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_Int hypre_DataExchangeList(HYPRE_Int num_contacts,
HYPRE_Int *contact_proc_list,
void *contact_send_buf,
HYPRE_Int *contact_send_buf_starts,
HYPRE_Int contact_obj_size,
HYPRE_Int response_obj_size,
hypre_DataExchangeResponse *response_obj,
HYPRE_Int max_response_size,
HYPRE_Int rnum, MPI_Comm comm,
void **p_response_recv_buf,
HYPRE_Int **p_response_recv_buf_starts)
{
/*-------------------------------------------
* parameters:
*
* num_contacts = how many procs to contact
* contact_proc_list = list of processors to contact
* contact_send_buf = array of data to send
* contact_send_buf_starts = index for contact_send_buf corresponding to
* contact_proc_list
* contact_obj_size = sizeof() one obj in contact list
* response_obj_size = sizeof() one obj in response_recv_buf
* response_obj = this will give us the function we need to
* fill the reponse as well as
* any data we might need to accomplish that
* max_response_size = max size of a single response expected (do NOT
* need to be an absolute upper bound)
* rnum = two consequentive exchanges should have different
* rnums. Alternate rnum = 1
* and rnum=2 - these flags will be even (so odd
* numbered tags could be used in calling code)
* p_response_recv_buf = where to receive the reponses - will be allocated
* in this function
* p_response_recv_buf_starts = index of p_response_buf corresponding to
* contact_buf_list - will be allocated here
*-------------------------------------------*/
HYPRE_Int num_procs, myid;
HYPRE_Int i;
HYPRE_Int terminate, responses_complete;
HYPRE_Int children_complete;
HYPRE_Int contact_flag;
HYPRE_Int proc;
HYPRE_Int contact_size;
HYPRE_Int size, post_size, copy_size;
HYPRE_Int total_size, count;
void *start_ptr = NULL, *index_ptr=NULL;
HYPRE_Int *int_ptr=NULL;
void *response_recv_buf = NULL;
void *send_response_buf = NULL;
HYPRE_Int *response_recv_buf_starts = NULL;
void *initial_recv_buf = NULL;
void *recv_contact_buf = NULL;
HYPRE_Int recv_contact_buf_size = 0;
HYPRE_Int response_message_size = 0;
HYPRE_Int overhead;
HYPRE_Int max_response_size_bytes;
HYPRE_Int max_response_total_bytes;
void **post_array = NULL; /*this must be set to null or realloc will crash */
HYPRE_Int post_array_storage = 0;
HYPRE_Int post_array_size = 0;
HYPRE_Int num_post_recvs =0;
void **contact_ptrs = NULL, **response_ptrs=NULL, **post_ptrs=NULL;
hypre_BinaryTree tree;
hypre_MPI_Request *response_requests, *contact_requests;
hypre_MPI_Status *response_statuses, *contact_statuses;
hypre_MPI_Request *post_send_requests = NULL, *post_recv_requests = NULL;
hypre_MPI_Status *post_send_statuses = NULL, *post_recv_statuses = NULL;
hypre_MPI_Request *term_requests, term_request1, request_parent;
hypre_MPI_Status *term_statuses, term_status1, status_parent;
hypre_MPI_Status status, fill_status;
const HYPRE_Int contact_tag = 1000*rnum;
const HYPRE_Int response_tag = 1002*rnum;
const HYPRE_Int term_tag = 1004*rnum;
const HYPRE_Int post_tag = 1006*rnum;
hypre_MPI_Comm_size(comm, &num_procs );
hypre_MPI_Comm_rank(comm, &myid );
/* ---------initializations ----------------*/
/* if the response_obj_size or contact_obj_size is 0, set to sizeof(HYPRE_Int) */
if (!response_obj_size) response_obj_size = sizeof(HYPRE_Int);
if (!contact_obj_size) contact_obj_size = sizeof(HYPRE_Int);
max_response_size_bytes = max_response_size*response_obj_size;
/* pre-allocate the max space for responding to contacts */
overhead = ceil((HYPRE_Real) sizeof(HYPRE_Int)/response_obj_size); /*for appending an integer*/
max_response_total_bytes = (max_response_size+overhead)*response_obj_size;
response_obj->send_response_overhead = overhead;
response_obj->send_response_storage = max_response_size;
/*send_response_buf = hypre_MAlloc(max_response_total_bytes);*/
send_response_buf = hypre_CAlloc(max_response_size+overhead, response_obj_size);
/*allocate space for inital recv array for the responses - give each processor
size max_response_size */
initial_recv_buf = hypre_MAlloc(max_response_total_bytes*num_contacts);
response_recv_buf_starts = hypre_CTAlloc(HYPRE_Int, num_contacts+1);
contact_ptrs = hypre_TAlloc( void *, num_contacts);
response_ptrs = hypre_TAlloc(void *, num_contacts);
/*-------------SEND CONTACTS AND POST RECVS FOR RESPONSES---*/
for (i=0; i<= num_contacts; i++)
{
response_recv_buf_starts[i] = i*(max_response_size+overhead);
}
/* Send "contact" messages to the list of processors and
pre-post receives to wait for their response*/
responses_complete = 1;
if (num_contacts > 0 )
{
responses_complete = 0;
response_requests = hypre_CTAlloc(hypre_MPI_Request, num_contacts);
response_statuses = hypre_CTAlloc(hypre_MPI_Status, num_contacts);
contact_requests = hypre_CTAlloc(hypre_MPI_Request, num_contacts);
contact_statuses = hypre_CTAlloc(hypre_MPI_Status, num_contacts);
/* post receives - could be confirmation or data*/
/* the size to post is max_response_total_bytes*/
for (i=0; i< num_contacts; i++)
{
/* response_ptrs[i] = initial_recv_buf + i*max_response_total_bytes ; */
response_ptrs[i] = (void *)((char *) initial_recv_buf +
i*max_response_total_bytes) ;
hypre_MPI_Irecv(response_ptrs[i], max_response_total_bytes,
hypre_MPI_BYTE, contact_proc_list[i],
response_tag, comm, &response_requests[i]);
}
/* send out contact messages */
start_ptr = contact_send_buf;
for (i=0; i< num_contacts; i++)
{
contact_ptrs[i] = start_ptr;
size = contact_send_buf_starts[i+1] - contact_send_buf_starts[i] ;
hypre_MPI_Isend(contact_ptrs[i], size*contact_obj_size,
hypre_MPI_BYTE, contact_proc_list[i],
contact_tag, comm, &contact_requests[i]);
/* start_ptr += (size*contact_obj_size); */
start_ptr = (void *) ((char *) start_ptr + (size*contact_obj_size));
}
}
/*------------BINARY TREE-----------------------*/
/*Now let's find out our binary tree information and
initialize for the termination check sweep */
terminate = 1; /*indicates whether we can stop probing for contact */
children_complete = 1;/*indicates whether we have recv. term messages
from our children*/
if (num_procs > 1)
{
hypre_CreateBinaryTree(myid, num_procs, &tree);
/* we will get a message from all of our children when they
have received responses for all of their contacts.
So post receives now */
term_requests = hypre_CTAlloc(hypre_MPI_Request, tree.num_child);
term_statuses = hypre_CTAlloc(hypre_MPI_Status, tree.num_child);
for (i=0; i< tree.num_child; i++)
{
hypre_MPI_Irecv(NULL, 0, HYPRE_MPI_INT, tree.child_id[i], term_tag, comm,
&term_requests[i]);
}
terminate = 0;
children_complete = 0;
}
else if (num_procs ==1 && num_contacts > 0 ) /* added 11/08 */
{
terminate = 0;
}
/*---------PROBE LOOP-----------------------------------------*/
/*Look for incoming contact messages - don't know how many I will get!*/
while (!terminate)
{
/* did I receive any contact messages? */
hypre_MPI_Iprobe(hypre_MPI_ANY_SOURCE, contact_tag, comm,
&contact_flag, &status);
while (contact_flag)
{
/* received contacts - from who and what do we do ?*/
proc = status.hypre_MPI_SOURCE;
hypre_MPI_Get_count(&status, hypre_MPI_BYTE, &contact_size);
contact_size = contact_size/contact_obj_size;
/*---------------FILL RESPONSE ------------------------*/
/*first receive the contact buffer - then call a function
to determine how to populate the send buffer for the reponse*/
/* do we have enough space to recv it? */
if(contact_size > recv_contact_buf_size)
{
recv_contact_buf = hypre_ReAlloc(recv_contact_buf,
contact_obj_size*contact_size);
recv_contact_buf_size = contact_size;
}
/* this must be blocking - can't fill recv without the buffer*/
hypre_MPI_Recv(recv_contact_buf, contact_size*contact_obj_size,
hypre_MPI_BYTE, proc, contact_tag, comm, &fill_status);
response_obj->fill_response(recv_contact_buf, contact_size, proc,
response_obj, comm, &send_response_buf,
&response_message_size );
/* we need to append the size of the send obj */
/* first we copy out any part that may be needed to send later so we don't overwrite */
post_size = response_message_size - max_response_size;
if (post_size > 0) /*we will need to send the extra information later */
{
/*hypre_printf("myid = %d, post_size = %d\n", myid, post_size);*/
if (post_array_size == post_array_storage)
{
/* allocate room for more posts - add 20*/
post_array_storage += 20;
post_array = hypre_TReAlloc(post_array, void *, post_array_storage);
post_send_requests =
hypre_TReAlloc(post_send_requests, hypre_MPI_Request,
post_array_storage);
}
/* allocate space for the data this post only*/
/* this should not happen often (unless a poor max_size has been chosen)
- so we will allocate space for the data as needed */
size = post_size*response_obj_size;
post_array[post_array_size] = hypre_MAlloc(size);
/* index_ptr = send_response_buf + max_response_size_bytes */;
index_ptr = (void *) ((char *) send_response_buf +
max_response_size_bytes);
memcpy(post_array[post_array_size], index_ptr, size);
/*now post any part of the message that is too long with a non-blocking
send and a different tag */
hypre_MPI_Isend(post_array[post_array_size], size,
hypre_MPI_BYTE, proc, post_tag,
/*hypre_MPI_COMM_WORLD, */
comm,
&post_send_requests[post_array_size]);
post_array_size++;
}
/*now append the size information into the overhead storage */
/* index_ptr = send_response_buf + max_response_size_bytes; */
index_ptr = (void *) ((char *) send_response_buf +
max_response_size_bytes);
memcpy(index_ptr, &response_message_size, sizeof(HYPRE_Int));
/*send the block of data that includes the overhead */
/* this is a blocking send - the recv has already been posted */
hypre_MPI_Send(send_response_buf, max_response_total_bytes,
hypre_MPI_BYTE, proc, response_tag, comm);
/*--------------------------------------------------------------*/
/* look for any more contact messages*/
hypre_MPI_Iprobe(hypre_MPI_ANY_SOURCE, contact_tag, comm,
&contact_flag, &status);
}
/* no more contact messages waiting - either
(1) check to see if we have received all of our response messages
(2) participate in termination (check for messages from children)
(3) participate in termination sweep (check for message from parent) */
if (!responses_complete)
{
hypre_MPI_Testall(num_contacts, response_requests, &responses_complete,
response_statuses);
if (responses_complete && num_procs == 1) terminate = 1; /*added 11/08 */
}
else if(!children_complete) /* have all of our children received all of their
response messages?*/
{
hypre_MPI_Testall(tree.num_child, term_requests, &children_complete,
term_statuses);
/* if we have gotten term messages from all of our children, send a term
message to our parent. Then post a receive to hear back from parent */
if (children_complete & (myid > 0)) /*root does not have a parent*/
{
hypre_MPI_Isend(NULL, 0, HYPRE_MPI_INT, tree.parent_id, term_tag,
comm, &request_parent);
hypre_MPI_Irecv(NULL, 0, HYPRE_MPI_INT, tree.parent_id, term_tag,
comm, &term_request1);
}
}
else /*have we gotten a term message from our parent? */
{
if (myid == 0) /* root doesn't have a parent */
{
terminate = 1;
}
else
{
hypre_MPI_Test(&term_request1, &terminate, &term_status1);
}
if (terminate) /*tell children to terminate */
{
if (myid > 0 ) hypre_MPI_Wait(&request_parent, &status_parent);
for (i=0; i< tree.num_child; i++)
{ /*a blocking send - recv has been posted already*/
hypre_MPI_Send(NULL, 0, HYPRE_MPI_INT, tree.child_id[i],
term_tag, comm);
}
}
}
}
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_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_Int HYPRE_ParCSRMLConstructMHMatrix(HYPRE_ParCSRMatrix A, MH_Matrix *mh_mat,
MPI_Comm comm, HYPRE_Int *partition,MH_Context *obj)
{
HYPRE_Int i, j, index, my_id, nprocs, msgid, *tempCnt;
HYPRE_Int sendProcCnt, *sendLeng, *sendProc, **sendList;
HYPRE_Int recvProcCnt, *recvLeng, *recvProc;
HYPRE_Int rowLeng, *colInd, startRow, endRow, localEqns;
HYPRE_Int *diagSize, *offdiagSize, externLeng, *externList, ncnt, nnz;
HYPRE_Int *rowptr, *columns, num_bdry;
double *colVal, *values;
hypre_MPI_Request *Request;
hypre_MPI_Status status;
/* -------------------------------------------------------- */
/* get machine information and local matrix information */
/* -------------------------------------------------------- */
hypre_MPI_Comm_rank(comm, &my_id);
hypre_MPI_Comm_size(comm, &nprocs);
startRow = partition[my_id];
endRow = partition[my_id+1] - 1;
localEqns = endRow - startRow + 1;
/* -------------------------------------------------------- */
/* probe A to find out about diagonal and off-diagonal */
/* block information */
/* -------------------------------------------------------- */
diagSize = (HYPRE_Int*) malloc( sizeof(HYPRE_Int) * localEqns );
offdiagSize = (HYPRE_Int*) malloc( sizeof(HYPRE_Int) * localEqns );
num_bdry = 0;
for ( i = startRow; i <= endRow; i++ )
{
diagSize[i-startRow] = offdiagSize[i-startRow] = 0;
HYPRE_ParCSRMatrixGetRow(A, i, &rowLeng, &colInd, &colVal);
for (j = 0; j < rowLeng; j++)
if ( colInd[j] < startRow || colInd[j] > endRow )
{
//if ( colVal[j] != 0.0 ) offdiagSize[i-startRow]++;
offdiagSize[i-startRow]++;
}
else
{
//if ( colVal[j] != 0.0 ) diagSize[i-startRow]++;
diagSize[i-startRow]++;
}
HYPRE_ParCSRMatrixRestoreRow(A, i, &rowLeng, &colInd, &colVal);
if ( diagSize[i-startRow] + offdiagSize[i-startRow] == 1 ) num_bdry++;
}
/* -------------------------------------------------------- */
/* construct external node list in global eqn numbers */
/* -------------------------------------------------------- */
externLeng = 0;
for ( i = 0; i < localEqns; i++ ) externLeng += offdiagSize[i];
if ( externLeng > 0 )
externList = (HYPRE_Int *) malloc( sizeof(HYPRE_Int) * externLeng);
else externList = NULL;
externLeng = 0;
for ( i = startRow; i <= endRow; i++ )
{
HYPRE_ParCSRMatrixGetRow(A, i, &rowLeng, &colInd, &colVal);
for (j = 0; j < rowLeng; j++)
{
if ( colInd[j] < startRow || colInd[j] > endRow )
//if ( colVal[j] != 0.0 ) externList[externLeng++] = colInd[j];
externList[externLeng++] = colInd[j];
}
HYPRE_ParCSRMatrixRestoreRow(A, i, &rowLeng, &colInd, &colVal);
}
qsort0( externList, 0, externLeng-1 );
ncnt = 0;
for ( i = 1; i < externLeng; i++ )
{
if ( externList[i] != externList[ncnt] )
externList[++ncnt] = externList[i];
}
externLeng = ncnt + 1;
/* -------------------------------------------------------- */
/* allocate the CSR matrix */
/* -------------------------------------------------------- */
nnz = 0;
for ( i = 0; i < localEqns; i++ ) nnz += diagSize[i] + offdiagSize[i];
rowptr = (HYPRE_Int *) malloc( (localEqns + 1) * sizeof(HYPRE_Int) );
columns = (HYPRE_Int *) malloc( nnz * sizeof(HYPRE_Int) );
values = (double *) malloc( nnz * sizeof(double) );
rowptr[0] = 0;
for ( i = 1; i <= localEqns; i++ )
rowptr[i] = rowptr[i-1] + diagSize[i-1] + offdiagSize[i-1];
free( diagSize );
free( offdiagSize );
/* -------------------------------------------------------- */
/* put the matrix data in the CSR matrix */
/* -------------------------------------------------------- */
rowptr[0] = 0;
ncnt = 0;
for ( i = startRow; i <= endRow; i++ )
{
HYPRE_ParCSRMatrixGetRow(A, i, &rowLeng, &colInd, &colVal);
for (j = 0; j < rowLeng; j++)
{
index = colInd[j];
//if ( colVal[j] != 0.0 )
{
if ( index < startRow || index > endRow )
{
columns[ncnt] = hypre_BinarySearch(externList,index,
externLeng );
columns[ncnt] += localEqns;
values [ncnt++] = colVal[j];
}
else
{
columns[ncnt] = index - startRow;
values[ncnt++] = colVal[j];
}
}
}
rowptr[i-startRow+1] = ncnt;
HYPRE_ParCSRMatrixRestoreRow(A, i, &rowLeng, &colInd, &colVal);
}
assert( ncnt == nnz );
/* -------------------------------------------------------- */
/* initialize the MH_Matrix data structure */
/* -------------------------------------------------------- */
mh_mat->Nrows = localEqns;
mh_mat->rowptr = rowptr;
mh_mat->colnum = columns;
mh_mat->values = values;
mh_mat->sendProcCnt = 0;
mh_mat->recvProcCnt = 0;
mh_mat->sendLeng = NULL;
mh_mat->recvLeng = NULL;
mh_mat->sendProc = NULL;
mh_mat->recvProc = NULL;
mh_mat->sendList = NULL;
mh_mat->map = externList;
/* -------------------------------------------------------- */
/* form the remote portion of the matrix */
/* -------------------------------------------------------- */
if ( nprocs > 1 )
{
/* ----------------------------------------------------- */
/* count number of elements to be received from each */
/* remote processor (assume sequential mapping) */
/* ----------------------------------------------------- */
tempCnt = (HYPRE_Int *) malloc( sizeof(HYPRE_Int) * nprocs );
for ( i = 0; i < nprocs; i++ ) tempCnt[i] = 0;
for ( i = 0; i < externLeng; i++ )
{
for ( j = 0; j < nprocs; j++ )
{
if ( externList[i] >= partition[j] &&
externList[i] < partition[j+1] )
{
tempCnt[j]++;
break;
}
}
}
/* ----------------------------------------------------- */
/* compile a list processors data is to be received from */
/* ----------------------------------------------------- */
recvProcCnt = 0;
for ( i = 0; i < nprocs; i++ )
if ( tempCnt[i] > 0 ) recvProcCnt++;
recvLeng = (HYPRE_Int*) malloc( sizeof(HYPRE_Int) * recvProcCnt );
recvProc = (HYPRE_Int*) malloc( sizeof(HYPRE_Int) * recvProcCnt );
recvProcCnt = 0;
for ( i = 0; i < nprocs; i++ )
{
if ( tempCnt[i] > 0 )
{
recvProc[recvProcCnt] = i;
recvLeng[recvProcCnt++] = tempCnt[i];
}
}
/* ----------------------------------------------------- */
/* each processor has to find out how many processors it */
/* has to send data to */
/* ----------------------------------------------------- */
sendLeng = (HYPRE_Int *) malloc( nprocs * sizeof(HYPRE_Int) );
for ( i = 0; i < nprocs; i++ ) tempCnt[i] = 0;
for ( i = 0; i < recvProcCnt; i++ ) tempCnt[recvProc[i]] = 1;
hypre_MPI_Allreduce(tempCnt, sendLeng, nprocs, HYPRE_MPI_INT, hypre_MPI_SUM, comm );
sendProcCnt = sendLeng[my_id];
free( sendLeng );
if ( sendProcCnt > 0 )
{
sendLeng = (HYPRE_Int *) malloc( sendProcCnt * sizeof(HYPRE_Int) );
sendProc = (HYPRE_Int *) malloc( sendProcCnt * sizeof(HYPRE_Int) );
sendList = (HYPRE_Int **) malloc( sendProcCnt * sizeof(HYPRE_Int*) );
}
else
{
sendLeng = sendProc = NULL;
sendList = NULL;
}
/* ----------------------------------------------------- */
/* each processor sends to all processors it expects to */
/* receive data about the lengths of data expected */
/* ----------------------------------------------------- */
msgid = 539;
for ( i = 0; i < recvProcCnt; i++ )
{
hypre_MPI_Send((void*) &recvLeng[i],1,HYPRE_MPI_INT,recvProc[i],msgid,comm);
}
for ( i = 0; i < sendProcCnt; i++ )
{
hypre_MPI_Recv((void*) &sendLeng[i],1,HYPRE_MPI_INT,hypre_MPI_ANY_SOURCE,msgid,
comm,&status);
sendProc[i] = status.hypre_MPI_SOURCE;
sendList[i] = (HYPRE_Int *) malloc( sendLeng[i] * sizeof(HYPRE_Int) );
if ( sendList[i] == NULL )
hypre_printf("allocate problem %d \n", sendLeng[i]);
}
/* ----------------------------------------------------- */
/* each processor sends to all processors it expects to */
/* receive data about the equation numbers */
/* ----------------------------------------------------- */
for ( i = 0; i < nprocs; i++ ) tempCnt[i] = 0;
ncnt = 1;
for ( i = 0; i < externLeng; i++ )
{
if ( externList[i] >= partition[ncnt] )
{
tempCnt[ncnt-1] = i;
i--;
ncnt++;
}
}
for ( i = ncnt-1; i < nprocs; i++ ) tempCnt[i] = externLeng;
/* ----------------------------------------------------- */
/* send the global equation numbers */
/* ----------------------------------------------------- */
msgid = 540;
for ( i = 0; i < recvProcCnt; i++ )
{
if ( recvProc[i] == 0 ) j = 0;
else j = tempCnt[recvProc[i]-1];
rowLeng = recvLeng[i];
hypre_MPI_Send((void*) &externList[j],rowLeng,HYPRE_MPI_INT,recvProc[i],
msgid,comm);
}
for ( i = 0; i < sendProcCnt; i++ )
{
rowLeng = sendLeng[i];
hypre_MPI_Recv((void*)sendList[i],rowLeng,HYPRE_MPI_INT,sendProc[i],
msgid,comm,&status);
}
/* ----------------------------------------------------- */
/* convert the send list from global to local numbers */
/* ----------------------------------------------------- */
for ( i = 0; i < sendProcCnt; i++ )
{
for ( j = 0; j < sendLeng[i]; j++ )
{
index = sendList[i][j] - startRow;
if ( index < 0 || index >= localEqns )
{
hypre_printf("%d : Construct MH matrix Error - index out ");
hypre_printf("of range%d\n", my_id, index);
}
sendList[i][j] = index;
}
}
/* ----------------------------------------------------- */
/* convert the send list from global to local numbers */
/* ----------------------------------------------------- */
mh_mat->sendProcCnt = sendProcCnt;
mh_mat->recvProcCnt = recvProcCnt;
mh_mat->sendLeng = sendLeng;
mh_mat->recvLeng = recvLeng;
mh_mat->sendProc = sendProc;
mh_mat->recvProc = recvProc;
mh_mat->sendList = sendList;
/* ----------------------------------------------------- */
/* clean up */
/* ----------------------------------------------------- */
free( tempCnt );
}
return 0;
}
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;
}