_data-> nrelax_type;
HYPRE_Int node_numlevs = maxwell_data-> node_numlevels;
hypre_ParCSRMatrix *Tgrad = maxwell_data-> Tgrad;
hypre_ParCSRMatrix *T_transpose = maxwell_data-> T_transpose;
hypre_ParCSRMatrix **Aee_l = maxwell_data-> Aee_l;
hypre_IJMatrix **Pe_l = maxwell_data-> Pe_l;
hypre_IJMatrix **ReT_l = maxwell_data-> ReT_l;
hypre_ParVector **be_l = maxwell_data-> be_l;
hypre_ParVector **xe_l = maxwell_data-> xe_l;
hypre_ParVector **rese_l = maxwell_data-> rese_l;
hypre_ParVector **ee_l = maxwell_data-> ee_l;
hypre_ParVector **eVtemp2_l = maxwell_data-> eVtemp2_l;
HYPRE_Int **eCF_marker_l = maxwell_data-> eCF_marker_l;
double *erelax_weight= maxwell_data-> erelax_weight;
double *eomega = maxwell_data-> eomega;
HYPRE_Int erelax_type = maxwell_data-> erelax_type;
HYPRE_Int edge_numlevs = maxwell_data-> edge_numlevels;
HYPRE_Int **BdryRanks_l = maxwell_data-> BdryRanks_l;
HYPRE_Int *BdryRanksCnts_l= maxwell_data-> BdryRanksCnts_l;
HYPRE_Int logging = maxwell_data-> logging;
double *norms = maxwell_data-> norms;
double *rel_norms = maxwell_data-> rel_norms;
HYPRE_Int Solve_err_flag;
HYPRE_Int relax_local, cycle_param;
double b_dot_b = 0, r_dot_r, eps = 0;
double e_dot_e, x_dot_x;
HYPRE_Int i, j;
HYPRE_Int level;
HYPRE_Int ierr= 0;
/* added for the relaxation routines */
hypre_ParVector *ze = NULL;
if (hypre_NumThreads() > 1)
{
/* Aee is always bigger than Ann */
ze = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(Aee_l[0]),
hypre_ParCSRMatrixGlobalNumRows(Aee_l[0]),
hypre_ParCSRMatrixRowStarts(Aee_l[0]));
hypre_ParVectorInitialize(ze);
hypre_ParVectorSetPartitioningOwner(ze,0);
}
hypre_BeginTiming(maxwell_data-> time_index);
hypre_SStructVectorConvert(f, &f_edge);
hypre_SStructVectorConvert(u, &u_edge);
hypre_ParVectorZeroBCValues(f_edge, BdryRanks_l[0], BdryRanksCnts_l[0]);
hypre_ParVectorZeroBCValues(u_edge, BdryRanks_l[0], BdryRanksCnts_l[0]);
be_l[0]= f_edge;
xe_l[0]= u_edge;
/* the nodal fine vectors: xn= 0. bn= T'*(be- Aee*xe) is updated in the cycle. */
hypre_ParVectorSetConstantValues(xn_l[0], 0.0);
relax_local= 0;
cycle_param= 0;
(maxwell_data-> num_iterations) = 0;
/* if max_iter is zero, return */
if (max_iter == 0)
{
/* if using a zero initial guess, return zero */
if (zero_guess)
{
hypre_ParVectorSetConstantValues(xe_l[0], 0.0);
}
hypre_EndTiming(maxwell_data -> time_index);
return ierr;
}
/* part of convergence check */
if (tol > 0.0)
{
/* eps = (tol^2) */
b_dot_b= hypre_ParVectorInnerProd(be_l[0], be_l[0]);
eps = tol*tol;
/* if rhs is zero, return a zero solution */
if (b_dot_b == 0.0)
{
hypre_ParVectorSetConstantValues(xe_l[0], 0.0);
if (logging > 0)
{
norms[0] = 0.0;
rel_norms[0] = 0.0;
}
hypre_EndTiming(maxwell_data -> time_index);
return ierr;
}
}
/*-----------------------------------------------------
* Do V-cycles:
* For each index l, "fine" = l, "coarse" = (l-1)
*
* solution update:
* edge_sol= edge_sol + T*node_sol
*-----------------------------------------------------*/
for (i = 0; i < max_iter; i++)
{
/* compute fine grid residual & nodal rhs. */
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
hypre_ParVectorZeroBCValues(rese_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]);
hypre_ParCSRMatrixMatvec(1.0, T_transpose, rese_l[0], 0.0, bn_l[0]);
/* convergence check */
if (tol > 0.0)
{
r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]);
if (logging > 0)
{
norms[i] = sqrt(r_dot_r);
if (b_dot_b > 0)
rel_norms[i] = sqrt(r_dot_r/b_dot_b);
else
rel_norms[i] = 0.0;
}
/* always do at least 1 V-cycle */
if ((r_dot_r/b_dot_b < eps) && (i > 0))
{
if (rel_change)
{
if ((e_dot_e/x_dot_x) < eps)
break;
}
else
{
break;
}
}
}
hypre_ParVectorCopy(bn_l[0], resn_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]);
r_dot_r= hypre_ParVectorInnerProd(resn_l[0], resn_l[0]);
for (level= 0; level<= node_numlevs-2; level++)
{
/*-----------------------------------------------
* Down cycle
*-----------------------------------------------*/
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
} /*for (j= 0; j< npre_relax; j++) */
/* compute residuals */
hypre_ParVectorCopy(bn_l[level], resn_l[level]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[level], xn_l[level],
1.0, resn_l[level]);
/* restrict residuals */
hypre_ParCSRMatrixMatvecT(1.0, RnT_l[level], resn_l[level],
0.0, bn_l[level+1]);
/* zero off initial guess for the next level */
hypre_ParVectorSetConstantValues(xn_l[level+1], 0.0);
} /* for (level= 0; level<= node_numlevs-2; level++) */
/* coarsest node solve */
level= node_numlevs-1;
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
/*---------------------------------------------------------------------
* Cycle up the levels.
*---------------------------------------------------------------------*/
for (level= (node_numlevs - 2); level>= 1; level--)
{
hypre_ParCSRMatrixMatvec(1.0, Pn_l[level], xn_l[level+1], 0.0,
en_l[level]);
hypre_ParVectorAxpy(1.0, en_l[level], xn_l[level]);
/* post smooth */
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
}
} /* for (level= (en_numlevs - 2); level>= 1; level--) */
/* interpolate error and correct on finest grids */
hypre_ParCSRMatrixMatvec(1.0, Pn_l[0], xn_l[1], 0.0, en_l[0]);
hypre_ParVectorAxpy(1.0, en_l[0], xn_l[0]);
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[0],
bn_l[0],
nCF_marker_l[0],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[0],
nomega[0],
NULL,
xn_l[0],
nVtemp2_l[0],
ze);
} /* for (j= 0; j< npost_relax; j++) */
hypre_ParVectorCopy(bn_l[0], resn_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]);
/* add the gradient solution component to xe_l[0] */
hypre_ParCSRMatrixMatvec(1.0, Tgrad, xn_l[0], 1.0, xe_l[0]);
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]);
for (level= 0; level<= edge_numlevs-2; level++)
{
/*-----------------------------------------------
* Down cycle
*-----------------------------------------------*/
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
} /*for (j= 0; j< npre_relax; j++) */
/* compute residuals */
hypre_ParVectorCopy(be_l[level], rese_l[level]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[level], xe_l[level],
1.0, rese_l[level]);
/* restrict residuals */
hypre_ParCSRMatrixMatvecT(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(ReT_l[level]),
rese_l[level], 0.0, be_l[level+1]);
hypre_ParVectorZeroBCValues(be_l[level+1], BdryRanks_l[level+1],
BdryRanksCnts_l[level+1]);
/* zero off initial guess for the next level */
hypre_ParVectorSetConstantValues(xe_l[level+1], 0.0);
} /* for (level= 1; level<= edge_numlevels-2; level++) */
/* coarsest edge solve */
level= edge_numlevs-1;
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
}
/*---------------------------------------------------------------------
* Up cycle.
*---------------------------------------------------------------------*/
for (level= (edge_numlevs - 2); level>= 1; level--)
{
hypre_ParCSRMatrixMatvec(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[level]),
xe_l[level+1], 0.0, ee_l[level]);
hypre_ParVectorZeroBCValues(ee_l[level], BdryRanks_l[level],
BdryRanksCnts_l[level]);
hypre_ParVectorAxpy(1.0, ee_l[level], xe_l[level]);
/* post smooth */
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
}
} /* for (level= (edge_numlevs - 2); level>= 1; level--) */
/* interpolate error and correct on finest grids */
hypre_ParCSRMatrixMatvec(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[0]),
xe_l[1], 0.0, ee_l[0]);
hypre_ParVectorZeroBCValues(ee_l[0], BdryRanks_l[0],
BdryRanksCnts_l[0]);
hypre_ParVectorAxpy(1.0, ee_l[0], xe_l[0]);
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[0],
be_l[0],
eCF_marker_l[0],
erelax_type,
relax_local,
cycle_param,
erelax_weight[0],
eomega[0],
NULL,
xe_l[0],
eVtemp2_l[0],
ze);
} /* for (j= 0; j< npost_relax; j++) */
e_dot_e= hypre_ParVectorInnerProd(ee_l[0], ee_l[0]);
x_dot_x= hypre_ParVectorInnerProd(xe_l[0], xe_l[0]);
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
(maxwell_data -> num_iterations) = (i + 1);
}
hypre_EndTiming(maxwell_data -> time_index);
if (ze)
hypre_ParVectorDestroy(ze);
return ierr;
}
HYPRE_Int
hypre_BoomerAMGCycle( void *amg_vdata,
hypre_ParVector **F_array,
hypre_ParVector **U_array )
{
hypre_ParAMGData *amg_data = amg_vdata;
HYPRE_Solver *smoother;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_ParCSRMatrix **R_array;
hypre_ParVector *Utemp;
hypre_ParVector *Vtemp;
hypre_ParVector *Rtemp;
hypre_ParVector *Ptemp;
hypre_ParVector *Ztemp;
hypre_ParVector *Aux_U;
hypre_ParVector *Aux_F;
hypre_ParCSRBlockMatrix **A_block_array;
hypre_ParCSRBlockMatrix **P_block_array;
hypre_ParCSRBlockMatrix **R_block_array;
HYPRE_Real *Ztemp_data;
HYPRE_Real *Ptemp_data;
HYPRE_Int **CF_marker_array;
/* HYPRE_Int **unknown_map_array;
HYPRE_Int **point_map_array;
HYPRE_Int **v_at_point_array; */
HYPRE_Real cycle_op_count;
HYPRE_Int cycle_type;
HYPRE_Int num_levels;
HYPRE_Int max_levels;
HYPRE_Real *num_coeffs;
HYPRE_Int *num_grid_sweeps;
HYPRE_Int *grid_relax_type;
HYPRE_Int **grid_relax_points;
HYPRE_Int block_mode;
HYPRE_Real *max_eig_est;
HYPRE_Real *min_eig_est;
HYPRE_Int cheby_order;
HYPRE_Real cheby_fraction;
/* Local variables */
HYPRE_Int *lev_counter;
HYPRE_Int Solve_err_flag;
HYPRE_Int k;
HYPRE_Int i, j, jj;
HYPRE_Int level;
HYPRE_Int cycle_param;
HYPRE_Int coarse_grid;
HYPRE_Int fine_grid;
HYPRE_Int Not_Finished;
HYPRE_Int num_sweep;
HYPRE_Int cg_num_sweep = 1;
HYPRE_Int relax_type;
HYPRE_Int relax_points;
HYPRE_Int relax_order;
HYPRE_Int relax_local;
HYPRE_Int old_version = 0;
HYPRE_Real *relax_weight;
HYPRE_Real *omega;
HYPRE_Real alfa, beta, gammaold;
HYPRE_Real gamma = 1.0;
HYPRE_Int local_size;
/* HYPRE_Int *smooth_option; */
HYPRE_Int smooth_type;
HYPRE_Int smooth_num_levels;
HYPRE_Int num_threads, my_id;
HYPRE_Real alpha;
HYPRE_Real **l1_norms = NULL;
HYPRE_Real *l1_norms_level;
HYPRE_Int seq_cg = 0;
MPI_Comm comm;
#if 0
HYPRE_Real *D_mat;
HYPRE_Real *S_vec;
#endif
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
R_array = hypre_ParAMGDataRArray(amg_data);
CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data);
Vtemp = hypre_ParAMGDataVtemp(amg_data);
Rtemp = hypre_ParAMGDataRtemp(amg_data);
Ptemp = hypre_ParAMGDataPtemp(amg_data);
Ztemp = hypre_ParAMGDataZtemp(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
max_levels = hypre_ParAMGDataMaxLevels(amg_data);
cycle_type = hypre_ParAMGDataCycleType(amg_data);
A_block_array = hypre_ParAMGDataABlockArray(amg_data);
P_block_array = hypre_ParAMGDataPBlockArray(amg_data);
R_block_array = hypre_ParAMGDataRBlockArray(amg_data);
block_mode = hypre_ParAMGDataBlockMode(amg_data);
num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data);
relax_order = hypre_ParAMGDataRelaxOrder(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
smooth_type = hypre_ParAMGDataSmoothType(amg_data);
smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data);
l1_norms = hypre_ParAMGDataL1Norms(amg_data);
/* smooth_option = hypre_ParAMGDataSmoothOption(amg_data); */
max_eig_est = hypre_ParAMGDataMaxEigEst(amg_data);
min_eig_est = hypre_ParAMGDataMinEigEst(amg_data);
cheby_order = hypre_ParAMGDataChebyOrder(amg_data);
cheby_fraction = hypre_ParAMGDataChebyFraction(amg_data);
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
lev_counter = hypre_CTAlloc(HYPRE_Int, num_levels);
if (hypre_ParAMGDataParticipate(amg_data)) seq_cg = 1;
/* Initialize */
Solve_err_flag = 0;
if (grid_relax_points) old_version = 1;
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]);
comm = hypre_ParCSRMatrixComm(A_array[0]);
hypre_MPI_Comm_rank(comm,&my_id);
if (block_mode)
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]);
}
else
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]);
}
/*---------------------------------------------------------------------
* Initialize cycling control counter
*
* Cycling is controlled using a level counter: lev_counter[k]
*
* Each time relaxation is performed on level k, the
* counter is decremented by 1. If the counter is then
* negative, we go to the next finer level. If non-
* negative, we go to the next coarser level. The
* following actions control cycling:
*
* a. lev_counter[0] is initialized to 1.
* b. lev_counter[k] is initialized to cycle_type for k>0.
*
* c. During cycling, when going down to level k, lev_counter[k]
* is set to the max of (lev_counter[k],cycle_type)
*---------------------------------------------------------------------*/
Not_Finished = 1;
lev_counter[0] = 1;
for (k = 1; k < num_levels; ++k)
{
lev_counter[k] = cycle_type;
}
level = 0;
cycle_param = 1;
smoother = hypre_ParAMGDataSmoother(amg_data);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8
|| smooth_type == 17 || smooth_type == 18
|| smooth_type == 9 || smooth_type == 19)
{
HYPRE_Int actual_local_size = hypre_ParVectorActualLocalSize(Vtemp);
Utemp = hypre_ParVectorCreate(comm,hypre_ParVectorGlobalSize(Vtemp),
hypre_ParVectorPartitioning(Vtemp));
hypre_ParVectorOwnsPartitioning(Utemp) = 0;
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp));
if (local_size < actual_local_size)
{
hypre_VectorData(hypre_ParVectorLocalVector(Utemp)) =
hypre_CTAlloc(HYPRE_Complex, actual_local_size);
hypre_ParVectorActualLocalSize(Utemp) = actual_local_size;
}
else
hypre_ParVectorInitialize(Utemp);
}
}
/*---------------------------------------------------------------------
* Main loop of cycling
*--------------------------------------------------------------------*/
while (Not_Finished)
{
if (num_levels > 1)
{
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(F_array[level]));
hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp)) = local_size;
if (smooth_num_levels <= level)
{
cg_num_sweep = 1;
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = U_array[level];
Aux_F = F_array[level];
}
else if (smooth_type > 9)
{
hypre_VectorSize(hypre_ParVectorLocalVector(Ztemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Rtemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Ptemp)) = local_size;
Ztemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ztemp));
Ptemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ptemp));
hypre_ParVectorSetConstantValues(Ztemp,0);
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, F_array[level], Rtemp);
cg_num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = Ztemp;
Aux_F = Rtemp;
}
else
{
cg_num_sweep = 1;
num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
Aux_U = U_array[level];
Aux_F = F_array[level];
}
relax_type = grid_relax_type[cycle_param];
}
else /* AB: 4/08: removed the max_levels > 1 check - should do this when max-levels = 1 also */
{
/* If no coarsening occurred, apply a simple smoother once */
Aux_U = U_array[level];
Aux_F = F_array[level];
num_sweep = 1;
/* TK: Use the user relax type (instead of 0) to allow for setting a
convergent smoother (e.g. in the solution of singular problems). */
relax_type = hypre_ParAMGDataUserRelaxType(amg_data);
}
if (l1_norms != NULL)
l1_norms_level = l1_norms[level];
else
l1_norms_level = NULL;
if (cycle_param == 3 && seq_cg)
{
hypre_seqAMGCycle(amg_data, level, F_array, U_array);
}
else
{
/*------------------------------------------------------------------
* Do the relaxation num_sweep times
*-----------------------------------------------------------------*/
for (jj = 0; jj < cg_num_sweep; jj++)
{
if (smooth_num_levels > level && smooth_type > 9)
hypre_ParVectorSetConstantValues(Aux_U,0);
for (j = 0; j < num_sweep; j++)
{
if (num_levels == 1 && max_levels > 1)
{
relax_points = 0;
relax_local = 0;
}
else
{
if (old_version)
relax_points = grid_relax_points[cycle_param][j];
relax_local = relax_order;
}
/*-----------------------------------------------
* VERY sloppy approximation to cycle complexity
*-----------------------------------------------*/
if (old_version && level < num_levels -1)
{
switch (relax_points)
{
case 1:
cycle_op_count += num_coeffs[level+1];
break;
case -1:
cycle_op_count += (num_coeffs[level]-num_coeffs[level+1]);
break;
}
}
else
{
cycle_op_count += num_coeffs[level];
}
/*-----------------------------------------------
Choose Smoother
-----------------------------------------------*/
if (smooth_num_levels > level &&
(smooth_type == 7 || smooth_type == 8 ||
smooth_type == 9 || smooth_type == 19 ||
smooth_type == 17 || smooth_type == 18))
{
hypre_VectorSize(hypre_ParVectorLocalVector(Utemp)) = local_size;
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, Aux_F, Vtemp);
if (smooth_type == 8 || smooth_type == 18)
HYPRE_ParCSRParaSailsSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 7 || smooth_type == 17)
HYPRE_ParCSRPilutSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 9 || smooth_type == 19)
HYPRE_EuclidSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
hypre_ParVectorAxpy(relax_weight[level],Utemp,Aux_U);
}
else if (smooth_num_levels > level &&
(smooth_type == 6 || smooth_type == 16))
{
HYPRE_SchwarzSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Aux_F,
(HYPRE_ParVector) Aux_U);
}
/*else if (relax_type == 99)*/
else if (relax_type == 9 || relax_type == 99)
{ /* Gaussian elimination */
hypre_GaussElimSolve(amg_data, level, relax_type);
}
else if (relax_type == 18)
{ /* L1 - Jacobi*/
if (relax_order == 1 && cycle_param < 3)
{
/* need to do CF - so can't use the AMS one */
HYPRE_Int i;
HYPRE_Int loc_relax_points[2];
if (cycle_type < 2)
{
loc_relax_points[0] = 1;
loc_relax_points[1] = -1;
}
else
{
loc_relax_points[0] = -1;
loc_relax_points[1] = 1;
}
for (i=0; i < 2; i++)
hypre_ParCSRRelax_L1_Jacobi(A_array[level],
Aux_F,
CF_marker_array[level],
loc_relax_points[i],
relax_weight[level],
l1_norms[level],
Aux_U,
Vtemp);
}
else /* not CF - so use through AMS */
{
if (num_threads == 1)
hypre_ParCSRRelax(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],0,0,0,0,
Aux_U,
Vtemp,
Ztemp);
else
hypre_ParCSRRelaxThreads(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],
Aux_U,
Vtemp,
Ztemp);
}
}
else if (relax_type == 15)
{ /* CG */
if (j ==0) /* do num sweep iterations of CG */
hypre_ParCSRRelax_CG( smoother[level],
A_array[level],
Aux_F,
Aux_U,
num_sweep);
}
else if (relax_type == 16)
{ /* scaled Chebyshev */
HYPRE_Int scale = 1;
HYPRE_Int variant = 0;
hypre_ParCSRRelax_Cheby(A_array[level],
Aux_F,
max_eig_est[level],
min_eig_est[level],
cheby_fraction, cheby_order, scale,
variant, Aux_U, Vtemp, Ztemp );
}
else if (relax_type ==17)
{
hypre_BoomerAMGRelax_FCFJacobi(A_array[level],
Aux_F,
CF_marker_array[level],
relax_weight[level],
Aux_U,
Vtemp);
}
else if (old_version)
{
Solve_err_flag = hypre_BoomerAMGRelax(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type, relax_points,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
else
{
/* smoother than can have CF ordering */
if (block_mode)
{
Solve_err_flag = hypre_BoomerAMGBlockRelaxIF(A_block_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
Aux_U,
Vtemp);
}
else
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
}
if (Solve_err_flag != 0)
return(Solve_err_flag);
}
if (smooth_num_levels > level && smooth_type > 9)
{
gammaold = gamma;
gamma = hypre_ParVectorInnerProd(Rtemp,Ztemp);
if (jj == 0)
hypre_ParVectorCopy(Ztemp,Ptemp);
else
{
beta = gamma/gammaold;
for (i=0; i < local_size; i++)
Ptemp_data[i] = Ztemp_data[i] + beta*Ptemp_data[i];
}
hypre_ParCSRMatrixMatvec(1.0,A_array[level],Ptemp,0.0,Vtemp);
alfa = gamma /hypre_ParVectorInnerProd(Ptemp,Vtemp);
hypre_ParVectorAxpy(alfa,Ptemp,U_array[level]);
hypre_ParVectorAxpy(-alfa,Vtemp,Rtemp);
}
}
}
/*------------------------------------------------------------------
* Decrement the control counter and determine which grid to visit next
*-----------------------------------------------------------------*/
--lev_counter[level];
if (lev_counter[level] >= 0 && level != num_levels-1)
{
/*---------------------------------------------------------------
* Visit coarser level next.
* Compute residual using hypre_ParCSRMatrixMatvec.
* Perform restriction using hypre_ParCSRMatrixMatvecT.
* Reset counters and cycling parameters for coarse level
*--------------------------------------------------------------*/
fine_grid = level;
coarse_grid = level + 1;
hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0);
alpha = -1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
hypre_ParCSRBlockMatrixMatvec(alpha, A_block_array[fine_grid], U_array[fine_grid],
beta, Vtemp);
}
else
{
// JSP: avoid unnecessary copy using out-of-place version of SpMV
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[fine_grid], U_array[fine_grid],
beta, F_array[fine_grid], Vtemp);
}
alpha = 1.0;
beta = 0.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvecT(alpha,R_block_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
else
{
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
++level;
lev_counter[level] = hypre_max(lev_counter[level],cycle_type);
cycle_param = 1;
if (level == num_levels-1) cycle_param = 3;
}
else if (level != 0)
{
/*---------------------------------------------------------------
* Visit finer level next.
* Interpolate and add correction using hypre_ParCSRMatrixMatvec.
* Reset counters and cycling parameters for finer level.
*--------------------------------------------------------------*/
fine_grid = level - 1;
coarse_grid = level;
alpha = 1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvec(alpha, P_block_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
else
{
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
--level;
cycle_param = 2;
}
else
{
Not_Finished = 0;
}
}
hypre_ParAMGDataCycleOpCount(amg_data) = cycle_op_count;
hypre_TFree(lev_counter);
hypre_TFree(num_coeffs);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8 || smooth_type == 9 ||
smooth_type == 17 || smooth_type == 18 || smooth_type == 19 )
hypre_ParVectorDestroy(Utemp);
}
return(Solve_err_flag);
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
hypre_ParVector *vector1;
hypre_ParVector *vector2;
hypre_ParVector *tmp_vector;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_size = 20;
HYPRE_Int local_size;
HYPRE_Int first_index;
HYPRE_Int i;
HYPRE_Int *partitioning;
HYPRE_Complex prod;
HYPRE_Complex *data, *data2;
hypre_Vector *vector;
hypre_Vector *local_vector;
hypre_Vector *local_vector2;
/* 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, &my_id );
hypre_printf(" my_id: %d num_procs: %d\n", my_id, num_procs);
partitioning = NULL;
vector1 = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
partitioning = hypre_ParVectorPartitioning(vector1);
hypre_ParVectorInitialize(vector1);
local_vector = hypre_ParVectorLocalVector(vector1);
data = hypre_VectorData(local_vector);
local_size = hypre_VectorSize(local_vector);
first_index = partitioning[my_id];
for (i=0; i < local_size; i++)
data[i] = first_index+i;
/*
hypre_ParVectorPrint(vector1, "Vector");
*/
local_vector2 = hypre_SeqVectorCreate(global_size);
hypre_SeqVectorInitialize(local_vector2);
data2 = hypre_VectorData(local_vector2);
for (i=0; i < global_size; i++)
data2[i] = i+1;
/* partitioning = hypre_CTAlloc(HYPRE_Int,4);
partitioning[0] = 0;
partitioning[1] = 10;
partitioning[2] = 10;
partitioning[3] = 20;
*/
vector2 = hypre_VectorToParVector(hypre_MPI_COMM_WORLD,local_vector2,partitioning);
hypre_ParVectorSetPartitioningOwner(vector2,0);
hypre_ParVectorPrint(vector2, "Convert");
vector = hypre_ParVectorToVectorAll(vector2);
/*-----------------------------------------------------------
* Copy the vector into tmp_vector
*-----------------------------------------------------------*/
tmp_vector = hypre_ParVectorRead(hypre_MPI_COMM_WORLD, "Convert");
/*
tmp_vector = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
hypre_ParVectorSetPartitioningOwner(tmp_vector,0);
hypre_ParVectorInitialize(tmp_vector);
hypre_ParVectorCopy(vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Copy");
*/
/*-----------------------------------------------------------
* Scale tmp_vector
*-----------------------------------------------------------*/
hypre_ParVectorScale(2.0, tmp_vector);
/*
hypre_ParVectorPrint(tmp_vector,"Scale");
*/
/*-----------------------------------------------------------
* Do an Axpy (2*vector - vector) = vector
*-----------------------------------------------------------*/
hypre_ParVectorAxpy(-1.0, vector1, tmp_vector);
/*
hypre_ParVectorPrint(tmp_vector,"Axpy");
*/
/*-----------------------------------------------------------
* Do an inner product vector* tmp_vector
*-----------------------------------------------------------*/
prod = hypre_ParVectorInnerProd(vector1, tmp_vector);
hypre_printf (" prod: %8.2f \n", prod);
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
hypre_ParVectorDestroy(vector1);
hypre_ParVectorDestroy(vector2);
hypre_ParVectorDestroy(tmp_vector);
hypre_SeqVectorDestroy(local_vector2);
if (vector) hypre_SeqVectorDestroy(vector);
/* Finalize MPI */
hypre_MPI_Finalize();
return 0;
}
