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 if (action_fp == NULL)
264 xbt_die("Cannot open %s: %s", path, strerror(errno));
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->flags & RMA || 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,
339 (request->flags & ACCUMULATE)? NULL : &smpi_comm_copy_buffer_callback,
342 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
343 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
345 simcall_process_sleep(sleeptime);
346 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
352 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
355 int rank = smpi_process_index();
356 if (TRACE_smpi_view_internals()) {
357 TRACE_smpi_send(rank, rank, receiver,request->size);
360 /* if(receiver == MPI_UNDEFINED) {*/
361 /* XBT_WARN("Trying to send a message to a wrong rank");*/
364 print_request("New send", request);
365 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
366 mailbox = smpi_process_remote_mailbox_small(receiver);
368 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
369 mailbox = smpi_process_remote_mailbox(receiver);
372 void* buf = request->buf;
373 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
375 request->detached = 1;
377 if(request->old_type->has_subtype == 0){
378 oldbuf = request->buf;
379 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
380 if((smpi_privatize_global_variables)
381 && ((char*)request->buf >= start_data_exe)
382 && ((char*)request->buf < start_data_exe + size_data_exe )){
383 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
384 switch_data_segment(smpi_process_index());
386 buf = xbt_malloc(request->size);
387 memcpy(buf,oldbuf,request->size);
390 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
393 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
394 request->real_size=request->size;
395 smpi_datatype_use(request->old_type);
396 smpi_comm_use(request->comm);
398 //if we are giving back the control to the user without waiting for completion, we have to inject timings
399 double sleeptime = 0.0;
400 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
401 //isend and send timings may be different
402 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
405 if(sleeptime != 0.0){
406 simcall_process_sleep(sleeptime);
407 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
411 simcall_comm_isend(mailbox, request->size, -1.0,
412 buf, request->real_size,
414 &xbt_free_f, // how to free the userdata if a detached send fails
415 (request->flags & ACCUMULATE)? NULL : &smpi_comm_copy_buffer_callback,
417 // detach if msg size < eager/rdv switch limit
421 /* FIXME: detached sends are not traceable (request->action == NULL) */
423 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
431 void smpi_mpi_startall(int count, MPI_Request * requests)
434 if(requests==NULL) return;
436 for(i = 0; i < count; i++) {
437 smpi_mpi_start(requests[i]);
441 void smpi_mpi_request_free(MPI_Request * request)
443 if((*request) != MPI_REQUEST_NULL){
444 (*request)->refcount--;
445 if((*request)->refcount<0) xbt_die("wrong refcount");
447 if((*request)->refcount==0){
448 print_request("Destroying", (*request));
450 *request = MPI_REQUEST_NULL;
452 print_request("Decrementing", (*request));
455 xbt_die("freeing an already free request");
460 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
461 int src, 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, src, dst, tag,
465 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
469 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
470 int src, 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, src, dst, tag,
474 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
479 MPI_Request smpi_isend_init(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, PERSISTENT | ISEND | SEND | PREPARED);
488 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
489 int dst, int tag, MPI_Comm comm)
491 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
492 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
493 comm, NON_PERSISTENT | ISEND | SEND);
494 smpi_mpi_start(request);
498 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
499 int dst, int tag, MPI_Comm comm)
501 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
502 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
503 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
504 smpi_mpi_start(request);
508 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
509 int src, int tag, MPI_Comm comm)
511 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
512 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,
513 comm, PERSISTENT | RECV | PREPARED);
517 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
518 int src, int tag, MPI_Comm comm)
520 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
521 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,
522 comm, NON_PERSISTENT | RECV);
523 smpi_mpi_start(request);
527 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
528 int tag, MPI_Comm comm, MPI_Status * status)
530 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
531 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
532 smpi_mpi_wait(&request, status);
538 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
539 int tag, MPI_Comm comm)
541 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
542 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
543 comm, NON_PERSISTENT | SEND);
545 smpi_mpi_start(request);
546 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
550 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
551 int dst, int tag, MPI_Comm comm)
553 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
554 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
555 comm, NON_PERSISTENT | SSEND | SEND);
557 smpi_mpi_start(request);
558 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
562 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
563 int dst, int sendtag, void *recvbuf, int recvcount,
564 MPI_Datatype recvtype, int src, int recvtag,
565 MPI_Comm comm, MPI_Status * status)
567 MPI_Request requests[2];
569 int myid=smpi_process_index();
570 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
571 smpi_datatype_copy(sendbuf, sendcount, sendtype,
572 recvbuf, recvcount, recvtype);
576 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
578 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
579 smpi_mpi_startall(2, requests);
580 smpi_mpi_waitall(2, requests, stats);
581 smpi_mpi_request_free(&requests[0]);
582 smpi_mpi_request_free(&requests[1]);
583 if(status != MPI_STATUS_IGNORE) {
584 // Copy receive status
589 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
591 return status->count / smpi_datatype_size(datatype);
594 static void finish_wait(MPI_Request * request, MPI_Status * status)
596 MPI_Request req = *request;
597 smpi_empty_status(status);
599 if(!(req->detached && req->flags & SEND)
600 && !(req->flags & PREPARED)){
601 if(status != MPI_STATUS_IGNORE) {
602 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
603 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
604 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
605 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
606 // this handles the case were size in receive differs from size in send
607 // FIXME: really this should just contain the count of receive-type blocks,
609 status->count = req->real_size;
612 print_request("Finishing", req);
613 MPI_Datatype datatype = req->old_type;
615 if(datatype->has_subtype == 1){
616 if (!_xbt_replay_is_active()){
617 if( smpi_privatize_global_variables
618 && ((char*)req->old_buf >= start_data_exe)
619 && ((char*)req->old_buf < start_data_exe + size_data_exe )
621 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
622 switch_data_segment(smpi_process_index());
625 // This part handles the problem of non-contignous memory
626 // the unserialization at the reception
627 s_smpi_subtype_t *subtype = datatype->substruct;
628 if(req->flags & RECV) {
629 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
631 if(req->detached == 0) free(req->buf);
633 smpi_comm_unuse(req->comm);
634 smpi_datatype_unuse(datatype);
639 if (TRACE_smpi_view_internals()) {
640 if(req->flags & RECV){
641 int rank = smpi_process_index();
642 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
643 TRACE_smpi_recv(rank, src_traced, rank);
648 if(req->detached_sender!=NULL){
649 smpi_mpi_request_free(&(req->detached_sender));
651 if(req->flags & PERSISTENT)
653 req->flags |= FINISHED;
655 smpi_mpi_request_free(request);
659 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
662 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
663 smpi_empty_status(status);
665 if (!((*request)->flags & PREPARED)) {
666 if ((*request)->action != NULL)
667 flag = simcall_comm_test((*request)->action);
669 finish_wait(request, status);
670 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
671 *request = MPI_REQUEST_NULL;
677 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
684 *index = MPI_UNDEFINED;
686 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
687 map = xbt_new(int, count);
689 for(i = 0; i < count; i++) {
690 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
691 !(requests[i]->flags & PREPARED)) {
692 xbt_dynar_push(comms, &requests[i]->action);
698 i = simcall_comm_testany(comms);
699 // not MPI_UNDEFINED, as this is a simix return code
702 finish_wait(&requests[*index], status);
703 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
704 requests[*index] = MPI_REQUEST_NULL;
708 //all requests are null or inactive, return true
710 smpi_empty_status(status);
713 xbt_dynar_free(&comms);
719 int smpi_mpi_testall(int count, MPI_Request requests[],
723 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
726 for(i=0; i<count; i++){
727 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
728 if (smpi_mpi_test(&requests[i], pstat)!=1){
731 requests[i]=MPI_REQUEST_NULL;
734 smpi_empty_status(pstat);
736 if(status != MPI_STATUSES_IGNORE) {
743 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
745 //FIXME find another wait to avoid busy waiting ?
746 // the issue here is that we have to wait on a nonexistent comm
748 smpi_mpi_iprobe(source, tag, comm, &flag, status);
749 XBT_DEBUG("Busy Waiting on probing : %d", flag);
753 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
755 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,
756 comm, PERSISTENT | RECV);
758 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
759 double sleeptime= sg_cfg_get_double("smpi/iprobe");
760 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
761 static int nsleeps = 1;
763 simcall_process_sleep(sleeptime);
765 // behave like a receive, but don't do it
768 print_request("New iprobe", request);
769 // We have to test both mailboxes as we don't know if we will receive one one or another
770 if (sg_cfg_get_int("smpi/async_small_thres")>0){
771 mailbox = smpi_process_mailbox_small();
772 XBT_DEBUG("trying to probe the perm recv mailbox");
773 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
775 if (request->action==NULL){
776 mailbox = smpi_process_mailbox();
777 XBT_DEBUG("trying to probe the other mailbox");
778 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
782 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
784 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
785 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
786 status->MPI_TAG = req->tag;
787 status->MPI_ERROR = MPI_SUCCESS;
788 status->count = req->real_size;
790 nsleeps=1;//reset the number of sleeps we will do next time
796 smpi_mpi_request_free(&request);
801 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
803 print_request("Waiting", *request);
804 if ((*request)->flags & PREPARED) {
805 smpi_empty_status(status);
809 if ((*request)->action != NULL) { // this is not a detached send
810 simcall_comm_wait((*request)->action, -1.0);
812 if(MC_is_active() && (*request)->action)
813 (*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
817 finish_wait(request, status);
818 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
819 *request = MPI_REQUEST_NULL;
820 // FIXME for a detached send, finish_wait is not called:
823 int smpi_mpi_waitany(int count, MPI_Request requests[],
830 index = MPI_UNDEFINED;
832 // Wait for a request to complete
833 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
834 map = xbt_new(int, count);
836 XBT_DEBUG("Wait for one of %d", count);
837 for(i = 0; i < count; i++) {
838 if (requests[i] != MPI_REQUEST_NULL
839 && !(requests[i]->flags & PREPARED)
840 && !(requests[i]->flags & FINISHED)) {
841 if (requests[i]->action != NULL) {
842 XBT_DEBUG("Waiting any %p ", requests[i]);
843 xbt_dynar_push(comms, &requests[i]->action);
847 //This is a finished detached request, let's return this one
848 size=0;//so we free the dynar but don't do the waitany call
850 finish_wait(&requests[i], status);//cleanup if refcount = 0
851 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
852 requests[i]=MPI_REQUEST_NULL;//set to null
858 i = simcall_comm_waitany(comms);
860 // not MPI_UNDEFINED, as this is a simix return code
863 finish_wait(&requests[index], status);
864 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
865 requests[index] = MPI_REQUEST_NULL;
869 xbt_dynar_free(&comms);
872 if (index==MPI_UNDEFINED)
873 smpi_empty_status(status);
878 int smpi_mpi_waitall(int count, MPI_Request requests[],
883 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
884 int retvalue = MPI_SUCCESS;
885 //tag invalid requests in the set
886 if (status != MPI_STATUSES_IGNORE) {
887 for (c = 0; c < count; c++) {
888 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
889 (requests[c]->flags & PREPARED)) {
890 smpi_empty_status(&status[c]);
891 } else if (requests[c]->src == MPI_PROC_NULL) {
892 smpi_empty_status(&status[c]);
893 status[c].MPI_SOURCE = MPI_PROC_NULL;
897 for(c = 0; c < count; c++) {
899 if (MC_is_active()) {
900 smpi_mpi_wait(&requests[c], pstat);
903 index = smpi_mpi_waitany(count, requests, pstat);
904 if (index == MPI_UNDEFINED)
906 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
907 requests[index]=MPI_REQUEST_NULL;
909 if (status != MPI_STATUSES_IGNORE) {
910 status[index] = *pstat;
911 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
912 retvalue = MPI_ERR_IN_STATUS;
919 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
924 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
927 for(i = 0; i < incount; i++)
929 index=smpi_mpi_waitany(incount, requests, pstat);
930 if(index!=MPI_UNDEFINED){
931 indices[count] = index;
933 if(status != MPI_STATUSES_IGNORE) {
934 status[index] = *pstat;
936 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
937 requests[index]=MPI_REQUEST_NULL;
939 return MPI_UNDEFINED;
945 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
948 int i, count, count_dead;
950 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
954 for(i = 0; i < incount; i++) {
955 if((requests[i] != MPI_REQUEST_NULL)) {
956 if(smpi_mpi_test(&requests[i], pstat)) {
959 if(status != MPI_STATUSES_IGNORE) {
962 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
963 requests[i]=MPI_REQUEST_NULL;
969 if(count_dead==incount)return MPI_UNDEFINED;
973 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
976 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
977 nary_tree_bcast(buf, count, datatype, root, comm, 4);
980 void smpi_mpi_barrier(MPI_Comm comm)
982 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
983 nary_tree_barrier(comm, 4);
986 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
987 void *recvbuf, int recvcount, MPI_Datatype recvtype,
988 int root, MPI_Comm comm)
990 int system_tag = COLL_TAG_GATHER;
991 int rank, size, src, index;
992 MPI_Aint lb = 0, recvext = 0;
993 MPI_Request *requests;
995 rank = smpi_comm_rank(comm);
996 size = smpi_comm_size(comm);
998 // Send buffer to root
999 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1001 // FIXME: check for errors
1002 smpi_datatype_extent(recvtype, &lb, &recvext);
1003 // Local copy from root
1004 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1005 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1006 // Receive buffers from senders
1007 requests = xbt_new(MPI_Request, size - 1);
1009 for(src = 0; src < size; src++) {
1011 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1012 recvcount, recvtype,
1013 src, system_tag, comm);
1017 // Wait for completion of irecv's.
1018 smpi_mpi_startall(size - 1, requests);
1019 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1020 for(src = 0; src < size-1; src++) {
1021 smpi_mpi_request_free(&requests[src]);
1028 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1029 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1033 int rank = smpi_process_index();
1036 /* arbitrarily choose root as rank 0 */
1037 size = smpi_comm_size(comm);
1039 displs = xbt_new(int, size);
1040 for (i = 0; i < size; i++) {
1042 count += recvcounts[i];
1044 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1045 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1046 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1047 recvcounts[rank], datatype, 0, comm);
1052 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1053 void *recvbuf, int *recvcounts, int *displs,
1054 MPI_Datatype recvtype, int root, MPI_Comm comm)
1056 int system_tag = COLL_TAG_GATHERV;
1057 int rank, size, src, index;
1058 MPI_Aint lb = 0, recvext = 0;
1059 MPI_Request *requests;
1061 rank = smpi_comm_rank(comm);
1062 size = smpi_comm_size(comm);
1064 // Send buffer to root
1065 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1067 // FIXME: check for errors
1068 smpi_datatype_extent(recvtype, &lb, &recvext);
1069 // Local copy from root
1070 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1071 (char *)recvbuf + displs[root] * recvext,
1072 recvcounts[root], recvtype);
1073 // Receive buffers from senders
1074 requests = xbt_new(MPI_Request, size - 1);
1076 for(src = 0; src < size; src++) {
1079 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1080 recvcounts[src], recvtype, src, system_tag, comm);
1084 // Wait for completion of irecv's.
1085 smpi_mpi_startall(size - 1, requests);
1086 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1087 for(src = 0; src < size-1; src++) {
1088 smpi_mpi_request_free(&requests[src]);
1094 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1095 MPI_Datatype sendtype, void *recvbuf,
1096 int recvcount, MPI_Datatype recvtype,
1099 int system_tag = COLL_TAG_ALLGATHER;
1100 int rank, size, other, index;
1101 MPI_Aint lb = 0, recvext = 0;
1102 MPI_Request *requests;
1104 rank = smpi_comm_rank(comm);
1105 size = smpi_comm_size(comm);
1106 // FIXME: check for errors
1107 smpi_datatype_extent(recvtype, &lb, &recvext);
1108 // Local copy from self
1109 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1110 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1112 // Send/Recv buffers to/from others;
1113 requests = xbt_new(MPI_Request, 2 * (size - 1));
1115 for(other = 0; other < size; other++) {
1118 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1121 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1122 recvcount, recvtype, other,
1127 // Wait for completion of all comms.
1128 smpi_mpi_startall(2 * (size - 1), requests);
1129 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1130 for(other = 0; other < 2*(size-1); other++) {
1131 smpi_mpi_request_free(&requests[other]);
1136 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1137 MPI_Datatype sendtype, void *recvbuf,
1138 int *recvcounts, int *displs,
1139 MPI_Datatype recvtype, MPI_Comm comm)
1141 int system_tag = COLL_TAG_ALLGATHERV;
1142 int rank, size, other, index;
1143 MPI_Aint lb = 0, recvext = 0;
1144 MPI_Request *requests;
1146 rank = smpi_comm_rank(comm);
1147 size = smpi_comm_size(comm);
1148 // FIXME: check for errors
1149 smpi_datatype_extent(recvtype, &lb, &recvext);
1150 // Local copy from self
1151 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1152 (char *)recvbuf + displs[rank] * recvext,
1153 recvcounts[rank], recvtype);
1154 // Send buffers to others;
1155 requests = xbt_new(MPI_Request, 2 * (size - 1));
1157 for(other = 0; other < size; other++) {
1160 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1164 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1165 recvtype, other, system_tag, comm);
1169 // Wait for completion of all comms.
1170 smpi_mpi_startall(2 * (size - 1), requests);
1171 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1172 for(other = 0; other < 2*(size-1); other++) {
1173 smpi_mpi_request_free(&requests[other]);
1178 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1179 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1180 int root, MPI_Comm comm)
1182 int system_tag = COLL_TAG_SCATTER;
1183 int rank, size, dst, index;
1184 MPI_Aint lb = 0, sendext = 0;
1185 MPI_Request *requests;
1187 rank = smpi_comm_rank(comm);
1188 size = smpi_comm_size(comm);
1190 // Recv buffer from root
1191 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1194 // FIXME: check for errors
1195 smpi_datatype_extent(sendtype, &lb, &sendext);
1196 // Local copy from root
1197 if(recvbuf!=MPI_IN_PLACE){
1198 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1199 sendcount, sendtype, recvbuf, recvcount, recvtype);
1201 // Send buffers to receivers
1202 requests = xbt_new(MPI_Request, size - 1);
1204 for(dst = 0; dst < size; dst++) {
1206 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1207 sendcount, sendtype, dst,
1212 // Wait for completion of isend's.
1213 smpi_mpi_startall(size - 1, requests);
1214 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1215 for(dst = 0; dst < size-1; dst++) {
1216 smpi_mpi_request_free(&requests[dst]);
1222 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1223 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1224 MPI_Datatype recvtype, int root, MPI_Comm comm)
1226 int system_tag = COLL_TAG_SCATTERV;
1227 int rank, size, dst, index;
1228 MPI_Aint lb = 0, sendext = 0;
1229 MPI_Request *requests;
1231 rank = smpi_comm_rank(comm);
1232 size = smpi_comm_size(comm);
1234 // Recv buffer from root
1235 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1238 // FIXME: check for errors
1239 smpi_datatype_extent(sendtype, &lb, &sendext);
1240 // Local copy from root
1241 if(recvbuf!=MPI_IN_PLACE){
1242 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1243 sendtype, recvbuf, recvcount, recvtype);
1245 // Send buffers to receivers
1246 requests = xbt_new(MPI_Request, size - 1);
1248 for(dst = 0; dst < size; dst++) {
1251 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1252 sendtype, dst, system_tag, comm);
1256 // Wait for completion of isend's.
1257 smpi_mpi_startall(size - 1, requests);
1258 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1259 for(dst = 0; dst < size-1; dst++) {
1260 smpi_mpi_request_free(&requests[dst]);
1266 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1267 MPI_Datatype datatype, MPI_Op op, int root,
1270 int system_tag = COLL_TAG_REDUCE;
1271 int rank, size, src, index;
1272 MPI_Aint lb = 0, dataext = 0;
1273 MPI_Request *requests;
1277 char* sendtmpbuf = (char*) sendbuf;
1278 if( sendbuf == MPI_IN_PLACE ) {
1279 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1280 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1283 rank = smpi_comm_rank(comm);
1284 size = smpi_comm_size(comm);
1285 //non commutative case, use a working algo from openmpi
1286 if(!smpi_op_is_commute(op)){
1287 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1288 datatype, op, root, comm);
1293 // Send buffer to root
1294 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1296 // FIXME: check for errors
1297 smpi_datatype_extent(datatype, &lb, &dataext);
1298 // Local copy from root
1299 if (sendtmpbuf && recvbuf)
1300 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1301 // Receive buffers from senders
1302 //TODO: make a MPI_barrier here ?
1303 requests = xbt_new(MPI_Request, size - 1);
1304 tmpbufs = xbt_new(void *, size - 1);
1306 for(src = 0; src < size; src++) {
1308 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1310 tmpbufs[index] = xbt_malloc(count * dataext);
1312 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1317 // Wait for completion of irecv's.
1318 smpi_mpi_startall(size - 1, requests);
1319 for(src = 0; src < size - 1; src++) {
1320 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1321 XBT_DEBUG("finished waiting any request with index %d", index);
1322 if(index == MPI_UNDEFINED) {
1325 smpi_mpi_request_free(&requests[index]);
1327 if(op) /* op can be MPI_OP_NULL that does nothing */
1328 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1330 for(index = 0; index < size - 1; index++) {
1331 xbt_free(tmpbufs[index]);
1336 if( sendbuf == MPI_IN_PLACE ) {
1337 xbt_free(sendtmpbuf);
1342 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1343 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1345 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1346 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1349 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1350 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1352 int system_tag = -888;
1353 int rank, size, other, index;
1354 MPI_Aint lb = 0, dataext = 0;
1355 MPI_Request *requests;
1358 rank = smpi_comm_rank(comm);
1359 size = smpi_comm_size(comm);
1361 // FIXME: check for errors
1362 smpi_datatype_extent(datatype, &lb, &dataext);
1364 // Local copy from self
1365 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1367 // Send/Recv buffers to/from others;
1368 requests = xbt_new(MPI_Request, size - 1);
1369 tmpbufs = xbt_new(void *, rank);
1371 for(other = 0; other < rank; other++) {
1372 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1374 tmpbufs[index] = xbt_malloc(count * dataext);
1376 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1380 for(other = rank + 1; other < size; other++) {
1382 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1385 // Wait for completion of all comms.
1386 smpi_mpi_startall(size - 1, requests);
1388 if(smpi_op_is_commute(op)){
1389 for(other = 0; other < size - 1; other++) {
1390 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1391 if(index == MPI_UNDEFINED) {
1395 // #Request is below rank: it's a irecv
1396 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1400 //non commutative case, wait in order
1401 for(other = 0; other < size - 1; other++) {
1402 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1404 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1408 for(index = 0; index < rank; index++) {
1409 xbt_free(tmpbufs[index]);
1411 for(index = 0; index < size-1; index++) {
1412 smpi_mpi_request_free(&requests[index]);
1418 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1419 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1421 int system_tag = -888;
1422 int rank, size, other, index;
1423 MPI_Aint lb = 0, dataext = 0;
1424 MPI_Request *requests;
1426 int recvbuf_is_empty=1;
1427 rank = smpi_comm_rank(comm);
1428 size = smpi_comm_size(comm);
1430 // FIXME: check for errors
1431 smpi_datatype_extent(datatype, &lb, &dataext);
1433 // Send/Recv buffers to/from others;
1434 requests = xbt_new(MPI_Request, size - 1);
1435 tmpbufs = xbt_new(void *, rank);
1437 for(other = 0; other < rank; other++) {
1438 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1440 tmpbufs[index] = xbt_malloc(count * dataext);
1442 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1446 for(other = rank + 1; other < size; other++) {
1448 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1451 // Wait for completion of all comms.
1452 smpi_mpi_startall(size - 1, requests);
1453 if(smpi_op_is_commute(op)){
1454 for(other = 0; other < size - 1; other++) {
1455 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1456 if(index == MPI_UNDEFINED) {
1460 if(recvbuf_is_empty){
1461 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1464 // #Request is below rank: it's a irecv
1465 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1469 //non commutative case, wait in order
1470 for(other = 0; other < size - 1; other++) {
1471 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1473 if(recvbuf_is_empty){
1474 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1476 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1480 for(index = 0; index < rank; index++) {
1481 xbt_free(tmpbufs[index]);
1483 for(index = 0; index < size-1; index++) {
1484 smpi_mpi_request_free(&requests[index]);