1 /* Copyright (c) 2007-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include "xbt/replay.h"
12 #include "simix/smx_private.h"
13 #include "surf/surf.h"
14 #include "simgrid/sg_config.h"
15 #include "colls/colls.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
20 static int match_recv(void* a, void* b, smx_action_t ignored) {
21 MPI_Request ref = (MPI_Request)a;
22 MPI_Request req = (MPI_Request)b;
23 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
25 xbt_assert(ref, "Cannot match recv against null reference");
26 xbt_assert(req, "Cannot match recv against null request");
27 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
28 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
29 //we match, we can transfer some values
30 // FIXME : move this to the copy function ?
31 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
32 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
33 if(ref->real_size < req->real_size) ref->truncated = 1;
35 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
37 XBT_DEBUG("match succeeded");
42 static int match_send(void* a, void* b,smx_action_t ignored) {
43 MPI_Request ref = (MPI_Request)a;
44 MPI_Request req = (MPI_Request)b;
45 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
46 xbt_assert(ref, "Cannot match send against null reference");
47 xbt_assert(req, "Cannot match send against null request");
49 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
50 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
52 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
53 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
54 if(req->real_size < ref->real_size) req->truncated = 1;
56 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
58 XBT_DEBUG("match succeeded");
64 typedef struct s_smpi_factor *smpi_factor_t;
65 typedef struct s_smpi_factor {
68 double values[4];//arbitrary set to 4
70 xbt_dynar_t smpi_os_values = NULL;
71 xbt_dynar_t smpi_or_values = NULL;
72 xbt_dynar_t smpi_ois_values = NULL;
74 // Methods used to parse and store the values for timing injections in smpi
75 // These are taken from surf/network.c and generalized to have more factors
76 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
78 static int factor_cmp(const void *pa, const void *pb)
80 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
81 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
85 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
88 unsigned int iter = 0;
92 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
94 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
95 radical_elements = xbt_str_split(smpi_coef_string, ";");
96 xbt_dynar_foreach(radical_elements, iter, value) {
97 memset(&fact, 0, sizeof(s_smpi_factor_t));
98 radical_elements2 = xbt_str_split(value, ":");
99 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
100 xbt_die("Malformed radical for smpi factor!");
101 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
103 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
105 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
110 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
111 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
112 xbt_dynar_free(&radical_elements2);
114 xbt_dynar_free(&radical_elements);
116 xbt_dynar_sort(smpi_factor, &factor_cmp);
117 xbt_dynar_foreach(smpi_factor, iter, fact) {
118 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
123 static double smpi_os(double size)
125 if (!smpi_os_values) {
126 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
127 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
129 unsigned int iter = 0;
130 s_smpi_factor_t fact;
132 xbt_dynar_foreach(smpi_os_values, iter, fact) {
133 if (size <= fact.factor) {
134 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
137 current=fact.values[0]+fact.values[1]*size;
140 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
145 static double smpi_ois(double size)
147 if (!smpi_ois_values) {
148 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
149 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
151 unsigned int iter = 0;
152 s_smpi_factor_t fact;
154 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
155 if (size <= fact.factor) {
156 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
159 current=fact.values[0]+fact.values[1]*size;
162 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
167 static double smpi_or(double size)
169 if (!smpi_or_values) {
170 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
171 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
173 unsigned int iter = 0;
174 s_smpi_factor_t fact;
176 xbt_dynar_foreach(smpi_or_values, iter, fact) {
177 if (size <= fact.factor) {
178 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
181 current=fact.values[0]+fact.values[1]*size;
183 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
188 static MPI_Request build_request(void *buf, int count,
189 MPI_Datatype datatype, int src, int dst,
190 int tag, MPI_Comm comm, unsigned flags)
192 MPI_Request request = NULL;
194 void *old_buf = NULL;
196 request = xbt_new(s_smpi_mpi_request_t, 1);
198 s_smpi_subtype_t *subtype = datatype->substruct;
200 if(datatype->has_subtype == 1){
201 // This part handles the problem of non-contiguous memory
203 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
205 subtype->serialize(old_buf, buf, count, datatype->substruct);
210 // This part handles the problem of non-contiguous memory (for the
211 // unserialisation at the reception)
212 request->old_buf = old_buf;
213 request->old_type = datatype;
215 request->size = smpi_datatype_size(datatype) * count;
219 request->comm = comm;
220 request->action = NULL;
221 request->flags = flags;
222 request->detached = 0;
223 request->detached_sender = NULL;
224 request->real_src = 0;
226 request->truncated = 0;
227 request->real_size = 0;
228 request->real_tag = 0;
229 if(flags & PERSISTENT)
230 request->refcount = 1;
232 request->refcount = 0;
238 if (flags & SEND) smpi_datatype_unuse(datatype);
244 void smpi_empty_status(MPI_Status * status)
246 if(status != MPI_STATUS_IGNORE) {
247 status->MPI_SOURCE = MPI_ANY_SOURCE;
248 status->MPI_TAG = MPI_ANY_TAG;
249 status->MPI_ERROR = MPI_SUCCESS;
254 void smpi_action_trace_run(char *path)
258 xbt_dict_cursor_t cursor;
262 action_fp = fopen(path, "r");
263 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
267 if (!xbt_dict_is_empty(action_queues)) {
269 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
272 xbt_dict_foreach(action_queues, cursor, name, todo) {
273 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
279 xbt_dict_free(&action_queues);
280 action_queues = xbt_dict_new_homogeneous(NULL);
283 static void smpi_mpi_request_free_voidp(void* request)
285 MPI_Request req = request;
286 smpi_mpi_request_free(&req);
289 /* MPI Low level calls */
290 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
291 int dst, int tag, MPI_Comm comm)
293 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
294 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
295 comm, PERSISTENT | SEND | PREPARED);
299 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
300 int dst, int tag, MPI_Comm comm)
302 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
303 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
304 comm, PERSISTENT | SSEND | SEND | PREPARED);
308 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
309 int src, int tag, MPI_Comm comm)
311 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
312 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
313 comm, PERSISTENT | RECV | PREPARED);
317 void smpi_mpi_start(MPI_Request request)
321 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
322 request->flags &= ~PREPARED;
323 request->flags &= ~FINISHED;
326 if (request->flags & RECV) {
327 print_request("New recv", request);
328 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
329 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
330 mailbox = smpi_process_mailbox_small();
332 mailbox = smpi_process_mailbox();
333 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
334 request->real_size=request->size;
335 smpi_datatype_use(request->old_type);
336 smpi_comm_use(request->comm);
337 request->action = simcall_comm_irecv(mailbox, request->buf,
338 &request->real_size, &match_recv,
341 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
342 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
344 simcall_process_sleep(sleeptime);
345 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
351 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
354 int rank = smpi_process_index();
355 if (TRACE_smpi_view_internals()) {
356 TRACE_smpi_send(rank, rank, receiver,request->size);
359 /* if(receiver == MPI_UNDEFINED) {*/
360 /* XBT_WARN("Trying to send a message to a wrong rank");*/
363 print_request("New send", request);
364 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
365 mailbox = smpi_process_remote_mailbox_small(receiver);
367 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
368 mailbox = smpi_process_remote_mailbox(receiver);
371 void* buf = request->buf;
372 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
374 request->detached = 1;
376 if(request->old_type->has_subtype == 0){
377 oldbuf = request->buf;
378 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
379 if((smpi_privatize_global_variables)
380 && ((char*)request->buf >= start_data_exe)
381 && ((char*)request->buf < start_data_exe + size_data_exe )
383 XBT_WARN("Privatization : We are sending from a zone inside global memory. Switch data segment ");
384 switch_data_segment(smpi_process_index());
388 buf = xbt_malloc(request->size);
389 memcpy(buf,oldbuf,request->size);
392 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
395 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
396 request->real_size=request->size;
397 smpi_datatype_use(request->old_type);
398 smpi_comm_use(request->comm);
400 //if we are giving back the control to the user without waiting for completion, we have to inject timings
401 double sleeptime = 0.0;
402 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
403 //isend and send timings may be different
404 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
407 if(sleeptime != 0.0){
408 simcall_process_sleep(sleeptime);
409 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
413 simcall_comm_isend(mailbox, request->size, -1.0,
414 buf, request->real_size,
416 &xbt_free, // how to free the userdata if a detached send fails
418 // detach if msg size < eager/rdv switch limit
422 /* FIXME: detached sends are not traceable (request->action == NULL) */
424 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
432 void smpi_mpi_startall(int count, MPI_Request * requests)
435 if(requests==NULL) return;
437 for(i = 0; i < count; i++) {
438 smpi_mpi_start(requests[i]);
442 void smpi_mpi_request_free(MPI_Request * request)
444 if((*request) != MPI_REQUEST_NULL){
445 (*request)->refcount--;
446 if((*request)->refcount<0) xbt_die("wrong refcount");
448 if((*request)->refcount==0){
449 print_request("Destroying", (*request));
451 *request = MPI_REQUEST_NULL;
453 print_request("Decrementing", (*request));
456 xbt_die("freeing an already free request");
460 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
461 int dst, int tag, MPI_Comm comm)
463 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
464 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
465 comm, PERSISTENT | ISEND | SEND | PREPARED);
469 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
470 int dst, int tag, MPI_Comm comm)
472 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
473 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
474 comm, NON_PERSISTENT | ISEND | SEND);
475 smpi_mpi_start(request);
479 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
480 int dst, int tag, MPI_Comm comm)
482 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
483 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
484 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
485 smpi_mpi_start(request);
491 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
492 int src, int tag, MPI_Comm comm)
494 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
495 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
496 comm, PERSISTENT | RECV | PREPARED);
500 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
501 int src, int tag, MPI_Comm comm)
503 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
504 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
505 comm, NON_PERSISTENT | RECV);
506 smpi_mpi_start(request);
510 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
511 int tag, MPI_Comm comm, MPI_Status * status)
513 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
514 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
515 smpi_mpi_wait(&request, status);
521 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
522 int tag, MPI_Comm comm)
524 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
525 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
526 comm, NON_PERSISTENT | SEND);
528 smpi_mpi_start(request);
529 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
533 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
534 int dst, int tag, MPI_Comm comm)
536 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
537 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
538 comm, NON_PERSISTENT | SSEND | SEND);
540 smpi_mpi_start(request);
541 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
545 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
546 int dst, int sendtag, void *recvbuf, int recvcount,
547 MPI_Datatype recvtype, int src, int recvtag,
548 MPI_Comm comm, MPI_Status * status)
550 MPI_Request requests[2];
552 int myid=smpi_process_index();
553 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
554 smpi_datatype_copy(sendbuf, sendcount, sendtype,
555 recvbuf, recvcount, recvtype);
559 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
561 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
562 smpi_mpi_startall(2, requests);
563 smpi_mpi_waitall(2, requests, stats);
564 smpi_mpi_request_free(&requests[0]);
565 smpi_mpi_request_free(&requests[1]);
566 if(status != MPI_STATUS_IGNORE) {
567 // Copy receive status
572 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
574 return status->count / smpi_datatype_size(datatype);
577 static void finish_wait(MPI_Request * request, MPI_Status * status)
579 MPI_Request req = *request;
580 smpi_empty_status(status);
582 if(!(req->detached && req->flags & SEND)
583 && !(req->flags & PREPARED)){
584 if(status != MPI_STATUS_IGNORE) {
585 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
586 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
587 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
588 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
589 // this handles the case were size in receive differs from size in send
590 // FIXME: really this should just contain the count of receive-type blocks,
592 status->count = req->real_size;
595 print_request("Finishing", req);
596 MPI_Datatype datatype = req->old_type;
598 if(datatype->has_subtype == 1){
599 if (!_xbt_replay_is_active()){
600 if( smpi_privatize_global_variables
601 && ((char*)req->old_buf >= start_data_exe)
602 && ((char*)req->old_buf < start_data_exe + size_data_exe )
604 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
605 switch_data_segment(smpi_process_index());
608 // This part handles the problem of non-contignous memory
609 // the unserialization at the reception
610 s_smpi_subtype_t *subtype = datatype->substruct;
611 if(req->flags & RECV) {
612 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
614 if(req->detached == 0) free(req->buf);
616 smpi_comm_unuse(req->comm);
617 smpi_datatype_unuse(datatype);
622 if (TRACE_smpi_view_internals()) {
623 if(req->flags & RECV){
624 int rank = smpi_process_index();
625 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
626 TRACE_smpi_recv(rank, src_traced, rank);
631 if(req->detached_sender!=NULL){
632 smpi_mpi_request_free(&(req->detached_sender));
634 if(req->flags & PERSISTENT)
636 req->flags |= FINISHED;
638 smpi_mpi_request_free(request);
642 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
645 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
646 smpi_empty_status(status);
648 if (!((*request)->flags & PREPARED)) {
649 if ((*request)->action != NULL)
650 flag = simcall_comm_test((*request)->action);
652 finish_wait(request, status);
653 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
654 *request = MPI_REQUEST_NULL;
660 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
667 *index = MPI_UNDEFINED;
669 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
670 map = xbt_new(int, count);
672 for(i = 0; i < count; i++) {
673 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
674 !(requests[i]->flags & PREPARED)) {
675 xbt_dynar_push(comms, &requests[i]->action);
681 i = simcall_comm_testany(comms);
682 // not MPI_UNDEFINED, as this is a simix return code
685 finish_wait(&requests[*index], status);
686 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
687 requests[*index] = MPI_REQUEST_NULL;
691 //all requests are null or inactive, return true
693 smpi_empty_status(status);
696 xbt_dynar_free(&comms);
702 int smpi_mpi_testall(int count, MPI_Request requests[],
706 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
709 for(i=0; i<count; i++){
710 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
711 if (smpi_mpi_test(&requests[i], pstat)!=1){
714 requests[i]=MPI_REQUEST_NULL;
717 smpi_empty_status(pstat);
719 if(status != MPI_STATUSES_IGNORE) {
726 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
728 //FIXME find another wait to avoid busy waiting ?
729 // the issue here is that we have to wait on a nonexistent comm
731 smpi_mpi_iprobe(source, tag, comm, &flag, status);
732 XBT_DEBUG("Busy Waiting on probing : %d", flag);
736 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
738 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
739 comm, PERSISTENT | RECV);
741 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
742 double sleeptime= sg_cfg_get_double("smpi/iprobe");
743 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
744 static int nsleeps = 1;
746 simcall_process_sleep(sleeptime);
748 // behave like a receive, but don't do it
751 print_request("New iprobe", request);
752 // We have to test both mailboxes as we don't know if we will receive one one or another
753 if (sg_cfg_get_int("smpi/async_small_thres")>0){
754 mailbox = smpi_process_mailbox_small();
755 XBT_DEBUG("trying to probe the perm recv mailbox");
756 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
758 if (request->action==NULL){
759 mailbox = smpi_process_mailbox();
760 XBT_DEBUG("trying to probe the other mailbox");
761 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
765 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
767 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
768 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
769 status->MPI_TAG = req->tag;
770 status->MPI_ERROR = MPI_SUCCESS;
771 status->count = req->real_size;
773 nsleeps=1;//reset the number of sleeps we will do next time
779 smpi_mpi_request_free(&request);
784 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
786 print_request("Waiting", *request);
787 if ((*request)->flags & PREPARED) {
788 smpi_empty_status(status);
792 if ((*request)->action != NULL) { // this is not a detached send
793 simcall_comm_wait((*request)->action, -1.0);
795 if(MC_is_active() && (*request)->action)
796 (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison
800 finish_wait(request, status);
801 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
802 *request = MPI_REQUEST_NULL;
803 // FIXME for a detached send, finish_wait is not called:
806 int smpi_mpi_waitany(int count, MPI_Request requests[],
813 index = MPI_UNDEFINED;
815 // Wait for a request to complete
816 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
817 map = xbt_new(int, count);
819 XBT_DEBUG("Wait for one of %d", count);
820 for(i = 0; i < count; i++) {
821 if (requests[i] != MPI_REQUEST_NULL
822 && !(requests[i]->flags & PREPARED)
823 && !(requests[i]->flags & FINISHED)) {
824 if (requests[i]->action != NULL) {
825 XBT_DEBUG("Waiting any %p ", requests[i]);
826 xbt_dynar_push(comms, &requests[i]->action);
830 //This is a finished detached request, let's return this one
831 size=0;//so we free the dynar but don't do the waitany call
833 finish_wait(&requests[i], status);//cleanup if refcount = 0
834 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
835 requests[i]=MPI_REQUEST_NULL;//set to null
841 i = simcall_comm_waitany(comms);
843 // not MPI_UNDEFINED, as this is a simix return code
846 finish_wait(&requests[index], status);
847 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
848 requests[index] = MPI_REQUEST_NULL;
852 xbt_dynar_free(&comms);
855 if (index==MPI_UNDEFINED)
856 smpi_empty_status(status);
861 int smpi_mpi_waitall(int count, MPI_Request requests[],
866 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
867 int retvalue = MPI_SUCCESS;
868 //tag invalid requests in the set
869 if (status != MPI_STATUSES_IGNORE) {
870 for (c = 0; c < count; c++) {
871 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
872 (requests[c]->flags & PREPARED)) {
873 smpi_empty_status(&status[c]);
874 } else if (requests[c]->src == MPI_PROC_NULL) {
875 smpi_empty_status(&status[c]);
876 status[c].MPI_SOURCE = MPI_PROC_NULL;
880 for(c = 0; c < count; c++) {
882 if (MC_is_active()) {
883 smpi_mpi_wait(&requests[c], pstat);
886 index = smpi_mpi_waitany(count, requests, pstat);
887 if (index == MPI_UNDEFINED)
889 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
890 requests[index]=MPI_REQUEST_NULL;
892 if (status != MPI_STATUSES_IGNORE) {
893 status[index] = *pstat;
894 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
895 retvalue = MPI_ERR_IN_STATUS;
902 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
907 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
910 for(i = 0; i < incount; i++)
912 index=smpi_mpi_waitany(incount, requests, pstat);
913 if(index!=MPI_UNDEFINED){
914 indices[count] = index;
916 if(status != MPI_STATUSES_IGNORE) {
917 status[index] = *pstat;
919 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
920 requests[index]=MPI_REQUEST_NULL;
922 return MPI_UNDEFINED;
928 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
931 int i, count, count_dead;
933 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
937 for(i = 0; i < incount; i++) {
938 if((requests[i] != MPI_REQUEST_NULL)) {
939 if(smpi_mpi_test(&requests[i], pstat)) {
942 if(status != MPI_STATUSES_IGNORE) {
945 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
946 requests[i]=MPI_REQUEST_NULL;
952 if(count_dead==incount)return MPI_UNDEFINED;
956 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
959 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
960 nary_tree_bcast(buf, count, datatype, root, comm, 4);
963 void smpi_mpi_barrier(MPI_Comm comm)
965 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
966 nary_tree_barrier(comm, 4);
969 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
970 void *recvbuf, int recvcount, MPI_Datatype recvtype,
971 int root, MPI_Comm comm)
973 int system_tag = COLL_TAG_GATHER;
974 int rank, size, src, index;
975 MPI_Aint lb = 0, recvext = 0;
976 MPI_Request *requests;
978 rank = smpi_comm_rank(comm);
979 size = smpi_comm_size(comm);
981 // Send buffer to root
982 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
984 // FIXME: check for errors
985 smpi_datatype_extent(recvtype, &lb, &recvext);
986 // Local copy from root
987 smpi_datatype_copy(sendbuf, sendcount, sendtype,
988 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
989 // Receive buffers from senders
990 requests = xbt_new(MPI_Request, size - 1);
992 for(src = 0; src < size; src++) {
994 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
996 src, system_tag, comm);
1000 // Wait for completion of irecv's.
1001 smpi_mpi_startall(size - 1, requests);
1002 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1003 for(src = 0; src < size-1; src++) {
1004 smpi_mpi_request_free(&requests[src]);
1011 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1012 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1016 int rank = smpi_process_index();
1019 /* arbitrarily choose root as rank 0 */
1020 size = smpi_comm_size(comm);
1022 displs = xbt_new(int, size);
1023 for (i = 0; i < size; i++) {
1025 count += recvcounts[i];
1027 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1028 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1029 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1030 recvcounts[rank], datatype, 0, comm);
1035 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1036 void *recvbuf, int *recvcounts, int *displs,
1037 MPI_Datatype recvtype, int root, MPI_Comm comm)
1039 int system_tag = COLL_TAG_GATHERV;
1040 int rank, size, src, index;
1041 MPI_Aint lb = 0, recvext = 0;
1042 MPI_Request *requests;
1044 rank = smpi_comm_rank(comm);
1045 size = smpi_comm_size(comm);
1047 // Send buffer to root
1048 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1050 // FIXME: check for errors
1051 smpi_datatype_extent(recvtype, &lb, &recvext);
1052 // Local copy from root
1053 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1054 (char *)recvbuf + displs[root] * recvext,
1055 recvcounts[root], recvtype);
1056 // Receive buffers from senders
1057 requests = xbt_new(MPI_Request, size - 1);
1059 for(src = 0; src < size; src++) {
1062 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1063 recvcounts[src], recvtype, src, system_tag, comm);
1067 // Wait for completion of irecv's.
1068 smpi_mpi_startall(size - 1, requests);
1069 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1070 for(src = 0; src < size-1; src++) {
1071 smpi_mpi_request_free(&requests[src]);
1077 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1078 MPI_Datatype sendtype, void *recvbuf,
1079 int recvcount, MPI_Datatype recvtype,
1082 int system_tag = COLL_TAG_ALLGATHER;
1083 int rank, size, other, index;
1084 MPI_Aint lb = 0, recvext = 0;
1085 MPI_Request *requests;
1087 rank = smpi_comm_rank(comm);
1088 size = smpi_comm_size(comm);
1089 // FIXME: check for errors
1090 smpi_datatype_extent(recvtype, &lb, &recvext);
1091 // Local copy from self
1092 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1093 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1095 // Send/Recv buffers to/from others;
1096 requests = xbt_new(MPI_Request, 2 * (size - 1));
1098 for(other = 0; other < size; other++) {
1101 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1104 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1105 recvcount, recvtype, other,
1110 // Wait for completion of all comms.
1111 smpi_mpi_startall(2 * (size - 1), requests);
1112 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1113 for(other = 0; other < 2*(size-1); other++) {
1114 smpi_mpi_request_free(&requests[other]);
1119 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1120 MPI_Datatype sendtype, void *recvbuf,
1121 int *recvcounts, int *displs,
1122 MPI_Datatype recvtype, MPI_Comm comm)
1124 int system_tag = COLL_TAG_ALLGATHERV;
1125 int rank, size, other, index;
1126 MPI_Aint lb = 0, recvext = 0;
1127 MPI_Request *requests;
1129 rank = smpi_comm_rank(comm);
1130 size = smpi_comm_size(comm);
1131 // FIXME: check for errors
1132 smpi_datatype_extent(recvtype, &lb, &recvext);
1133 // Local copy from self
1134 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1135 (char *)recvbuf + displs[rank] * recvext,
1136 recvcounts[rank], recvtype);
1137 // Send buffers to others;
1138 requests = xbt_new(MPI_Request, 2 * (size - 1));
1140 for(other = 0; other < size; other++) {
1143 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1147 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1148 recvtype, other, system_tag, comm);
1152 // Wait for completion of all comms.
1153 smpi_mpi_startall(2 * (size - 1), requests);
1154 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1155 for(other = 0; other < 2*(size-1); other++) {
1156 smpi_mpi_request_free(&requests[other]);
1161 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1162 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1163 int root, MPI_Comm comm)
1165 int system_tag = COLL_TAG_SCATTER;
1166 int rank, size, dst, index;
1167 MPI_Aint lb = 0, sendext = 0;
1168 MPI_Request *requests;
1170 rank = smpi_comm_rank(comm);
1171 size = smpi_comm_size(comm);
1173 // Recv buffer from root
1174 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1177 // FIXME: check for errors
1178 smpi_datatype_extent(sendtype, &lb, &sendext);
1179 // Local copy from root
1180 if(recvbuf!=MPI_IN_PLACE){
1181 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1182 sendcount, sendtype, recvbuf, recvcount, recvtype);
1184 // Send buffers to receivers
1185 requests = xbt_new(MPI_Request, size - 1);
1187 for(dst = 0; dst < size; dst++) {
1189 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1190 sendcount, sendtype, dst,
1195 // Wait for completion of isend's.
1196 smpi_mpi_startall(size - 1, requests);
1197 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1198 for(dst = 0; dst < size-1; dst++) {
1199 smpi_mpi_request_free(&requests[dst]);
1205 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1206 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1207 MPI_Datatype recvtype, int root, MPI_Comm comm)
1209 int system_tag = COLL_TAG_SCATTERV;
1210 int rank, size, dst, index;
1211 MPI_Aint lb = 0, sendext = 0;
1212 MPI_Request *requests;
1214 rank = smpi_comm_rank(comm);
1215 size = smpi_comm_size(comm);
1217 // Recv buffer from root
1218 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1221 // FIXME: check for errors
1222 smpi_datatype_extent(sendtype, &lb, &sendext);
1223 // Local copy from root
1224 if(recvbuf!=MPI_IN_PLACE){
1225 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1226 sendtype, recvbuf, recvcount, recvtype);
1228 // Send buffers to receivers
1229 requests = xbt_new(MPI_Request, size - 1);
1231 for(dst = 0; dst < size; dst++) {
1234 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1235 sendtype, dst, system_tag, comm);
1239 // Wait for completion of isend's.
1240 smpi_mpi_startall(size - 1, requests);
1241 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1242 for(dst = 0; dst < size-1; dst++) {
1243 smpi_mpi_request_free(&requests[dst]);
1249 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1250 MPI_Datatype datatype, MPI_Op op, int root,
1253 int system_tag = COLL_TAG_REDUCE;
1254 int rank, size, src, index;
1255 MPI_Aint lb = 0, dataext = 0;
1256 MPI_Request *requests;
1260 char* sendtmpbuf = (char*) sendbuf;
1261 if( sendbuf == MPI_IN_PLACE ) {
1262 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1263 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1266 rank = smpi_comm_rank(comm);
1267 size = smpi_comm_size(comm);
1268 //non commutative case, use a working algo from openmpi
1269 if(!smpi_op_is_commute(op)){
1270 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1271 datatype, op, root, comm);
1276 // Send buffer to root
1277 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1279 // FIXME: check for errors
1280 smpi_datatype_extent(datatype, &lb, &dataext);
1281 // Local copy from root
1282 if (sendtmpbuf && recvbuf)
1283 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1284 // Receive buffers from senders
1285 //TODO: make a MPI_barrier here ?
1286 requests = xbt_new(MPI_Request, size - 1);
1287 tmpbufs = xbt_new(void *, size - 1);
1289 for(src = 0; src < size; src++) {
1291 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1293 tmpbufs[index] = xbt_malloc(count * dataext);
1295 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1300 // Wait for completion of irecv's.
1301 smpi_mpi_startall(size - 1, requests);
1302 for(src = 0; src < size - 1; src++) {
1303 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1304 XBT_DEBUG("finished waiting any request with index %d", index);
1305 if(index == MPI_UNDEFINED) {
1308 smpi_mpi_request_free(&requests[index]);
1310 if(op) /* op can be MPI_OP_NULL that does nothing */
1311 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1313 for(index = 0; index < size - 1; index++) {
1314 xbt_free(tmpbufs[index]);
1319 if( sendbuf == MPI_IN_PLACE ) {
1320 xbt_free(sendtmpbuf);
1325 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1326 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1328 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1329 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1332 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1333 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1335 int system_tag = -888;
1336 int rank, size, other, index;
1337 MPI_Aint lb = 0, dataext = 0;
1338 MPI_Request *requests;
1341 rank = smpi_comm_rank(comm);
1342 size = smpi_comm_size(comm);
1344 // FIXME: check for errors
1345 smpi_datatype_extent(datatype, &lb, &dataext);
1347 // Local copy from self
1348 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1350 // Send/Recv buffers to/from others;
1351 requests = xbt_new(MPI_Request, size - 1);
1352 tmpbufs = xbt_new(void *, rank);
1354 for(other = 0; other < rank; other++) {
1355 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1357 tmpbufs[index] = xbt_malloc(count * dataext);
1359 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1363 for(other = rank + 1; other < size; other++) {
1365 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1368 // Wait for completion of all comms.
1369 smpi_mpi_startall(size - 1, requests);
1371 if(smpi_op_is_commute(op)){
1372 for(other = 0; other < size - 1; other++) {
1373 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1374 if(index == MPI_UNDEFINED) {
1378 // #Request is below rank: it's a irecv
1379 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1383 //non commutative case, wait in order
1384 for(other = 0; other < size - 1; other++) {
1385 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1387 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1391 for(index = 0; index < rank; index++) {
1392 xbt_free(tmpbufs[index]);
1394 for(index = 0; index < size-1; index++) {
1395 smpi_mpi_request_free(&requests[index]);
1401 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1402 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1404 int system_tag = -888;
1405 int rank, size, other, index;
1406 MPI_Aint lb = 0, dataext = 0;
1407 MPI_Request *requests;
1409 int recvbuf_is_empty=1;
1410 rank = smpi_comm_rank(comm);
1411 size = smpi_comm_size(comm);
1413 // FIXME: check for errors
1414 smpi_datatype_extent(datatype, &lb, &dataext);
1416 // Send/Recv buffers to/from others;
1417 requests = xbt_new(MPI_Request, size - 1);
1418 tmpbufs = xbt_new(void *, rank);
1420 for(other = 0; other < rank; other++) {
1421 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1423 tmpbufs[index] = xbt_malloc(count * dataext);
1425 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1429 for(other = rank + 1; other < size; other++) {
1431 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1434 // Wait for completion of all comms.
1435 smpi_mpi_startall(size - 1, requests);
1436 if(smpi_op_is_commute(op)){
1437 for(other = 0; other < size - 1; other++) {
1438 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1439 if(index == MPI_UNDEFINED) {
1443 if(recvbuf_is_empty){
1444 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1447 // #Request is below rank: it's a irecv
1448 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1452 //non commutative case, wait in order
1453 for(other = 0; other < size - 1; other++) {
1454 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1456 if(recvbuf_is_empty){
1457 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1459 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1463 for(index = 0; index < rank; index++) {
1464 xbt_free(tmpbufs[index]);
1466 for(index = 0; index < size-1; index++) {
1467 smpi_mpi_request_free(&requests[index]);