int32_t
impl_bHYPRE_IJParCSRMatrix_SetDiagOffdSizes(
/* in */ bHYPRE_IJParCSRMatrix self,
/* in rarray[local_nrows] */ int32_t* diag_sizes,
/* in rarray[local_nrows] */ int32_t* offdiag_sizes,
/* in */ int32_t local_nrows,
/* out */ sidl_BaseInterface *_ex)
{
*_ex = 0;
{
/* DO-NOT-DELETE splicer.begin(bHYPRE.IJParCSRMatrix.SetDiagOffdSizes) */
/* Insert the implementation of the SetDiagOffdSizes method here... */
int ierr=0;
struct bHYPRE_IJParCSRMatrix__data * data;
HYPRE_IJMatrix ij_A;
data = bHYPRE_IJParCSRMatrix__get_data( self );
ij_A = data -> ij_A;
ierr = HYPRE_IJMatrixSetDiagOffdSizes( ij_A,
diag_sizes,
offdiag_sizes );
return( ierr );
/* DO-NOT-DELETE splicer.end(bHYPRE.IJParCSRMatrix.SetDiagOffdSizes) */
}
}
HYPRE_Int
hypre_ParaSailsBuildIJMatrix(hypre_ParaSails obj, HYPRE_IJMatrix *pij_A)
{
hypre_ParaSails_struct *internal = (hypre_ParaSails_struct *) obj;
ParaSails *ps = internal->ps;
Matrix *mat = internal->ps->M;
HYPRE_Int *diag_sizes, *offdiag_sizes, local_row, i, j;
HYPRE_Int size;
HYPRE_Int *col_inds;
HYPRE_Real *values;
HYPRE_IJMatrixCreate( ps->comm, ps->beg_row, ps->end_row,
ps->beg_row, ps->end_row,
pij_A );
HYPRE_IJMatrixSetObjectType( *pij_A, HYPRE_PARCSR );
diag_sizes = hypre_CTAlloc(HYPRE_Int, ps->end_row - ps->beg_row + 1);
offdiag_sizes = hypre_CTAlloc(HYPRE_Int, ps->end_row - ps->beg_row + 1);
local_row = 0;
for (i=ps->beg_row; i<= ps->end_row; i++)
{
MatrixGetRow(mat, local_row, &size, &col_inds, &values);
NumberingLocalToGlobal(ps->numb, size, col_inds, col_inds);
for (j=0; j < size; j++)
{
if (col_inds[j] < ps->beg_row || col_inds[j] > ps->end_row)
offdiag_sizes[local_row]++;
else
diag_sizes[local_row]++;
}
local_row++;
}
HYPRE_IJMatrixSetDiagOffdSizes( *pij_A, (const HYPRE_Int *) diag_sizes,
(const HYPRE_Int *) offdiag_sizes );
hypre_TFree(diag_sizes);
hypre_TFree(offdiag_sizes);
HYPRE_IJMatrixInitialize( *pij_A );
local_row = 0;
for (i=ps->beg_row; i<= ps->end_row; i++)
{
MatrixGetRow(mat, local_row, &size, &col_inds, &values);
HYPRE_IJMatrixSetValues( *pij_A, 1, &size, &i, (const HYPRE_Int *) col_inds,
(const HYPRE_Real *) values );
NumberingGlobalToLocal(ps->numb, size, col_inds, col_inds);
local_row++;
}
HYPRE_IJMatrixAssemble( *pij_A );
return hypre_error_flag;
}
inline void numfact(unsigned int ncol, int* I, int* loc2glob, int* J, K* C) {
static_assert(std::is_same<double, K>::value, "Hypre only supports double-precision floating-point real numbers");
static_assert(S == 'G', "Hypre only supports nonsymmetric matrices");
HYPRE_IJMatrixCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], loc2glob[0], loc2glob[1], &_A);
HYPRE_IJMatrixSetObjectType(_A, HYPRE_PARCSR);
HYPRE_IJMatrixSetRowSizes(_A, I + 1);
_local = ncol;
int* rows = new int[3 * _local]();
int* diag_sizes = rows + _local;
int* offdiag_sizes = diag_sizes + _local;
rows[0] = I[0];
for(unsigned int i = 0; i < _local; ++i) {
std::for_each(J + rows[0], J + rows[0] + I[i + 1], [&](int& j) { (j < loc2glob[0] || loc2glob[1] < j) ? ++offdiag_sizes[i] : ++diag_sizes[i]; });
rows[0] += I[i + 1];
}
HYPRE_IJMatrixSetDiagOffdSizes(_A, diag_sizes, offdiag_sizes);
HYPRE_IJMatrixSetMaxOffProcElmts(_A, 0);
HYPRE_IJMatrixInitialize(_A);
std::iota(rows, rows + _local, loc2glob[0]);
HYPRE_IJMatrixSetValues(_A, _local, I + 1, rows, J, C);
HYPRE_IJMatrixAssemble(_A);
HYPRE_IJVectorCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], &_b);
HYPRE_IJVectorSetObjectType(_b, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(_b);
HYPRE_IJVectorCreate(DMatrix::_communicator, loc2glob[0], loc2glob[1], &_x);
HYPRE_IJVectorSetObjectType(_x, HYPRE_PARCSR);
HYPRE_IJVectorInitialize(_x);
delete [] rows;
delete [] I;
delete [] loc2glob;
HYPRE_BoomerAMGCreate(_strategy == 1 ? &_solver : &_precond);
HYPRE_BoomerAMGSetCoarsenType(_strategy == 1 ? _solver : _precond, 6); /* Falgout coarsening */
HYPRE_BoomerAMGSetRelaxType(_strategy == 1 ? _solver : _precond, 6); /* G-S/Jacobi hybrid relaxation */
HYPRE_BoomerAMGSetNumSweeps(_strategy == 1 ? _solver : _precond, 1); /* sweeps on each level */
HYPRE_BoomerAMGSetMaxLevels(_strategy == 1 ? _solver : _precond, 10); /* maximum number of levels */
HYPRE_ParCSRMatrix parcsr_A;
HYPRE_IJMatrixGetObject(_A, reinterpret_cast<void**>(&parcsr_A));
HYPRE_ParVector par_b;
HYPRE_IJVectorGetObject(_b, reinterpret_cast<void**>(&par_b));
HYPRE_ParVector par_x;
HYPRE_IJVectorGetObject(_x, reinterpret_cast<void**>(&par_x));
if(_strategy == 1) {
HYPRE_BoomerAMGSetTol(_solver, 1.0e-8);
HYPRE_BoomerAMGSetMaxIter(_solver, 1000);
HYPRE_BoomerAMGSetPrintLevel(_solver, 1);
HYPRE_BoomerAMGSetup(_solver, parcsr_A, nullptr, nullptr);
}
else {
HYPRE_BoomerAMGSetTol(_precond, 0.0);
HYPRE_BoomerAMGSetMaxIter(_precond, 1);
HYPRE_BoomerAMGSetPrintLevel(_precond, 1);
if(_strategy == 2) {
HYPRE_ParCSRPCGCreate(DMatrix::_communicator, &_solver);
HYPRE_PCGSetMaxIter(_solver, 500);
HYPRE_PCGSetTol(_solver, 1.0e-8);
HYPRE_PCGSetTwoNorm(_solver, 1);
HYPRE_PCGSetPrintLevel(_solver, 1);
HYPRE_PCGSetLogging(_solver, 1);
HYPRE_PCGSetPrecond(_solver, reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSolve), reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSetup), _precond);
HYPRE_ParCSRPCGSetup(_solver, parcsr_A, par_b, par_x);
}
else {
HYPRE_ParCSRFlexGMRESCreate(DMatrix::_communicator, &_solver);
HYPRE_FlexGMRESSetKDim(_solver, 50);
HYPRE_FlexGMRESSetMaxIter(_solver, 500);
HYPRE_FlexGMRESSetTol(_solver, 1.0e-8);
HYPRE_FlexGMRESSetPrintLevel(_solver, 1);
HYPRE_FlexGMRESSetLogging(_solver, 1);
HYPRE_FlexGMRESSetPrecond(_solver, reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSolve), reinterpret_cast<HYPRE_PtrToSolverFcn>(HYPRE_BoomerAMGSetup), _precond);
HYPRE_ParCSRFlexGMRESSetup(_solver, parcsr_A, par_b, par_x);
}
}
}
HYPRE_Int AmgCGCGraphAssemble (hypre_ParCSRMatrix *S,HYPRE_Int *vertexrange,HYPRE_Int *CF_marker,HYPRE_Int *CF_marker_offd,HYPRE_Int coarsen_type,
HYPRE_IJMatrix *ijG)
/* assemble a graph representing the connections between the grids
* ================================================================================================
* S : the strength matrix
* vertexrange : the parallel layout of the candidate coarse grid vertices
* CF_marker, CF_marker_offd : the coarse/fine markers
* coarsen_type : the coarsening type
* ijG : the created graph
* ================================================================================================*/
{
HYPRE_Int ierr=0;
HYPRE_Int i,/* ii,*/ip,j,jj,m,n,p;
HYPRE_Int mpisize,mpirank;
HYPRE_Real weight;
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
/* hypre_MPI_Status status; */
HYPRE_IJMatrix ijmatrix;
hypre_CSRMatrix *S_diag = hypre_ParCSRMatrixDiag (S);
hypre_CSRMatrix *S_offd = hypre_ParCSRMatrixOffd (S);
/* HYPRE_Int *S_i = hypre_CSRMatrixI(S_diag); */
/* HYPRE_Int *S_j = hypre_CSRMatrixJ(S_diag); */
HYPRE_Int *S_offd_i = hypre_CSRMatrixI(S_offd);
HYPRE_Int *S_offd_j = NULL;
HYPRE_Int num_variables = hypre_CSRMatrixNumRows (S_diag);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols (S_offd);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd (S);
HYPRE_Int pointrange_start,pointrange_end;
HYPRE_Int *pointrange,*pointrange_nonlocal,*pointrange_strong=NULL;
HYPRE_Int vertexrange_start,vertexrange_end;
HYPRE_Int *vertexrange_strong= NULL;
HYPRE_Int *vertexrange_nonlocal;
HYPRE_Int num_recvs,num_recvs_strong;
HYPRE_Int *recv_procs,*recv_procs_strong=NULL;
HYPRE_Int /* *zeros,*rownz,*/*rownz_diag,*rownz_offd;
HYPRE_Int nz;
HYPRE_Int nlocal;
HYPRE_Int one=1;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg (S);
hypre_MPI_Comm_size (comm,&mpisize);
hypre_MPI_Comm_rank (comm,&mpirank);
/* determine neighbor processors */
num_recvs = hypre_ParCSRCommPkgNumRecvs (comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs (comm_pkg);
pointrange = hypre_ParCSRMatrixRowStarts (S);
pointrange_nonlocal = hypre_CTAlloc (HYPRE_Int, 2*num_recvs);
vertexrange_nonlocal = hypre_CTAlloc (HYPRE_Int, 2*num_recvs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
{
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends (comm_pkg);
HYPRE_Int *send_procs = hypre_ParCSRCommPkgSendProcs (comm_pkg);
HYPRE_Int *int_buf_data = hypre_CTAlloc (HYPRE_Int,4*num_sends);
HYPRE_Int *int_buf_data2 = int_buf_data + 2*num_sends;
hypre_MPI_Request *sendrequest,*recvrequest;
nlocal = vertexrange[1] - vertexrange[0];
pointrange_start = pointrange[0];
pointrange_end = pointrange[1];
vertexrange_start = vertexrange[0];
vertexrange_end = vertexrange[1];
sendrequest = hypre_CTAlloc (hypre_MPI_Request,2*(num_sends+num_recvs));
recvrequest = sendrequest+2*num_sends;
for (i=0;i<num_recvs;i++) {
hypre_MPI_Irecv (pointrange_nonlocal+2*i,2,HYPRE_MPI_INT,recv_procs[i],tag_pointrange,comm,&recvrequest[2*i]);
hypre_MPI_Irecv (vertexrange_nonlocal+2*i,2,HYPRE_MPI_INT,recv_procs[i],tag_vertexrange,comm,&recvrequest[2*i+1]);
}
for (i=0;i<num_sends;i++) {
int_buf_data[2*i] = pointrange_start;
int_buf_data[2*i+1] = pointrange_end;
int_buf_data2[2*i] = vertexrange_start;
int_buf_data2[2*i+1] = vertexrange_end;
hypre_MPI_Isend (int_buf_data+2*i,2,HYPRE_MPI_INT,send_procs[i],tag_pointrange,comm,&sendrequest[2*i]);
hypre_MPI_Isend (int_buf_data2+2*i,2,HYPRE_MPI_INT,send_procs[i],tag_vertexrange,comm,&sendrequest[2*i+1]);
}
hypre_MPI_Waitall (2*(num_sends+num_recvs),sendrequest,hypre_MPI_STATUSES_IGNORE);
hypre_TFree (int_buf_data);
hypre_TFree (sendrequest);
}
#else
nlocal = vertexrange[mpirank+1] - vertexrange[mpirank];
pointrange_start = pointrange[mpirank];
pointrange_end = pointrange[mpirank+1];
vertexrange_start = vertexrange[mpirank];
vertexrange_end = vertexrange[mpirank+1];
for (i=0;i<num_recvs;i++) {
pointrange_nonlocal[2*i] = pointrange[recv_procs[i]];
pointrange_nonlocal[2*i+1] = pointrange[recv_procs[i]+1];
vertexrange_nonlocal[2*i] = vertexrange[recv_procs[i]];
vertexrange_nonlocal[2*i+1] = vertexrange[recv_procs[i]+1];
}
#endif
/* now we have the array recv_procs. However, it may contain too many entries as it is
inherited from A. We now have to determine the subset which contains only the
strongly connected neighbors */
if (num_cols_offd) {
S_offd_j = hypre_CSRMatrixJ(S_offd);
recv_procs_strong = hypre_CTAlloc (HYPRE_Int,num_recvs);
memset (recv_procs_strong,0,num_recvs*sizeof(HYPRE_Int));
/* don't forget to shorten the pointrange and vertexrange arrays accordingly */
pointrange_strong = hypre_CTAlloc (HYPRE_Int,2*num_recvs);
memset (pointrange_strong,0,2*num_recvs*sizeof(HYPRE_Int));
vertexrange_strong = hypre_CTAlloc (HYPRE_Int,2*num_recvs);
memset (vertexrange_strong,0,2*num_recvs*sizeof(HYPRE_Int));
for (i=0;i<num_variables;i++)
for (j=S_offd_i[i];j<S_offd_i[i+1];j++) {
jj = col_map_offd[S_offd_j[j]];
for (p=0;p<num_recvs;p++) /* S_offd_j is NOT sorted! */
if (jj >= pointrange_nonlocal[2*p] && jj < pointrange_nonlocal[2*p+1]) break;
#if 0
hypre_printf ("Processor %d, remote point %d on processor %d\n",mpirank,jj,recv_procs[p]);
#endif
recv_procs_strong [p]=1;
}
for (p=0,num_recvs_strong=0;p<num_recvs;p++) {
if (recv_procs_strong[p]) {
recv_procs_strong[num_recvs_strong]=recv_procs[p];
pointrange_strong[2*num_recvs_strong] = pointrange_nonlocal[2*p];
pointrange_strong[2*num_recvs_strong+1] = pointrange_nonlocal[2*p+1];
vertexrange_strong[2*num_recvs_strong] = vertexrange_nonlocal[2*p];
vertexrange_strong[2*num_recvs_strong+1] = vertexrange_nonlocal[2*p+1];
num_recvs_strong++;
}
}
}
else num_recvs_strong=0;
hypre_TFree (pointrange_nonlocal);
hypre_TFree (vertexrange_nonlocal);
rownz_diag = hypre_CTAlloc (HYPRE_Int,2*nlocal);
rownz_offd = rownz_diag + nlocal;
for (p=0,nz=0;p<num_recvs_strong;p++) {
nz += vertexrange_strong[2*p+1]-vertexrange_strong[2*p];
}
for (m=0;m<nlocal;m++) {
rownz_diag[m]=nlocal-1;
rownz_offd[m]=nz;
}
HYPRE_IJMatrixCreate(comm, vertexrange_start, vertexrange_end-1, vertexrange_start, vertexrange_end-1, &ijmatrix);
HYPRE_IJMatrixSetObjectType(ijmatrix, HYPRE_PARCSR);
HYPRE_IJMatrixSetDiagOffdSizes (ijmatrix, rownz_diag, rownz_offd);
HYPRE_IJMatrixInitialize(ijmatrix);
hypre_TFree (rownz_diag);
/* initialize graph */
weight = -1;
for (m=vertexrange_start;m<vertexrange_end;m++) {
for (p=0;p<num_recvs_strong;p++) {
for (n=vertexrange_strong[2*p];n<vertexrange_strong[2*p+1];n++) {
ierr = HYPRE_IJMatrixAddToValues (ijmatrix,1,&one,&m,&n,&weight);
#if 0
if (ierr) hypre_printf ("Processor %d: error %d while initializing graphs at (%d, %d)\n",mpirank,ierr,m,n);
#endif
}
}
}
/* weight graph */
for (i=0;i<num_variables;i++) {
for (j=S_offd_i[i];j<S_offd_i[i+1];j++) {
jj = S_offd_j[j]; /* jj is not a global index!!! */
/* determine processor */
for (p=0;p<num_recvs_strong;p++)
if (col_map_offd[jj] >= pointrange_strong[2*p] && col_map_offd[jj] < pointrange_strong[2*p+1]) break;
ip=recv_procs_strong[p];
/* loop over all coarse grids constructed on this processor domain */
for (m=vertexrange_start;m<vertexrange_end;m++) {
/* loop over all coarse grids constructed on neighbor processor domain */
for (n=vertexrange_strong[2*p];n<vertexrange_strong[2*p+1];n++) {
/* coarse grid counting inside gridpartition->local/gridpartition->nonlocal starts with one
while counting inside range starts with zero */
if (CF_marker[i]-1==m && CF_marker_offd[jj]-1==n)
/* C-C-coupling */
weight = -1;
else if ( (CF_marker[i]-1==m && (CF_marker_offd[jj]==0 || CF_marker_offd[jj]-1!=n) )
|| ( (CF_marker[i]==0 || CF_marker[i]-1!=m) && CF_marker_offd[jj]-1==n ) )
/* C-F-coupling */
weight = 0;
else weight = -8; /* F-F-coupling */
ierr = HYPRE_IJMatrixAddToValues (ijmatrix,1,&one,&m,&n,&weight);
#if 0
if (ierr) hypre_printf ("Processor %d: error %d while adding %lf to entry (%d, %d)\n",mpirank,ierr,weight,m,n);
#endif
}
}
}
}
/* assemble */
HYPRE_IJMatrixAssemble (ijmatrix);
/*if (num_recvs_strong) {*/
hypre_TFree (recv_procs_strong);
hypre_TFree (pointrange_strong);
hypre_TFree (vertexrange_strong);
/*} */
*ijG = ijmatrix;
return (ierr);
}
