HYPRE_Int
hypre_ParCSRCommMultiHandleDestroy(hypre_ParCSRCommMultiHandle *comm_handle)
{
hypre_MPI_Status *status0;
HYPRE_Int ierr = 0;
if (hypre_ParCSRCommMultiHandleNumRequests(comm_handle))
{
status0 = hypre_CTAlloc(hypre_MPI_Status,
hypre_ParCSRCommMultiHandleNumRequests(comm_handle));
hypre_MPI_Waitall(hypre_ParCSRCommMultiHandleNumRequests(comm_handle),
hypre_ParCSRCommMultiHandleRequests(comm_handle), status0);
hypre_TFree(status0);
}
hypre_TFree(hypre_ParCSRCommMultiHandleRequests(comm_handle));
hypre_TFree(comm_handle);
return ierr;
}
hypre_ParCSRBooleanMatrix *
hypre_CSRBooleanMatrixToParCSRBooleanMatrix
( MPI_Comm comm, hypre_CSRBooleanMatrix *A,
HYPRE_Int *row_starts, HYPRE_Int *col_starts )
{
HYPRE_Int global_data[2];
HYPRE_Int global_num_rows;
HYPRE_Int global_num_cols;
HYPRE_Int *local_num_rows;
HYPRE_Int num_procs, my_id;
HYPRE_Int *local_num_nonzeros;
HYPRE_Int num_nonzeros;
HYPRE_Int *a_i;
HYPRE_Int *a_j;
hypre_CSRBooleanMatrix *local_A;
hypre_MPI_Request *requests;
hypre_MPI_Status *status, status0;
hypre_MPI_Datatype *csr_matrix_datatypes;
hypre_ParCSRBooleanMatrix *par_matrix;
HYPRE_Int first_col_diag;
HYPRE_Int last_col_diag;
HYPRE_Int i, j, ind;
hypre_MPI_Comm_rank(comm, &my_id);
hypre_MPI_Comm_size(comm, &num_procs);
if (my_id == 0)
{
global_data[0] = hypre_CSRBooleanMatrix_Get_NRows(A);
global_data[1] = hypre_CSRBooleanMatrix_Get_NCols(A);
a_i = hypre_CSRBooleanMatrix_Get_I(A);
a_j = hypre_CSRBooleanMatrix_Get_J(A);
}
hypre_MPI_Bcast(global_data,2,HYPRE_MPI_INT,0,comm);
global_num_rows = global_data[0];
global_num_cols = global_data[1];
local_num_rows = hypre_CTAlloc(HYPRE_Int, num_procs);
csr_matrix_datatypes = hypre_CTAlloc(hypre_MPI_Datatype, num_procs);
par_matrix = hypre_ParCSRBooleanMatrixCreate (comm, global_num_rows,
global_num_cols,row_starts,col_starts,0,0,0);
row_starts = hypre_ParCSRBooleanMatrix_Get_RowStarts(par_matrix);
col_starts = hypre_ParCSRBooleanMatrix_Get_ColStarts(par_matrix);
for (i=0; i < num_procs; i++)
local_num_rows[i] = row_starts[i+1] - row_starts[i];
if (my_id == 0)
{
local_num_nonzeros = hypre_CTAlloc(HYPRE_Int, num_procs);
for (i=0; i < num_procs-1; i++)
local_num_nonzeros[i] = a_i[row_starts[i+1]]
- a_i[row_starts[i]];
local_num_nonzeros[num_procs-1] = a_i[global_num_rows]
- a_i[row_starts[num_procs-1]];
}
hypre_MPI_Scatter(local_num_nonzeros,1,HYPRE_MPI_INT,&num_nonzeros,1,HYPRE_MPI_INT,0,comm);
if (my_id == 0) num_nonzeros = local_num_nonzeros[0];
local_A = hypre_CSRBooleanMatrixCreate(local_num_rows[my_id], global_num_cols,
num_nonzeros);
if (my_id == 0)
{
requests = hypre_CTAlloc (hypre_MPI_Request, num_procs-1);
status = hypre_CTAlloc(hypre_MPI_Status, num_procs-1);
j=0;
for (i=1; i < num_procs; i++)
{
ind = a_i[row_starts[i]];
hypre_BuildCSRBooleanMatrixMPIDataType(local_num_nonzeros[i],
local_num_rows[i],
&a_i[row_starts[i]],
&a_j[ind],
&csr_matrix_datatypes[i]);
hypre_MPI_Isend(hypre_MPI_BOTTOM, 1, csr_matrix_datatypes[i], i, 0, comm,
&requests[j++]);
hypre_MPI_Type_free(&csr_matrix_datatypes[i]);
}
hypre_CSRBooleanMatrix_Get_I(local_A) = a_i;
hypre_CSRBooleanMatrix_Get_J(local_A) = a_j;
hypre_MPI_Waitall(num_procs-1,requests,status);
hypre_TFree(requests);
hypre_TFree(status);
hypre_TFree(local_num_nonzeros);
}
else
{
hypre_CSRBooleanMatrixInitialize(local_A);
hypre_BuildCSRBooleanMatrixMPIDataType(num_nonzeros,
local_num_rows[my_id],
hypre_CSRBooleanMatrix_Get_I(local_A),
hypre_CSRBooleanMatrix_Get_J(local_A),
csr_matrix_datatypes);
hypre_MPI_Recv(hypre_MPI_BOTTOM,1,csr_matrix_datatypes[0],0,0,comm,&status0);
hypre_MPI_Type_free(csr_matrix_datatypes);
}
first_col_diag = col_starts[my_id];
last_col_diag = col_starts[my_id+1]-1;
BooleanGenerateDiagAndOffd(local_A, par_matrix, first_col_diag, last_col_diag);
/* set pointers back to NULL before destroying */
if (my_id == 0)
{
hypre_CSRBooleanMatrix_Get_I(local_A) = NULL;
hypre_CSRBooleanMatrix_Get_J(local_A) = NULL;
}
hypre_CSRBooleanMatrixDestroy(local_A);
hypre_TFree(local_num_rows);
hypre_TFree(csr_matrix_datatypes);
return par_matrix;
}
hypre_ParVector *
hypre_VectorToParVector (MPI_Comm comm, hypre_Vector *v, HYPRE_Int *vec_starts)
{
HYPRE_Int global_size;
HYPRE_Int local_size;
HYPRE_Int num_vectors;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_vecstride, vecstride, idxstride;
hypre_ParVector *par_vector;
hypre_Vector *local_vector;
double *v_data;
double *local_data;
hypre_MPI_Request *requests;
hypre_MPI_Status *status, status0;
HYPRE_Int i, j, k, p;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
if (my_id == 0)
{
global_size = hypre_VectorSize(v);
v_data = hypre_VectorData(v);
num_vectors = hypre_VectorNumVectors(v); /* for multivectors */
global_vecstride = hypre_VectorVectorStride(v);
}
hypre_MPI_Bcast(&global_size,1,HYPRE_MPI_INT,0,comm);
hypre_MPI_Bcast(&num_vectors,1,HYPRE_MPI_INT,0,comm);
hypre_MPI_Bcast(&global_vecstride,1,HYPRE_MPI_INT,0,comm);
if ( num_vectors==1 )
par_vector = hypre_ParVectorCreate(comm, global_size, vec_starts);
else
par_vector = hypre_ParMultiVectorCreate(comm, global_size, vec_starts, num_vectors);
vec_starts = hypre_ParVectorPartitioning(par_vector);
local_size = vec_starts[my_id+1] - vec_starts[my_id];
hypre_ParVectorInitialize(par_vector);
local_vector = hypre_ParVectorLocalVector(par_vector);
local_data = hypre_VectorData(local_vector);
vecstride = hypre_VectorVectorStride(local_vector);
idxstride = hypre_VectorIndexStride(local_vector);
hypre_assert( idxstride==1 ); /* <<< so far only the only implemented multivector StorageMethod is 0 <<< */
if (my_id == 0)
{
requests = hypre_CTAlloc(hypre_MPI_Request,num_vectors*(num_procs-1));
status = hypre_CTAlloc(hypre_MPI_Status,num_vectors*(num_procs-1));
k = 0;
for ( p=1; p<num_procs; p++)
for ( j=0; j<num_vectors; ++j )
{
hypre_MPI_Isend( &v_data[vec_starts[p]]+j*global_vecstride,
(vec_starts[p+1]-vec_starts[p]),
hypre_MPI_DOUBLE, p, 0, comm, &requests[k++] );
}
if ( num_vectors==1 )
{
for (i=0; i < local_size; i++)
local_data[i] = v_data[i];
}
else
for ( j=0; j<num_vectors; ++j )
{
for (i=0; i < local_size; i++)
local_data[i+j*vecstride] = v_data[i+j*global_vecstride];
}
hypre_MPI_Waitall(num_procs-1,requests, status);
hypre_TFree(requests);
hypre_TFree(status);
}
else
{
for ( j=0; j<num_vectors; ++j )
hypre_MPI_Recv( local_data+j*vecstride, local_size, hypre_MPI_DOUBLE, 0, 0, comm,&status0 );
}
return par_vector;
}
hypre_Vector *
hypre_ParVectorToVectorAll (hypre_ParVector *par_v)
{
MPI_Comm comm = hypre_ParVectorComm(par_v);
HYPRE_Int global_size = hypre_ParVectorGlobalSize(par_v);
#ifndef HYPRE_NO_GLOBAL_PARTITION
HYPRE_Int *vec_starts = hypre_ParVectorPartitioning(par_v);
#endif
hypre_Vector *local_vector = hypre_ParVectorLocalVector(par_v);
HYPRE_Int num_procs, my_id;
HYPRE_Int num_vectors = hypre_ParVectorNumVectors(par_v);
hypre_Vector *vector;
double *vector_data;
double *local_data;
HYPRE_Int local_size;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int i, j;
HYPRE_Int *used_procs;
HYPRE_Int num_types, num_requests;
HYPRE_Int vec_len, proc_id;
#ifdef HYPRE_NO_GLOBAL_PARTITION
HYPRE_Int *new_vec_starts;
HYPRE_Int num_contacts;
HYPRE_Int contact_proc_list[1];
HYPRE_Int contact_send_buf[1];
HYPRE_Int contact_send_buf_starts[2];
HYPRE_Int max_response_size;
HYPRE_Int *response_recv_buf=NULL;
HYPRE_Int *response_recv_buf_starts = NULL;
hypre_DataExchangeResponse response_obj;
hypre_ProcListElements send_proc_obj;
HYPRE_Int *send_info = NULL;
hypre_MPI_Status status1;
HYPRE_Int count, tag1 = 112, tag2 = 223;
HYPRE_Int start;
#endif
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm, &my_id);
#ifdef HYPRE_NO_GLOBAL_PARTITION
local_size = hypre_ParVectorLastIndex(par_v) -
hypre_ParVectorFirstIndex(par_v) + 1;
/* determine procs which hold data of par_v and store ids in used_procs */
/* we need to do an exchange data for this. If I own row then I will contact
processor 0 with the endpoint of my local range */
if (local_size > 0)
{
num_contacts = 1;
contact_proc_list[0] = 0;
contact_send_buf[0] = hypre_ParVectorLastIndex(par_v);
contact_send_buf_starts[0] = 0;
contact_send_buf_starts[1] = 1;
}
else
{
num_contacts = 0;
contact_send_buf_starts[0] = 0;
contact_send_buf_starts[1] = 0;
}
/*build the response object*/
/*send_proc_obj will be for saving info from contacts */
send_proc_obj.length = 0;
send_proc_obj.storage_length = 10;
send_proc_obj.id = hypre_CTAlloc(HYPRE_Int, send_proc_obj.storage_length);
send_proc_obj.vec_starts = hypre_CTAlloc(HYPRE_Int, send_proc_obj.storage_length + 1);
send_proc_obj.vec_starts[0] = 0;
send_proc_obj.element_storage_length = 10;
send_proc_obj.elements = hypre_CTAlloc(HYPRE_Int, send_proc_obj.element_storage_length);
max_response_size = 0; /* each response is null */
response_obj.fill_response = hypre_FillResponseParToVectorAll;
response_obj.data1 = NULL;
response_obj.data2 = &send_proc_obj; /*this is where we keep info from contacts*/
hypre_DataExchangeList(num_contacts,
contact_proc_list, contact_send_buf,
contact_send_buf_starts, sizeof(HYPRE_Int),
sizeof(HYPRE_Int), &response_obj,
max_response_size, 1,
comm, (void**) &response_recv_buf,
&response_recv_buf_starts);
/* now processor 0 should have a list of ranges for processors that have rows -
these are in send_proc_obj - it needs to create the new list of processors
and also an array of vec starts - and send to those who own row*/
if (my_id)
{
if (local_size)
{
/* look for a message from processor 0 */
hypre_MPI_Probe(0, tag1, comm, &status1);
hypre_MPI_Get_count(&status1, HYPRE_MPI_INT, &count);
send_info = hypre_CTAlloc(HYPRE_Int, count);
hypre_MPI_Recv(send_info, count, HYPRE_MPI_INT, 0, tag1, comm, &status1);
/* now unpack */
num_types = send_info[0];
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
new_vec_starts = hypre_CTAlloc(HYPRE_Int, num_types+1);
for (i=1; i<= num_types; i++)
{
used_procs[i-1] = send_info[i];
}
for (i=num_types+1; i< count; i++)
{
new_vec_starts[i-num_types-1] = send_info[i] ;
}
}
else /* clean up and exit */
{
hypre_TFree(send_proc_obj.vec_starts);
hypre_TFree(send_proc_obj.id);
hypre_TFree(send_proc_obj.elements);
if(response_recv_buf) hypre_TFree(response_recv_buf);
if(response_recv_buf_starts) hypre_TFree(response_recv_buf_starts);
return NULL;
}
}
else /* my_id ==0 */
{
num_types = send_proc_obj.length;
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
new_vec_starts = hypre_CTAlloc(HYPRE_Int, num_types+1);
new_vec_starts[0] = 0;
for (i=0; i< num_types; i++)
{
used_procs[i] = send_proc_obj.id[i];
new_vec_starts[i+1] = send_proc_obj.elements[i]+1;
}
qsort0(used_procs, 0, num_types-1);
qsort0(new_vec_starts, 0, num_types);
/*now we need to put into an array to send */
count = 2*num_types+2;
send_info = hypre_CTAlloc(HYPRE_Int, count);
send_info[0] = num_types;
for (i=1; i<= num_types; i++)
{
send_info[i] = used_procs[i-1];
}
for (i=num_types+1; i< count; i++)
{
send_info[i] = new_vec_starts[i-num_types-1];
}
requests = hypre_CTAlloc(hypre_MPI_Request, num_types);
status = hypre_CTAlloc(hypre_MPI_Status, num_types);
/* don't send to myself - these are sorted so my id would be first*/
start = 0;
if (used_procs[0] == 0)
{
start = 1;
}
for (i=start; i < num_types; i++)
{
hypre_MPI_Isend(send_info, count, HYPRE_MPI_INT, used_procs[i], tag1, comm, &requests[i-start]);
}
hypre_MPI_Waitall(num_types-start, requests, status);
hypre_TFree(status);
hypre_TFree(requests);
}
/* clean up */
hypre_TFree(send_proc_obj.vec_starts);
hypre_TFree(send_proc_obj.id);
hypre_TFree(send_proc_obj.elements);
hypre_TFree(send_info);
if(response_recv_buf) hypre_TFree(response_recv_buf);
if(response_recv_buf_starts) hypre_TFree(response_recv_buf_starts);
/* now proc 0 can exit if it has no rows */
if (!local_size) {
hypre_TFree(used_procs);
hypre_TFree(new_vec_starts);
return NULL;
}
/* everyone left has rows and knows: new_vec_starts, num_types, and used_procs */
/* this vector should be rather small */
local_data = hypre_VectorData(local_vector);
vector = hypre_SeqVectorCreate(global_size);
hypre_VectorNumVectors(vector) = num_vectors;
hypre_SeqVectorInitialize(vector);
vector_data = hypre_VectorData(vector);
num_requests = 2*num_types;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
/* initialize data exchange among used_procs and generate vector - here we
send to ourself also*/
j = 0;
for (i = 0; i < num_types; i++)
{
proc_id = used_procs[i];
vec_len = new_vec_starts[i+1] - new_vec_starts[i];
hypre_MPI_Irecv(&vector_data[new_vec_starts[i]], num_vectors*vec_len, hypre_MPI_DOUBLE,
proc_id, tag2, comm, &requests[j++]);
}
for (i = 0; i < num_types; i++)
{
hypre_MPI_Isend(local_data, num_vectors*local_size, hypre_MPI_DOUBLE, used_procs[i],
tag2, comm, &requests[j++]);
}
hypre_MPI_Waitall(num_requests, requests, status);
if (num_requests)
{
hypre_TFree(requests);
hypre_TFree(status);
hypre_TFree(used_procs);
}
hypre_TFree(new_vec_starts);
#else
local_size = vec_starts[my_id+1] - vec_starts[my_id];
/* if my_id contains no data, return NULL */
if (!local_size)
return NULL;
local_data = hypre_VectorData(local_vector);
vector = hypre_SeqVectorCreate(global_size);
hypre_VectorNumVectors(vector) = num_vectors;
hypre_SeqVectorInitialize(vector);
vector_data = hypre_VectorData(vector);
/* determine procs which hold data of par_v and store ids in used_procs */
num_types = -1;
for (i=0; i < num_procs; i++)
if (vec_starts[i+1]-vec_starts[i])
num_types++;
num_requests = 2*num_types;
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
j = 0;
for (i=0; i < num_procs; i++)
if (vec_starts[i+1]-vec_starts[i] && i-my_id)
used_procs[j++] = i;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
/* initialize data exchange among used_procs and generate vector */
j = 0;
for (i = 0; i < num_types; i++)
{
proc_id = used_procs[i];
vec_len = vec_starts[proc_id+1] - vec_starts[proc_id];
hypre_MPI_Irecv(&vector_data[vec_starts[proc_id]], num_vectors*vec_len, hypre_MPI_DOUBLE,
proc_id, 0, comm, &requests[j++]);
}
for (i = 0; i < num_types; i++)
{
hypre_MPI_Isend(local_data, num_vectors*local_size, hypre_MPI_DOUBLE, used_procs[i],
0, comm, &requests[j++]);
}
for (i=0; i < num_vectors*local_size; i++)
vector_data[vec_starts[my_id]+i] = local_data[i];
hypre_MPI_Waitall(num_requests, requests, status);
if (num_requests)
{
hypre_TFree(used_procs);
hypre_TFree(requests);
hypre_TFree(status);
}
#endif
return vector;
}
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);
}
HYPRE_Int
hypre_GenerateSendMapAndCommPkg(MPI_Comm comm, HYPRE_Int num_sends, HYPRE_Int num_recvs,
HYPRE_Int *recv_procs, HYPRE_Int *send_procs,
HYPRE_Int *recv_vec_starts, hypre_ParCSRMatrix *A)
{
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int i, j;
HYPRE_Int num_requests = num_sends+num_recvs;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int vec_len, vec_start;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(A);
/*--------------------------------------------------------------------------
* generate send_map_starts and send_map_elmts
*--------------------------------------------------------------------------*/
requests = hypre_CTAlloc(hypre_MPI_Request,num_requests);
status = hypre_CTAlloc(hypre_MPI_Status,num_requests);
send_map_starts = hypre_CTAlloc(HYPRE_Int, num_sends+1);
j = 0;
for (i=0; i < num_sends; i++)
hypre_MPI_Irecv(&send_map_starts[i+1],1,HYPRE_MPI_INT,send_procs[i],0,comm,
&requests[j++]);
for (i=0; i < num_recvs; i++)
{
vec_len = recv_vec_starts[i+1] - recv_vec_starts[i];
hypre_MPI_Isend(&vec_len,1,HYPRE_MPI_INT, recv_procs[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
send_map_starts[0] = 0;
for (i=0; i < num_sends; i++)
send_map_starts[i+1] += send_map_starts[i];
send_map_elmts = hypre_CTAlloc(HYPRE_Int,send_map_starts[num_sends]);
j = 0;
for (i=0; i < num_sends; i++)
{
vec_start = send_map_starts[i];
vec_len = send_map_starts[i+1]-vec_start;
hypre_MPI_Irecv(&send_map_elmts[vec_start],vec_len,HYPRE_MPI_INT,
send_procs[i],0,comm,&requests[j++]);
}
for (i=0; i < num_recvs; i++)
{
vec_start = recv_vec_starts[i];
vec_len = recv_vec_starts[i+1] - vec_start;
hypre_MPI_Isend(&col_map_offd[vec_start],vec_len,HYPRE_MPI_INT,
recv_procs[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
for (i=0; i < send_map_starts[num_sends]; i++)
send_map_elmts[i] -= first_col_diag;
comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(comm_pkg) = num_sends;
hypre_ParCSRCommPkgNumRecvs(comm_pkg) = num_recvs;
hypre_ParCSRCommPkgSendProcs(comm_pkg) = send_procs;
hypre_ParCSRCommPkgRecvProcs(comm_pkg) = recv_procs;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg) = recv_vec_starts;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg) = send_map_starts;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = send_map_elmts;
hypre_TFree(status);
hypre_TFree(requests);
hypre_ParCSRMatrixCommPkg(A) = comm_pkg;
return 0;
}
HYPRE_Int
hypre_GetCommPkgRTFromCommPkgA( hypre_ParCSRMatrix *RT,
hypre_ParCSRMatrix *A,
HYPRE_Int *fine_to_coarse_offd)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(RT);
hypre_ParCSRCommPkg *comm_pkg_A = hypre_ParCSRMatrixCommPkg(A);
HYPRE_Int num_recvs_A = hypre_ParCSRCommPkgNumRecvs(comm_pkg_A);
HYPRE_Int *recv_procs_A = hypre_ParCSRCommPkgRecvProcs(comm_pkg_A);
HYPRE_Int *recv_vec_starts_A = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_A);
HYPRE_Int num_sends_A = hypre_ParCSRCommPkgNumSends(comm_pkg_A);
HYPRE_Int *send_procs_A = hypre_ParCSRCommPkgSendProcs(comm_pkg_A);
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int num_recvs_RT;
HYPRE_Int *recv_procs_RT;
HYPRE_Int *recv_vec_starts_RT;
HYPRE_Int num_sends_RT;
HYPRE_Int *send_procs_RT;
HYPRE_Int *send_map_starts_RT;
HYPRE_Int *send_map_elmts_RT;
HYPRE_Int *col_map_offd_RT = hypre_ParCSRMatrixColMapOffd(RT);
HYPRE_Int num_cols_offd_RT = hypre_CSRMatrixNumCols( hypre_ParCSRMatrixOffd(RT));
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(RT);
HYPRE_Int i, j;
HYPRE_Int vec_len, vec_start;
HYPRE_Int num_procs, my_id;
HYPRE_Int ierr = 0;
HYPRE_Int num_requests;
HYPRE_Int offd_col, proc_num;
HYPRE_Int *proc_mark;
HYPRE_Int *change_array;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
/*--------------------------------------------------------------------------
* determine num_recvs, recv_procs and recv_vec_starts for RT
*--------------------------------------------------------------------------*/
proc_mark = hypre_CTAlloc(HYPRE_Int, num_recvs_A);
for (i=0; i < num_recvs_A; i++)
proc_mark[i] = 0;
proc_num = 0;
num_recvs_RT = 0;
if (num_cols_offd_RT)
{
for (i=0; i < num_recvs_A; i++)
{
for (j=recv_vec_starts_A[i]; j<recv_vec_starts_A[i+1]; j++)
{
offd_col = col_map_offd_RT[proc_num];
if (offd_col == j)
{
proc_mark[i]++;
proc_num++;
if (proc_num == num_cols_offd_RT) break;
}
}
if (proc_mark[i]) num_recvs_RT++;
if (proc_num == num_cols_offd_RT) break;
}
}
for (i=0; i < num_cols_offd_RT; i++)
col_map_offd_RT[i] = fine_to_coarse_offd[col_map_offd_RT[i]];
recv_procs_RT = hypre_CTAlloc(HYPRE_Int,num_recvs_RT);
recv_vec_starts_RT = hypre_CTAlloc(HYPRE_Int, num_recvs_RT+1);
j = 0;
recv_vec_starts_RT[0] = 0;
for (i=0; i < num_recvs_A; i++)
if (proc_mark[i])
{
recv_procs_RT[j] = recv_procs_A[i];
recv_vec_starts_RT[j+1] = recv_vec_starts_RT[j]+proc_mark[i];
j++;
}
/*--------------------------------------------------------------------------
* send num_changes to recv_procs_A and receive change_array from send_procs_A
*--------------------------------------------------------------------------*/
num_requests = num_recvs_A+num_sends_A;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
change_array = hypre_CTAlloc(HYPRE_Int, num_sends_A);
j = 0;
for (i=0; i < num_sends_A; i++)
hypre_MPI_Irecv(&change_array[i],1,HYPRE_MPI_INT,send_procs_A[i],0,comm,
&requests[j++]);
for (i=0; i < num_recvs_A; i++)
hypre_MPI_Isend(&proc_mark[i],1,HYPRE_MPI_INT,recv_procs_A[i],0,comm,
&requests[j++]);
hypre_MPI_Waitall(num_requests,requests,status);
hypre_TFree(proc_mark);
/*--------------------------------------------------------------------------
* if change_array[i] is 0 , omit send_procs_A[i] in send_procs_RT
*--------------------------------------------------------------------------*/
num_sends_RT = 0;
for (i=0; i < num_sends_A; i++)
if (change_array[i])
{
num_sends_RT++;
}
send_procs_RT = hypre_CTAlloc(HYPRE_Int, num_sends_RT);
send_map_starts_RT = hypre_CTAlloc(HYPRE_Int, num_sends_RT+1);
j = 0;
send_map_starts_RT[0] = 0;
for (i=0; i < num_sends_A; i++)
if (change_array[i])
{
send_procs_RT[j] = send_procs_A[i];
send_map_starts_RT[j+1] = send_map_starts_RT[j]+change_array[i];
j++;
}
/*--------------------------------------------------------------------------
* generate send_map_elmts
*--------------------------------------------------------------------------*/
send_map_elmts_RT = hypre_CTAlloc(HYPRE_Int,send_map_starts_RT[num_sends_RT]);
j = 0;
for (i=0; i < num_sends_RT; i++)
{
vec_start = send_map_starts_RT[i];
vec_len = send_map_starts_RT[i+1]-vec_start;
hypre_MPI_Irecv(&send_map_elmts_RT[vec_start],vec_len,HYPRE_MPI_INT,
send_procs_RT[i],0,comm,&requests[j++]);
}
for (i=0; i < num_recvs_RT; i++)
{
vec_start = recv_vec_starts_RT[i];
vec_len = recv_vec_starts_RT[i+1] - vec_start;
hypre_MPI_Isend(&col_map_offd_RT[vec_start],vec_len,HYPRE_MPI_INT,
recv_procs_RT[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
for (i=0; i < send_map_starts_RT[num_sends_RT]; i++)
send_map_elmts_RT[i] -= first_col_diag;
comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(comm_pkg) = num_sends_RT;
hypre_ParCSRCommPkgNumRecvs(comm_pkg) = num_recvs_RT;
hypre_ParCSRCommPkgSendProcs(comm_pkg) = send_procs_RT;
hypre_ParCSRCommPkgRecvProcs(comm_pkg) = recv_procs_RT;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg) = recv_vec_starts_RT;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg) = send_map_starts_RT;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = send_map_elmts_RT;
hypre_TFree(status);
hypre_TFree(requests);
hypre_ParCSRMatrixCommPkg(RT) = comm_pkg;
hypre_TFree(change_array);
return ierr;
}
void hypre_ParChordMatrix_RowStarts(
hypre_ParChordMatrix *Ac, MPI_Comm comm,
HYPRE_Int ** row_starts, HYPRE_Int * global_num_cols )
/* This function computes the ParCSRMatrix-style row_starts from a chord matrix.
It requires the the idofs of the chord matrix be partitioned among
processors, so their numbering is monotonic with the processor number;
see below.
The algorithm: each proc. p knows its min & max global row & col numbers.
Mins are first_index_rdof[p], first_index_idof[p]
***IF*** these were in proper order (see below),
first_index_rdof[p] is row_starts[p].
Add num_rdofs-1 to get the max, i.e. add num_rdofs
to get row_starts[p+1] (IF the processors are ordered thus).
Compute these, then broadcast to the other processors to form
row_starts.
(We also could get global_num_rows by an AllReduce num_idofs.)
We get global_num_cols by taking the min and max over processors of
the min and max col no.s on each processor.
If the chord matrix is not ordered so the above will work, then we
would need to to completely move matrices around sometimes, a very expensive
operation.
The problem is that the chord matrix format makes no assumptions about
processor order, but the ParCSR format assumes that
p<q => (local row numbers of p) < (local row numbers of q)
Maybe instead I could change the global numbering scheme as part of this
conversion.
A closely related ordering-type problem to watch for: row_starts must be
a partition for a ParCSRMatrix. In a ChordMatrix, the struct itself
makes no guarantees, but Panayot said, in essence, that row_starts will
be a partition.
col_starts should be NULL; later we shall let the Create function compute one.
*/
{
HYPRE_Int * fis_idof = hypre_ParChordMatrixFirstindexIdof(Ac);
HYPRE_Int * fis_rdof = hypre_ParChordMatrixFirstindexRdof(Ac);
HYPRE_Int my_id, num_procs;
HYPRE_Int num_idofs = hypre_ParChordMatrixNumIdofs(Ac);
HYPRE_Int num_rdofs = hypre_ParChordMatrixNumRdofs(Ac);
HYPRE_Int min_rdof, max_rdof, global_min_rdof, global_max_rdof;
HYPRE_Int p, lens[2], lastlens[2];
hypre_MPI_Status *status;
hypre_MPI_Request *request;
hypre_MPI_Comm_rank(comm, &my_id);
hypre_MPI_Comm_size(comm, &num_procs);
request = hypre_CTAlloc(hypre_MPI_Request, 1 );
status = hypre_CTAlloc(hypre_MPI_Status, 1 );
min_rdof = fis_rdof[my_id];
max_rdof = min_rdof + num_rdofs;
lens[0] = num_idofs;
lens[1] = num_rdofs;
/* row_starts (except last value */
*row_starts = hypre_CTAlloc( HYPRE_Int, num_procs+1 );
for ( p=0; p<num_procs; ++p ) {
(*row_starts)[p] = fis_idof[p];
}
/* check that ordering and partitioning of rows is as expected
(much is missing, and even then not perfect)... */
if ( my_id<num_procs-1 )
hypre_MPI_Isend( lens, 2, HYPRE_MPI_INT, my_id+1, 0, comm, request );
if ( my_id>0 )
hypre_MPI_Recv( lastlens, 2, HYPRE_MPI_INT, my_id-1, 0, comm, status );
if ( my_id<num_procs-1 )
hypre_MPI_Waitall( 1, request, status);
if ( my_id>0 )
hypre_assert( (*row_starts)[my_id] == (*row_starts)[my_id-1] + lastlens[0] );
hypre_TFree( request );
hypre_TFree( status );
/* Get the upper bound for all the rows */
hypre_MPI_Bcast( lens, 2, HYPRE_MPI_INT, num_procs-1, comm );
(*row_starts)[num_procs] = (*row_starts)[num_procs-1] + lens[0];
/* Global number of columns */
/* hypre_MPI_Allreduce( &num_rdofs, global_num_cols, 1, HYPRE_MPI_INT, hypre_MPI_SUM, comm );*/
hypre_MPI_Allreduce( &min_rdof, &global_min_rdof, 1, HYPRE_MPI_INT, hypre_MPI_MIN, comm );
hypre_MPI_Allreduce( &max_rdof, &global_max_rdof, 1, HYPRE_MPI_INT, hypre_MPI_MAX, comm );
*global_num_cols = global_max_rdof - global_min_rdof;
}
HYPRE_Int
hypre_ParCSRMatrixToParChordMatrix(
hypre_ParCSRMatrix *Ap,
MPI_Comm comm,
hypre_ParChordMatrix **pAc )
{
HYPRE_Int * row_starts = hypre_ParCSRMatrixRowStarts(Ap);
HYPRE_Int * col_starts = hypre_ParCSRMatrixColStarts(Ap);
hypre_CSRMatrix * diag = hypre_ParCSRMatrixDiag(Ap);
hypre_CSRMatrix * offd = hypre_ParCSRMatrixOffd(Ap);
HYPRE_Int * offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Int * diag_j = hypre_CSRMatrixJ(diag);
HYPRE_Int * col_map_offd = hypre_ParCSRMatrixColMapOffd(Ap);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(Ap);
hypre_ParChordMatrix * Ac;
hypre_NumbersNode * rdofs, * offd_cols_me;
hypre_NumbersNode ** offd_cols;
HYPRE_Int ** offd_col_array;
HYPRE_Int * len_offd_col_array, * offd_col_array_me;
HYPRE_Int len_offd_col_array_me;
HYPRE_Int num_idofs, num_rdofs, j_local, j_global, row_global;
HYPRE_Int i, j, jj, p, pto, q, qto, my_id, my_q, row, ireq;
HYPRE_Int num_inprocessors, num_toprocessors, num_procs, len_num_rdofs_toprocessor;
HYPRE_Int *inprocessor, *toprocessor, *pcr, *qcr, *num_inchords, *chord, *chordto;
HYPRE_Int *inproc, *toproc, *num_rdofs_toprocessor;
HYPRE_Int **inchord_idof, **inchord_rdof, **rdof_toprocessor;
double **inchord_data;
double data;
HYPRE_Int *first_index_idof, *first_index_rdof;
hypre_MPI_Request * request;
hypre_MPI_Status * status;
hypre_MPI_Comm_rank(comm, &my_id);
hypre_MPI_Comm_size(comm, &num_procs);
num_idofs = row_starts[my_id+1] - row_starts[my_id];
num_rdofs = col_starts[my_id+1] - col_starts[my_id];
hypre_ParChordMatrixCreate( pAc, comm, num_idofs, num_rdofs );
Ac = *pAc;
/* The following block sets Inprocessor:
On each proc. my_id, we find the columns in the offd and diag blocks
(global no.s). The columns are rdofs (contrary to what I wrote in
ChordMatrix.txt).
For each such col/rdof r, find the proc. p which owns row/idof r.
We set the temporary array pcr[p]=1 for such p.
An MPI all-to-all will exchange such arrays so my_id's array qcr has
qcr[q]=1 iff, on proc. q, pcr[my_id]=1. In other words, qcr[q]=1 if
my_id owns a row/idof i which is the same as a col/rdof owned by q.
Collect all such q's into in the array Inprocessor.
While constructing pcr, we also construct pj such that for any index jj
into offd_j,offd_data, pj[jj] is the processor which owns jj as a row (idof)
(the number jj is local to this processor).
*/
pcr = hypre_CTAlloc( HYPRE_Int, num_procs );
qcr = hypre_CTAlloc( HYPRE_Int, num_procs );
for ( p=0; p<num_procs; ++p ) pcr[p]=0;
for ( jj=0; jj<hypre_CSRMatrixNumNonzeros(offd); ++jj ) {
j_local = offd_j[jj];
j_global = col_map_offd[j_local];
for ( p=0; p<num_procs; ++p ) {
if ( j_global >= row_starts[p] && j_global<row_starts[p+1] ) {
pcr[p]=1;
/* not used yet... pj[jj] = p;*/
break;
}
}
}
/* jjd = jj; ...not used yet */
/* pcr[my_id] = 1; ...for square matrices (with nonzero diag block)
this one line would do the job of the following nested loop.
For non-square matrices, the data distribution is too arbitrary. */
for ( jj=0; jj<hypre_CSRMatrixNumNonzeros(diag); ++jj ) {
j_local = diag_j[jj];
j_global = j_local + first_col_diag;
for ( p=0; p<num_procs; ++p ) {
if ( j_global >= row_starts[p] && j_global<row_starts[p+1] ) {
pcr[p]=1;
/* not used yet... pj[jj+jjd] = p;*/
break;
}
}
}
/* Now pcr[p]=1 iff my_id owns a col/rdof r which proc. p owns as a row/idof */
hypre_MPI_Alltoall( pcr, 1, HYPRE_MPI_INT, qcr, 1, HYPRE_MPI_INT, comm );
/* Now qcr[q]==1 if my_id owns a row/idof i which is a col/rdof of proc. q
The array of such q's is the array Inprocessor. */
num_inprocessors = 0;
for ( q=0; q<num_procs; ++q ) if ( qcr[q]==1 ) ++num_inprocessors;
inprocessor = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
p = 0;
for ( q=0; q<num_procs; ++q ) if ( qcr[q]==1 ) inprocessor[ p++ ] = q;
num_toprocessors = 0;
for ( q=0; q<num_procs; ++q ) if ( pcr[q]==1 ) ++num_toprocessors;
toprocessor = hypre_CTAlloc( HYPRE_Int, num_toprocessors );
p = 0;
for ( q=0; q<num_procs; ++q ) if ( pcr[q]==1 ) toprocessor[ p++ ] = q;
hypre_ParChordMatrixNumInprocessors(Ac) = num_inprocessors;
hypre_ParChordMatrixInprocessor(Ac) = inprocessor;
hypre_ParChordMatrixNumToprocessors(Ac) = num_toprocessors;
hypre_ParChordMatrixToprocessor(Ac) = toprocessor;
hypre_TFree( qcr );
/* FirstIndexIdof[p] is the global index of proc. p's row 0 */
/* FirstIndexRdof[p] is the global index of proc. p's col 0 */
/* Fir FirstIndexIdof, we copy the array row_starts rather than its pointers,
because the chord matrix will think it's free to delete FirstIndexIdof */
/* col_starts[p] contains the global index of the first column
in the diag block of p. But for first_index_rdof we want the global
index of the first column in p (whether that's in the diag or offd block).
So it's more involved than row/idof: we also check the offd block, and
have to do a gather to get first_index_rdof for every proc. on every proc. */
first_index_idof = hypre_CTAlloc( HYPRE_Int, num_procs+1 );
first_index_rdof = hypre_CTAlloc( HYPRE_Int, num_procs+1 );
for ( p=0; p<=num_procs; ++p ) {
first_index_idof[p] = row_starts[p];
first_index_rdof[p] = col_starts[p];
};
if ( hypre_CSRMatrixNumRows(offd) > 0 && hypre_CSRMatrixNumCols(offd) > 0 )
first_index_rdof[my_id] =
col_starts[my_id]<col_map_offd[0] ? col_starts[my_id] : col_map_offd[0];
hypre_MPI_Allgather( &first_index_rdof[my_id], 1, HYPRE_MPI_INT,
first_index_rdof, 1, HYPRE_MPI_INT, comm );
/* Set num_inchords: num_inchords[p] is no. chords on my_id connected to p.
Set each chord (idof,jdof,data).
We go through each matrix element in the diag block, find what processor
owns its column no. as a row, then update num_inchords[p], inchord_idof[p],
inchord_rdof[p], inchord_data[p].
*/
inchord_idof = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
inchord_rdof = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
inchord_data = hypre_CTAlloc( double*, num_inprocessors );
num_inchords = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
chord = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
chordto = hypre_CTAlloc( HYPRE_Int, num_toprocessors );
num_rdofs = 0;
for ( q=0; q<num_inprocessors; ++q ) num_inchords[q] = 0;
my_q = -1;
for ( q=0; q<num_inprocessors; ++q ) if ( inprocessor[q]==my_id ) my_q = q;
hypre_assert( my_q>=0 );
/* diag block: first count chords (from my_id to my_id),
then set them from diag block's CSR data structure */
num_idofs = hypre_CSRMatrixNumRows(diag);
rdofs = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(diag); ++row ) {
for ( i=hypre_CSRMatrixI(diag)[row]; i<hypre_CSRMatrixI(diag)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(diag)[i];
hypre_NumbersEnter( rdofs, j_local );
++num_inchords[my_q];
}
};
num_rdofs = hypre_NumbersNEntered( rdofs );
inchord_idof[my_q] = hypre_CTAlloc( HYPRE_Int, num_inchords[my_q] );
inchord_rdof[my_q] = hypre_CTAlloc( HYPRE_Int, num_inchords[my_q] );
inchord_data[my_q] = hypre_CTAlloc( double, num_inchords[my_q] );
chord[0] = 0;
for ( row=0; row<hypre_CSRMatrixNumRows(diag); ++row ) {
for ( i=hypre_CSRMatrixI(diag)[row]; i<hypre_CSRMatrixI(diag)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(diag)[i];
data = hypre_CSRMatrixData(diag)[i];
inchord_idof[my_q][chord[0]] = row;
/* Here We need to convert from j_local - a column local to
the diag of this proc., to a j which is local only to this
processor - a column (rdof) numbering scheme to be shared by the
diag and offd blocks... */
j_global = j_local + hypre_ParCSRMatrixColStarts(Ap)[my_q];
j = j_global - first_index_rdof[my_q];
inchord_rdof[my_q][chord[0]] = j;
inchord_data[my_q][chord[0]] = data;
hypre_assert( chord[0] < num_inchords[my_q] );
++chord[0];
}
};
hypre_NumbersDeleteNode(rdofs);
/* offd block: */
/* >>> offd_cols_me duplicates rdofs */
offd_cols_me = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(offd); ++row ) {
for ( i=hypre_CSRMatrixI(offd)[row]; i<hypre_CSRMatrixI(offd)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(offd)[i];
j_global = col_map_offd[j_local];
hypre_NumbersEnter( offd_cols_me, j_global );
}
}
offd_col_array = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
len_offd_col_array = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
offd_col_array_me = hypre_NumbersArray( offd_cols_me );
len_offd_col_array_me = hypre_NumbersNEntered( offd_cols_me );
request = hypre_CTAlloc(hypre_MPI_Request, 2*num_procs );
ireq = 0;
for ( q=0; q<num_inprocessors; ++q )
hypre_MPI_Irecv( &len_offd_col_array[q], 1, HYPRE_MPI_INT,
inprocessor[q], 0, comm, &request[ireq++] );
for ( p=0; p<num_procs; ++p ) if ( pcr[p]==1 ) {
hypre_MPI_Isend( &len_offd_col_array_me, 1, HYPRE_MPI_INT, p, 0, comm, &request[ireq++] );
}
status = hypre_CTAlloc(hypre_MPI_Status, ireq );
hypre_MPI_Waitall( ireq, request, status );
hypre_TFree(status);
ireq = 0;
for ( q=0; q<num_inprocessors; ++q )
offd_col_array[q] = hypre_CTAlloc( HYPRE_Int, len_offd_col_array[q] );
for ( q=0; q<num_inprocessors; ++q )
hypre_MPI_Irecv( offd_col_array[q], len_offd_col_array[q], HYPRE_MPI_INT,
inprocessor[q], 0, comm, &request[ireq++] );
for ( p=0; p<num_procs; ++p ) if ( pcr[p]==1 ) {
hypre_MPI_Isend( offd_col_array_me, len_offd_col_array_me,
HYPRE_MPI_INT, p, 0, comm, &request[ireq++] );
}
status = hypre_CTAlloc(hypre_MPI_Status, ireq );
hypre_MPI_Waitall( ireq, request, status );
hypre_TFree(request);
hypre_TFree(status);
offd_cols = hypre_CTAlloc( hypre_NumbersNode *, num_inprocessors );
for ( q=0; q<num_inprocessors; ++q ) {
offd_cols[q] = hypre_NumbersNewNode();
for ( i=0; i<len_offd_col_array[q]; ++i )
hypre_NumbersEnter( offd_cols[q], offd_col_array[q][i] );
}
len_num_rdofs_toprocessor = 1 + hypre_CSRMatrixI(offd)
[hypre_CSRMatrixNumRows(offd)];
inproc = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
toproc = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
num_rdofs_toprocessor = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
for ( qto=0; qto<len_num_rdofs_toprocessor; ++qto ) {
inproc[qto] = -1;
toproc[qto] = -1;
num_rdofs_toprocessor[qto] = 0;
};
rdofs = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(offd); ++row ) {
for ( i=hypre_CSRMatrixI(offd)[row]; i<hypre_CSRMatrixI(offd)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(offd)[i];
j_global = col_map_offd[j_local];
hypre_NumbersEnter( rdofs, j_local );
/* TO DO: find faster ways to do the two processor lookups below.*/
/* Find a processor p (local index q) from the inprocessor list,
which owns the column(rdof) whichis the same as this processor's
row(idof) row. Update num_inchords for p.
Save q as inproc[i] for quick recall later. It represents
an inprocessor (not unique) connected to a chord i.
*/
inproc[i] = -1;
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
p = inprocessor[q];
if ( hypre_NumbersQuery( offd_cols[q],
row+hypre_ParCSRMatrixFirstRowIndex(Ap) )
== 1 ) {
/* row is one of the offd columns of p */
++num_inchords[q];
inproc[i] = q;
break;
}
}
if ( inproc[i]<0 ) {
/* For square matrices, we would have found the column in some
other processor's offd. But for non-square matrices it could
exist only in some other processor's diag...*/
/* Note that all data in a diag block is stored. We don't check
whether the value of a data entry is zero. */
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
p = inprocessor[q];
row_global = row+hypre_ParCSRMatrixFirstRowIndex(Ap);
if ( row_global>=col_starts[p] &&
row_global< col_starts[p+1] ) {
/* row is one of the diag columns of p */
++num_inchords[q];
inproc[i] = q;
break;
}
}
}
hypre_assert( inproc[i]>=0 );
/* Find the processor pto (local index qto) from the toprocessor list,
which owns the row(idof) which is the same as this processor's
column(rdof) j_global. Update num_rdofs_toprocessor for pto.
Save pto as toproc[i] for quick recall later. It represents
the toprocessor connected to a chord i. */
for ( qto=0; qto<num_toprocessors; ++qto ) {
pto = toprocessor[qto];
if ( j_global >= row_starts[pto] && j_global<row_starts[pto+1] ) {
hypre_assert( qto < len_num_rdofs_toprocessor );
++num_rdofs_toprocessor[qto];
/* ... an overestimate, as if two chords share an rdof, that
rdof will be counted twice in num_rdofs_toprocessor.
It can be fixed up later.*/
toproc[i] = qto;
break;
}
}
}
};
num_rdofs += hypre_NumbersNEntered(rdofs);
hypre_NumbersDeleteNode(rdofs);
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
inchord_idof[q] = hypre_CTAlloc( HYPRE_Int, num_inchords[q] );
inchord_rdof[q] = hypre_CTAlloc( HYPRE_Int, num_inchords[q] );
inchord_data[q] = hypre_CTAlloc( double, num_inchords[q] );
chord[q] = 0;
};
