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(((flags & RECV) && (flags & ACCUMULATE)) || (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));
204 if ((datatype->has_subtype == 1) && (flags & SEND)) {
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;
233 request->op = MPI_REPLACE;
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 &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 &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, MPI_Op op)
463 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
465 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
466 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
468 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
469 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
475 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
476 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
478 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
480 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
481 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
483 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
484 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
491 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
492 int dst, 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, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
496 comm, PERSISTENT | ISEND | SEND | PREPARED);
500 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
501 int dst, 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, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
505 comm, NON_PERSISTENT | ISEND | SEND);
506 smpi_mpi_start(request);
510 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
511 int dst, int tag, MPI_Comm comm)
513 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
514 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
515 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
516 smpi_mpi_start(request);
520 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
521 int src, int tag, MPI_Comm comm)
523 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
524 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,
525 comm, PERSISTENT | RECV | PREPARED);
529 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
530 int src, int tag, MPI_Comm comm)
532 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
533 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,
534 comm, NON_PERSISTENT | RECV);
535 smpi_mpi_start(request);
539 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
540 int tag, MPI_Comm comm, MPI_Status * status)
542 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
543 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
544 smpi_mpi_wait(&request, status);
550 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
551 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 | SEND);
557 smpi_mpi_start(request);
558 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
562 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
563 int dst, int tag, MPI_Comm comm)
565 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
566 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
567 comm, NON_PERSISTENT | SSEND | SEND);
569 smpi_mpi_start(request);
570 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
574 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
575 int dst, int sendtag, void *recvbuf, int recvcount,
576 MPI_Datatype recvtype, int src, int recvtag,
577 MPI_Comm comm, MPI_Status * status)
579 MPI_Request requests[2];
581 int myid=smpi_process_index();
582 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
583 smpi_datatype_copy(sendbuf, sendcount, sendtype,
584 recvbuf, recvcount, recvtype);
588 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
590 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
591 smpi_mpi_startall(2, requests);
592 smpi_mpi_waitall(2, requests, stats);
593 smpi_mpi_request_free(&requests[0]);
594 smpi_mpi_request_free(&requests[1]);
595 if(status != MPI_STATUS_IGNORE) {
596 // Copy receive status
601 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
603 return status->count / smpi_datatype_size(datatype);
606 static void finish_wait(MPI_Request * request, MPI_Status * status)
608 MPI_Request req = *request;
609 smpi_empty_status(status);
611 if(!(req->detached && req->flags & SEND)
612 && !(req->flags & PREPARED)){
613 if(status != MPI_STATUS_IGNORE) {
614 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
615 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
616 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
617 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
618 // this handles the case were size in receive differs from size in send
619 // FIXME: really this should just contain the count of receive-type blocks,
621 status->count = req->real_size;
624 print_request("Finishing", req);
625 MPI_Datatype datatype = req->old_type;
627 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
628 if (!_xbt_replay_is_active()){
629 if( smpi_privatize_global_variables
630 && ((char*)req->old_buf >= start_data_exe)
631 && ((char*)req->old_buf < start_data_exe + size_data_exe )
633 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
634 switch_data_segment(smpi_process_index());
637 if(req->flags & RECV) {
639 if(datatype->has_subtype == 1){
640 // This part handles the problem of non-contignous memory
641 // the unserialization at the reception
642 s_smpi_subtype_t *subtype = datatype->substruct;
643 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
644 if(req->detached == 0) free(req->buf);
645 }else{//apply op on contiguous buffer for accumulate
646 int n =req->real_size/smpi_datatype_size(datatype);
647 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
651 smpi_comm_unuse(req->comm);
652 smpi_datatype_unuse(datatype);
657 if (TRACE_smpi_view_internals()) {
658 if(req->flags & RECV){
659 int rank = smpi_process_index();
660 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
661 TRACE_smpi_recv(rank, src_traced, rank);
666 if(req->detached_sender!=NULL){
667 smpi_mpi_request_free(&(req->detached_sender));
669 if(req->flags & PERSISTENT)
671 req->flags |= FINISHED;
673 smpi_mpi_request_free(request);
677 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
680 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
681 smpi_empty_status(status);
683 if (!((*request)->flags & PREPARED)) {
684 if ((*request)->action != NULL)
685 flag = simcall_comm_test((*request)->action);
687 finish_wait(request, status);
688 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
689 *request = MPI_REQUEST_NULL;
695 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
702 *index = MPI_UNDEFINED;
704 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
705 map = xbt_new(int, count);
707 for(i = 0; i < count; i++) {
708 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
709 !(requests[i]->flags & PREPARED)) {
710 xbt_dynar_push(comms, &requests[i]->action);
716 i = simcall_comm_testany(comms);
717 // not MPI_UNDEFINED, as this is a simix return code
720 finish_wait(&requests[*index], status);
721 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
722 requests[*index] = MPI_REQUEST_NULL;
726 //all requests are null or inactive, return true
728 smpi_empty_status(status);
731 xbt_dynar_free(&comms);
737 int smpi_mpi_testall(int count, MPI_Request requests[],
741 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
744 for(i=0; i<count; i++){
745 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
746 if (smpi_mpi_test(&requests[i], pstat)!=1){
749 requests[i]=MPI_REQUEST_NULL;
752 smpi_empty_status(pstat);
754 if(status != MPI_STATUSES_IGNORE) {
761 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
763 //FIXME find another wait to avoid busy waiting ?
764 // the issue here is that we have to wait on a nonexistent comm
766 smpi_mpi_iprobe(source, tag, comm, &flag, status);
767 XBT_DEBUG("Busy Waiting on probing : %d", flag);
771 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
773 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,
774 comm, PERSISTENT | RECV);
776 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
777 double sleeptime= sg_cfg_get_double("smpi/iprobe");
778 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
779 static int nsleeps = 1;
781 simcall_process_sleep(sleeptime);
783 // behave like a receive, but don't do it
786 print_request("New iprobe", request);
787 // We have to test both mailboxes as we don't know if we will receive one one or another
788 if (sg_cfg_get_int("smpi/async_small_thres")>0){
789 mailbox = smpi_process_mailbox_small();
790 XBT_DEBUG("trying to probe the perm recv mailbox");
791 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
793 if (request->action==NULL){
794 mailbox = smpi_process_mailbox();
795 XBT_DEBUG("trying to probe the other mailbox");
796 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
800 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
802 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
803 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
804 status->MPI_TAG = req->tag;
805 status->MPI_ERROR = MPI_SUCCESS;
806 status->count = req->real_size;
808 nsleeps=1;//reset the number of sleeps we will do next time
814 smpi_mpi_request_free(&request);
819 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
821 print_request("Waiting", *request);
822 if ((*request)->flags & PREPARED) {
823 smpi_empty_status(status);
827 if ((*request)->action != NULL) { // this is not a detached send
828 simcall_comm_wait((*request)->action, -1.0);
830 if(MC_is_active() && (*request)->action)
831 (*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
835 finish_wait(request, status);
836 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
837 *request = MPI_REQUEST_NULL;
838 // FIXME for a detached send, finish_wait is not called:
841 int smpi_mpi_waitany(int count, MPI_Request requests[],
848 index = MPI_UNDEFINED;
850 // Wait for a request to complete
851 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
852 map = xbt_new(int, count);
854 XBT_DEBUG("Wait for one of %d", count);
855 for(i = 0; i < count; i++) {
856 if (requests[i] != MPI_REQUEST_NULL
857 && !(requests[i]->flags & PREPARED)
858 && !(requests[i]->flags & FINISHED)) {
859 if (requests[i]->action != NULL) {
860 XBT_DEBUG("Waiting any %p ", requests[i]);
861 xbt_dynar_push(comms, &requests[i]->action);
865 //This is a finished detached request, let's return this one
866 size=0;//so we free the dynar but don't do the waitany call
868 finish_wait(&requests[i], status);//cleanup if refcount = 0
869 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
870 requests[i]=MPI_REQUEST_NULL;//set to null
876 i = simcall_comm_waitany(comms);
878 // not MPI_UNDEFINED, as this is a simix return code
881 finish_wait(&requests[index], status);
882 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
883 requests[index] = MPI_REQUEST_NULL;
887 xbt_dynar_free(&comms);
890 if (index==MPI_UNDEFINED)
891 smpi_empty_status(status);
896 int smpi_mpi_waitall(int count, MPI_Request requests[],
901 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
902 int retvalue = MPI_SUCCESS;
903 //tag invalid requests in the set
904 if (status != MPI_STATUSES_IGNORE) {
905 for (c = 0; c < count; c++) {
906 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
907 (requests[c]->flags & PREPARED)) {
908 smpi_empty_status(&status[c]);
909 } else if (requests[c]->src == MPI_PROC_NULL) {
910 smpi_empty_status(&status[c]);
911 status[c].MPI_SOURCE = MPI_PROC_NULL;
915 for(c = 0; c < count; c++) {
917 if (MC_is_active()) {
918 smpi_mpi_wait(&requests[c], pstat);
921 index = smpi_mpi_waitany(count, requests, pstat);
922 if (index == MPI_UNDEFINED)
924 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
925 requests[index]=MPI_REQUEST_NULL;
927 if (status != MPI_STATUSES_IGNORE) {
928 status[index] = *pstat;
929 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
930 retvalue = MPI_ERR_IN_STATUS;
937 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
942 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
945 for(i = 0; i < incount; i++)
947 index=smpi_mpi_waitany(incount, requests, pstat);
948 if(index!=MPI_UNDEFINED){
949 indices[count] = index;
951 if(status != MPI_STATUSES_IGNORE) {
952 status[index] = *pstat;
954 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
955 requests[index]=MPI_REQUEST_NULL;
957 return MPI_UNDEFINED;
963 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
966 int i, count, count_dead;
968 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
972 for(i = 0; i < incount; i++) {
973 if((requests[i] != MPI_REQUEST_NULL)) {
974 if(smpi_mpi_test(&requests[i], pstat)) {
977 if(status != MPI_STATUSES_IGNORE) {
980 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
981 requests[i]=MPI_REQUEST_NULL;
987 if(count_dead==incount)return MPI_UNDEFINED;
991 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
994 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
995 nary_tree_bcast(buf, count, datatype, root, comm, 4);
998 void smpi_mpi_barrier(MPI_Comm comm)
1000 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1001 nary_tree_barrier(comm, 4);
1004 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1005 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1006 int root, MPI_Comm comm)
1008 int system_tag = COLL_TAG_GATHER;
1009 int rank, size, src, index;
1010 MPI_Aint lb = 0, recvext = 0;
1011 MPI_Request *requests;
1013 rank = smpi_comm_rank(comm);
1014 size = smpi_comm_size(comm);
1016 // Send buffer to root
1017 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1019 // FIXME: check for errors
1020 smpi_datatype_extent(recvtype, &lb, &recvext);
1021 // Local copy from root
1022 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1023 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1024 // Receive buffers from senders
1025 requests = xbt_new(MPI_Request, size - 1);
1027 for(src = 0; src < size; src++) {
1029 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1030 recvcount, recvtype,
1031 src, system_tag, comm);
1035 // Wait for completion of irecv's.
1036 smpi_mpi_startall(size - 1, requests);
1037 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1038 for(src = 0; src < size-1; src++) {
1039 smpi_mpi_request_free(&requests[src]);
1046 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1047 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1051 int rank = smpi_process_index();
1054 /* arbitrarily choose root as rank 0 */
1055 size = smpi_comm_size(comm);
1057 displs = xbt_new(int, size);
1058 for (i = 0; i < size; i++) {
1060 count += recvcounts[i];
1062 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1063 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1064 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1065 recvcounts[rank], datatype, 0, comm);
1070 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1071 void *recvbuf, int *recvcounts, int *displs,
1072 MPI_Datatype recvtype, int root, MPI_Comm comm)
1074 int system_tag = COLL_TAG_GATHERV;
1075 int rank, size, src, index;
1076 MPI_Aint lb = 0, recvext = 0;
1077 MPI_Request *requests;
1079 rank = smpi_comm_rank(comm);
1080 size = smpi_comm_size(comm);
1082 // Send buffer to root
1083 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1085 // FIXME: check for errors
1086 smpi_datatype_extent(recvtype, &lb, &recvext);
1087 // Local copy from root
1088 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1089 (char *)recvbuf + displs[root] * recvext,
1090 recvcounts[root], recvtype);
1091 // Receive buffers from senders
1092 requests = xbt_new(MPI_Request, size - 1);
1094 for(src = 0; src < size; src++) {
1097 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1098 recvcounts[src], recvtype, src, system_tag, comm);
1102 // Wait for completion of irecv's.
1103 smpi_mpi_startall(size - 1, requests);
1104 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1105 for(src = 0; src < size-1; src++) {
1106 smpi_mpi_request_free(&requests[src]);
1112 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1113 MPI_Datatype sendtype, void *recvbuf,
1114 int recvcount, MPI_Datatype recvtype,
1117 int system_tag = COLL_TAG_ALLGATHER;
1118 int rank, size, other, index;
1119 MPI_Aint lb = 0, recvext = 0;
1120 MPI_Request *requests;
1122 rank = smpi_comm_rank(comm);
1123 size = smpi_comm_size(comm);
1124 // FIXME: check for errors
1125 smpi_datatype_extent(recvtype, &lb, &recvext);
1126 // Local copy from self
1127 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1128 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1130 // Send/Recv buffers to/from others;
1131 requests = xbt_new(MPI_Request, 2 * (size - 1));
1133 for(other = 0; other < size; other++) {
1136 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1139 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1140 recvcount, recvtype, other,
1145 // Wait for completion of all comms.
1146 smpi_mpi_startall(2 * (size - 1), requests);
1147 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1148 for(other = 0; other < 2*(size-1); other++) {
1149 smpi_mpi_request_free(&requests[other]);
1154 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1155 MPI_Datatype sendtype, void *recvbuf,
1156 int *recvcounts, int *displs,
1157 MPI_Datatype recvtype, MPI_Comm comm)
1159 int system_tag = COLL_TAG_ALLGATHERV;
1160 int rank, size, other, index;
1161 MPI_Aint lb = 0, recvext = 0;
1162 MPI_Request *requests;
1164 rank = smpi_comm_rank(comm);
1165 size = smpi_comm_size(comm);
1166 // FIXME: check for errors
1167 smpi_datatype_extent(recvtype, &lb, &recvext);
1168 // Local copy from self
1169 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1170 (char *)recvbuf + displs[rank] * recvext,
1171 recvcounts[rank], recvtype);
1172 // Send buffers to others;
1173 requests = xbt_new(MPI_Request, 2 * (size - 1));
1175 for(other = 0; other < size; other++) {
1178 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1182 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1183 recvtype, other, system_tag, comm);
1187 // Wait for completion of all comms.
1188 smpi_mpi_startall(2 * (size - 1), requests);
1189 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1190 for(other = 0; other < 2*(size-1); other++) {
1191 smpi_mpi_request_free(&requests[other]);
1196 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1197 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1198 int root, MPI_Comm comm)
1200 int system_tag = COLL_TAG_SCATTER;
1201 int rank, size, dst, index;
1202 MPI_Aint lb = 0, sendext = 0;
1203 MPI_Request *requests;
1205 rank = smpi_comm_rank(comm);
1206 size = smpi_comm_size(comm);
1208 // Recv buffer from root
1209 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1212 // FIXME: check for errors
1213 smpi_datatype_extent(sendtype, &lb, &sendext);
1214 // Local copy from root
1215 if(recvbuf!=MPI_IN_PLACE){
1216 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1217 sendcount, sendtype, recvbuf, recvcount, recvtype);
1219 // Send buffers to receivers
1220 requests = xbt_new(MPI_Request, size - 1);
1222 for(dst = 0; dst < size; dst++) {
1224 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1225 sendcount, sendtype, dst,
1230 // Wait for completion of isend's.
1231 smpi_mpi_startall(size - 1, requests);
1232 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1233 for(dst = 0; dst < size-1; dst++) {
1234 smpi_mpi_request_free(&requests[dst]);
1240 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1241 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1242 MPI_Datatype recvtype, int root, MPI_Comm comm)
1244 int system_tag = COLL_TAG_SCATTERV;
1245 int rank, size, dst, index;
1246 MPI_Aint lb = 0, sendext = 0;
1247 MPI_Request *requests;
1249 rank = smpi_comm_rank(comm);
1250 size = smpi_comm_size(comm);
1252 // Recv buffer from root
1253 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1256 // FIXME: check for errors
1257 smpi_datatype_extent(sendtype, &lb, &sendext);
1258 // Local copy from root
1259 if(recvbuf!=MPI_IN_PLACE){
1260 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1261 sendtype, recvbuf, recvcount, recvtype);
1263 // Send buffers to receivers
1264 requests = xbt_new(MPI_Request, size - 1);
1266 for(dst = 0; dst < size; dst++) {
1269 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1270 sendtype, dst, system_tag, comm);
1274 // Wait for completion of isend's.
1275 smpi_mpi_startall(size - 1, requests);
1276 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1277 for(dst = 0; dst < size-1; dst++) {
1278 smpi_mpi_request_free(&requests[dst]);
1284 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1285 MPI_Datatype datatype, MPI_Op op, int root,
1288 int system_tag = COLL_TAG_REDUCE;
1289 int rank, size, src, index;
1290 MPI_Aint lb = 0, dataext = 0;
1291 MPI_Request *requests;
1295 char* sendtmpbuf = (char*) sendbuf;
1296 if( sendbuf == MPI_IN_PLACE ) {
1297 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1298 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1301 rank = smpi_comm_rank(comm);
1302 size = smpi_comm_size(comm);
1303 //non commutative case, use a working algo from openmpi
1304 if(!smpi_op_is_commute(op)){
1305 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1306 datatype, op, root, comm);
1311 // Send buffer to root
1312 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1314 // FIXME: check for errors
1315 smpi_datatype_extent(datatype, &lb, &dataext);
1316 // Local copy from root
1317 if (sendtmpbuf && recvbuf)
1318 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1319 // Receive buffers from senders
1320 //TODO: make a MPI_barrier here ?
1321 requests = xbt_new(MPI_Request, size - 1);
1322 tmpbufs = xbt_new(void *, size - 1);
1324 for(src = 0; src < size; src++) {
1326 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1328 tmpbufs[index] = xbt_malloc(count * dataext);
1330 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1335 // Wait for completion of irecv's.
1336 smpi_mpi_startall(size - 1, requests);
1337 for(src = 0; src < size - 1; src++) {
1338 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1339 XBT_DEBUG("finished waiting any request with index %d", index);
1340 if(index == MPI_UNDEFINED) {
1343 smpi_mpi_request_free(&requests[index]);
1345 if(op) /* op can be MPI_OP_NULL that does nothing */
1346 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1348 for(index = 0; index < size - 1; index++) {
1349 xbt_free(tmpbufs[index]);
1354 if( sendbuf == MPI_IN_PLACE ) {
1355 xbt_free(sendtmpbuf);
1360 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1361 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1363 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1364 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1367 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1368 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1370 int system_tag = -888;
1371 int rank, size, other, index;
1372 MPI_Aint lb = 0, dataext = 0;
1373 MPI_Request *requests;
1376 rank = smpi_comm_rank(comm);
1377 size = smpi_comm_size(comm);
1379 // FIXME: check for errors
1380 smpi_datatype_extent(datatype, &lb, &dataext);
1382 // Local copy from self
1383 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1385 // Send/Recv buffers to/from others;
1386 requests = xbt_new(MPI_Request, size - 1);
1387 tmpbufs = xbt_new(void *, rank);
1389 for(other = 0; other < rank; other++) {
1390 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1392 tmpbufs[index] = xbt_malloc(count * dataext);
1394 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1398 for(other = rank + 1; other < size; other++) {
1400 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1403 // Wait for completion of all comms.
1404 smpi_mpi_startall(size - 1, requests);
1406 if(smpi_op_is_commute(op)){
1407 for(other = 0; other < size - 1; other++) {
1408 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1409 if(index == MPI_UNDEFINED) {
1413 // #Request is below rank: it's a irecv
1414 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1418 //non commutative case, wait in order
1419 for(other = 0; other < size - 1; other++) {
1420 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1422 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1426 for(index = 0; index < rank; index++) {
1427 xbt_free(tmpbufs[index]);
1429 for(index = 0; index < size-1; index++) {
1430 smpi_mpi_request_free(&requests[index]);
1436 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1437 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1439 int system_tag = -888;
1440 int rank, size, other, index;
1441 MPI_Aint lb = 0, dataext = 0;
1442 MPI_Request *requests;
1444 int recvbuf_is_empty=1;
1445 rank = smpi_comm_rank(comm);
1446 size = smpi_comm_size(comm);
1448 // FIXME: check for errors
1449 smpi_datatype_extent(datatype, &lb, &dataext);
1451 // Send/Recv buffers to/from others;
1452 requests = xbt_new(MPI_Request, size - 1);
1453 tmpbufs = xbt_new(void *, rank);
1455 for(other = 0; other < rank; other++) {
1456 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1458 tmpbufs[index] = xbt_malloc(count * dataext);
1460 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1464 for(other = rank + 1; other < size; other++) {
1466 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1469 // Wait for completion of all comms.
1470 smpi_mpi_startall(size - 1, requests);
1471 if(smpi_op_is_commute(op)){
1472 for(other = 0; other < size - 1; other++) {
1473 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1474 if(index == MPI_UNDEFINED) {
1478 if(recvbuf_is_empty){
1479 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1482 // #Request is below rank: it's a irecv
1483 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1487 //non commutative case, wait in order
1488 for(other = 0; other < size - 1; other++) {
1489 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1491 if(recvbuf_is_empty){
1492 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1494 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1498 for(index = 0; index < rank; index++) {
1499 xbt_free(tmpbufs[index]);
1501 for(index = 0; index < size-1; index++) {
1502 smpi_mpi_request_free(&requests[index]);