4 #include "smpi_coll_private.h"
6 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi,
7 "Logging specific to SMPI (mpi)");
9 int SMPI_MPI_Init(int *argc, char ***argv)
11 smpi_process_init(argc, argv);
16 int SMPI_MPI_Finalize()
19 smpi_process_finalize();
23 // right now this just exits the current node, should send abort signal to all
24 // hosts in the communicator (TODO)
25 int SMPI_MPI_Abort(MPI_Comm comm, int errorcode)
31 int SMPI_MPI_Comm_size(MPI_Comm comm, int *size)
33 int retval = MPI_SUCCESS;
38 retval = MPI_ERR_COMM;
39 } else if (NULL == size) {
50 int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank)
52 int retval = MPI_SUCCESS;
57 retval = MPI_ERR_COMM;
58 } else if (NULL == rank) {
61 *rank = smpi_mpi_comm_rank(comm);
69 int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size)
71 int retval = MPI_SUCCESS;
75 if (NULL == datatype) {
76 retval = MPI_ERR_TYPE;
77 } else if (NULL == size) {
80 *size = datatype->size;
88 int SMPI_MPI_Barrier(MPI_Comm comm)
90 int retval = MPI_SUCCESS;
96 retval = MPI_ERR_COMM;
100 * original implemantation:
101 * retval = smpi_mpi_barrier(comm);
102 * this one is unrealistic: it just cond_waits, means no time.
104 retval = nary_tree_barrier( comm, arity );
112 int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
113 int tag, MPI_Comm comm, MPI_Request * request)
115 int retval = MPI_SUCCESS;
119 retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
121 if (NULL != *request && MPI_SUCCESS == retval) {
122 retval = smpi_mpi_irecv(*request);
130 int SMPI_MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src,
131 int tag, MPI_Comm comm, MPI_Status * status)
133 int retval = MPI_SUCCESS;
134 smpi_mpi_request_t request;
138 retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
140 if (NULL != request && MPI_SUCCESS == retval) {
141 retval = smpi_mpi_irecv(request);
142 if (MPI_SUCCESS == retval) {
143 retval = smpi_mpi_wait(request, status);
145 xbt_mallocator_release(smpi_global->request_mallocator, request);
153 int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
154 int tag, MPI_Comm comm, MPI_Request * request)
156 int retval = MPI_SUCCESS;
160 retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
162 if (NULL != *request && MPI_SUCCESS == retval) {
163 retval = smpi_mpi_isend(*request);
171 int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
172 int tag, MPI_Comm comm)
174 int retval = MPI_SUCCESS;
175 smpi_mpi_request_t request;
179 retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
181 if (NULL != request && MPI_SUCCESS == retval) {
182 retval = smpi_mpi_isend(request);
183 if (MPI_SUCCESS == retval) {
184 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
186 xbt_mallocator_release(smpi_global->request_mallocator, request);
195 * MPI_Wait and friends
197 int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status)
199 return smpi_mpi_wait(*request, status);
202 int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
204 return smpi_mpi_waitall(count, requests, status);
207 int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index,
210 return smpi_mpi_waitany(count, requests, index, status);
220 int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
221 int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
225 int retval = MPI_SUCCESS;
226 smpi_mpi_request_t request;
228 rank = smpi_mpi_comm_rank(comm);
230 retval = smpi_create_request(buf, count, datatype, root,
231 (root + 1) % comm->size, 0, comm, &request);
232 request->forward = comm->size - 1;
233 smpi_mpi_isend(request);
235 retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
237 smpi_mpi_irecv(request);
240 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
241 xbt_mallocator_release(smpi_global->request_mallocator, request);
248 * Bcast user entry point
250 int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
253 int retval = MPI_SUCCESS;
257 //retval = flat_tree_bcast(buf, count, datatype, root, comm);
258 retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 );
267 //#ifdef DEBUG_REDUCE
269 * debugging helper function
271 static void print_buffer_int(void *buf, int len, char *msg, int rank)
274 printf("**[%d] %s: ", rank, msg);
275 for (tmp = 0; tmp < len; tmp++) {
277 printf("[%d]", v[tmp]);
282 static void print_buffer_double(void *buf, int len, char *msg, int rank)
286 printf("**[%d] %s: ", rank, msg);
287 for (tmp = 0; tmp < len; tmp++) {
289 printf("[%lf]", v[tmp]);
300 int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
301 MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
303 int retval = MPI_SUCCESS;
308 smpi_mpi_request_t *requests;
309 smpi_mpi_request_t request;
313 rank = smpi_mpi_comm_rank(comm);
316 if (rank != root) { // if i am not ROOT, simply send my buffer to root
319 print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank);
322 smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm,
324 smpi_mpi_isend(request);
325 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
326 xbt_mallocator_release(smpi_global->request_mallocator, request);
329 // i am the ROOT: wait for all buffers by creating one request by sender
331 requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t));
333 void **tmpbufs = xbt_malloc((size-1) * sizeof(void *));
334 for (i = 0; i < size-1; i++) {
335 // we need 1 buffer per request to store intermediate receptions
336 tmpbufs[i] = xbt_malloc(count * datatype->size);
338 // root: initiliaze recv buf with my own snd buf
339 memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char));
341 // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
342 // since we should op values as soon as one receiving request matches.
343 for (i = 0; i < size-1; i++) {
344 // reminder: for smpi_create_request() the src is always the process sending.
345 src = i < root ? i : i + 1;
346 retval = smpi_create_request(tmpbufs[i], count, datatype,
347 src, root, tag, comm, &(requests[i]));
348 if (NULL != requests[i] && MPI_SUCCESS == retval) {
349 if (MPI_SUCCESS == retval) {
350 smpi_mpi_irecv(requests[i]);
354 // now, wait for completion of all irecv's.
355 for (i = 0; i < size-1; i++) {
356 int index = MPI_UNDEFINED;
357 smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE);
359 printf ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index);
360 print_buffer_int(tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)", index),
365 op->func(tmpbufs[index], recvbuf, &count, &datatype);
367 print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank);
369 xbt_free(tmpbufs[index]);
370 /* FIXME: with the following line, it generates an
371 * [xbt_ex/CRITICAL] Conditional list not empty 162518800.
373 // xbt_mallocator_release(smpi_global->request_mallocator, requests[index]);
385 * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members.
387 int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
388 MPI_Op op, MPI_Comm comm );
389 int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
390 MPI_Op op, MPI_Comm comm )
392 int retval = MPI_SUCCESS;
393 int root=1; // arbitrary choice
397 retval = SMPI_MPI_Reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
398 if (MPI_SUCCESS != retval)
401 retval = SMPI_MPI_Bcast( sendbuf, count, datatype, root, comm);
408 * MPI_Scatter user entry point
410 //int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
411 // void *recvbuf, int recvcount, MPI_Datatype recvtype,int root,
413 int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
414 void *recvbuf, int recvcount, MPI_Datatype recvtype,
415 int root, MPI_Comm comm)
417 int retval = MPI_SUCCESS;
422 char *cptr; // to manipulate the void * buffers
423 smpi_mpi_request_t *requests;
424 smpi_mpi_request_t request;
425 smpi_mpi_status_t status;
430 rank = smpi_mpi_comm_rank(comm);
432 requests = xbt_malloc((comm->size-1) * sizeof(smpi_mpi_request_t));
434 // i am the root: distribute my sendbuf
435 //print_buffer_int(sendbuf, comm->size, xbt_strdup("rcvbuf"), rank);
437 for (i=0; i < comm->size; i++) {
438 if ( i!=root ) { // send to processes ...
440 retval = smpi_create_request((void *)cptr, sendcount,
441 datatype, root, i, tag, comm, &(requests[cnt]));
442 if (NULL != requests[cnt] && MPI_SUCCESS == retval) {
443 if (MPI_SUCCESS == retval) {
444 smpi_mpi_isend(requests[cnt]);
449 else { // ... except if it's me.
450 memcpy(recvbuf, (void *)cptr, recvcount*recvtype->size*sizeof(char));
452 cptr += sendcount*datatype->size;
454 for(i=0; i<cnt; i++) { // wait for send to complete
455 /* FIXME: waitall() should be slightly better */
456 smpi_mpi_wait(requests[i], &status);
457 xbt_mallocator_release(smpi_global->request_mallocator, requests[i]);
461 else { // i am a non-root process: wait data from the root
462 retval = smpi_create_request(recvbuf,recvcount,
463 recvtype, root, rank, tag, comm, &request);
464 if (NULL != request && MPI_SUCCESS == retval) {
465 if (MPI_SUCCESS == retval) {
466 smpi_mpi_irecv(request);
469 smpi_mpi_wait(request, &status);
470 xbt_mallocator_release(smpi_global->request_mallocator, request);
484 // used by comm_split to sort ranks based on key values
485 int smpi_compare_rankkeys(const void *a, const void *b);
486 int smpi_compare_rankkeys(const void *a, const void *b)
505 int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key,
508 int retval = MPI_SUCCESS;
511 smpi_mpi_request_t request;
513 smpi_mpi_status_t status;
517 // FIXME: need to test parameters
519 index = smpi_process_index();
520 rank = comm->index_to_rank_map[index];
525 // root node does most of the real work
527 int colormap[comm->size];
528 int keymap[comm->size];
529 int rankkeymap[comm->size * 2];
531 smpi_mpi_communicator_t tempcomm = NULL;
538 // FIXME: use scatter/gather or similar instead of individual comms
539 for (i = 1; i < comm->size; i++) {
540 retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE,
541 rank, MPI_ANY_TAG, comm, &request);
542 smpi_mpi_irecv(request);
543 smpi_mpi_wait(request, &status);
544 colormap[status.MPI_SOURCE] = colorkey[0];
545 keymap[status.MPI_SOURCE] = colorkey[1];
546 xbt_mallocator_release(smpi_global->request_mallocator, request);
549 for (i = 0; i < comm->size; i++) {
550 if (MPI_UNDEFINED == colormap[i]) {
553 // make a list of nodes with current color and sort by keys
555 for (j = i; j < comm->size; j++) {
556 if (colormap[i] == colormap[j]) {
557 colormap[j] = MPI_UNDEFINED;
558 rankkeymap[count * 2] = j;
559 rankkeymap[count * 2 + 1] = keymap[j];
563 qsort(rankkeymap, count, sizeof(int) * 2, &smpi_compare_rankkeys);
566 tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
567 tempcomm->barrier_count = 0;
568 tempcomm->size = count;
569 tempcomm->barrier_mutex = SIMIX_mutex_init();
570 tempcomm->barrier_cond = SIMIX_cond_init();
571 tempcomm->rank_to_index_map = xbt_new(int, count);
572 tempcomm->index_to_rank_map = xbt_new(int, smpi_global->process_count);
573 for (j = 0; j < smpi_global->process_count; j++) {
574 tempcomm->index_to_rank_map[j] = -1;
576 for (j = 0; j < count; j++) {
577 indextmp = comm->rank_to_index_map[rankkeymap[j * 2]];
578 tempcomm->rank_to_index_map[j] = indextmp;
579 tempcomm->index_to_rank_map[indextmp] = j;
581 for (j = 0; j < count; j++) {
582 if (rankkeymap[j * 2]) {
583 retval = smpi_create_request(&j, 1, MPI_INT, 0,
584 rankkeymap[j * 2], 0, comm, &request);
585 request->data = tempcomm;
586 smpi_mpi_isend(request);
587 smpi_mpi_wait(request, &status);
588 xbt_mallocator_release(smpi_global->request_mallocator, request);
590 *comm_out = tempcomm;
597 retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm,
599 smpi_mpi_isend(request);
600 smpi_mpi_wait(request, &status);
601 xbt_mallocator_release(smpi_global->request_mallocator, request);
602 if (MPI_UNDEFINED != color) {
603 retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm,
605 smpi_mpi_irecv(request);
606 smpi_mpi_wait(request, &status);
607 *comm_out = request->data;
616 double SMPI_MPI_Wtime(void)
618 return (SIMIX_get_clock());