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 = request->src;
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;
376 XBT_DEBUG("Send request %p is detached", request);
378 if(request->old_type->has_subtype == 0){
379 oldbuf = request->buf;
380 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
381 if((smpi_privatize_global_variables)
382 && ((char*)request->buf >= start_data_exe)
383 && ((char*)request->buf < start_data_exe + size_data_exe )){
384 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
385 switch_data_segment(request->src);
387 buf = xbt_malloc(request->size);
388 memcpy(buf,oldbuf,request->size);
389 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
394 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
395 request->real_size=request->size;
396 smpi_datatype_use(request->old_type);
397 smpi_comm_use(request->comm);
399 //if we are giving back the control to the user without waiting for completion, we have to inject timings
400 double sleeptime = 0.0;
401 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
402 //isend and send timings may be different
403 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
406 if(sleeptime != 0.0){
407 simcall_process_sleep(sleeptime);
408 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
411 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), 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());
638 if(datatype->has_subtype == 1){
639 // This part handles the problem of non-contignous memory
640 // the unserialization at the reception
641 s_smpi_subtype_t *subtype = datatype->substruct;
642 if(req->flags & RECV)
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 if(req->flags & RECV){//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);
650 smpi_comm_unuse(req->comm);
651 smpi_datatype_unuse(datatype);
656 if (TRACE_smpi_view_internals()) {
657 if(req->flags & RECV){
658 int rank = smpi_process_index();
659 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
660 TRACE_smpi_recv(rank, src_traced, rank);
665 if(req->detached_sender!=NULL){
666 smpi_mpi_request_free(&(req->detached_sender));
668 if(req->flags & PERSISTENT)
670 req->flags |= FINISHED;
672 smpi_mpi_request_free(request);
676 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
679 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
680 smpi_empty_status(status);
682 if (!((*request)->flags & PREPARED)) {
683 if ((*request)->action != NULL)
684 flag = simcall_comm_test((*request)->action);
686 finish_wait(request, status);
687 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
688 *request = MPI_REQUEST_NULL;
694 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
701 *index = MPI_UNDEFINED;
703 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
704 map = xbt_new(int, count);
706 for(i = 0; i < count; i++) {
707 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
708 !(requests[i]->flags & PREPARED)) {
709 xbt_dynar_push(comms, &requests[i]->action);
715 i = simcall_comm_testany(comms);
716 // not MPI_UNDEFINED, as this is a simix return code
719 finish_wait(&requests[*index], status);
720 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
721 requests[*index] = MPI_REQUEST_NULL;
725 //all requests are null or inactive, return true
727 smpi_empty_status(status);
730 xbt_dynar_free(&comms);
736 int smpi_mpi_testall(int count, MPI_Request requests[],
740 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
743 for(i=0; i<count; i++){
744 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
745 if (smpi_mpi_test(&requests[i], pstat)!=1){
748 requests[i]=MPI_REQUEST_NULL;
751 smpi_empty_status(pstat);
753 if(status != MPI_STATUSES_IGNORE) {
760 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
762 //FIXME find another wait to avoid busy waiting ?
763 // the issue here is that we have to wait on a nonexistent comm
765 smpi_mpi_iprobe(source, tag, comm, &flag, status);
766 XBT_DEBUG("Busy Waiting on probing : %d", flag);
770 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
772 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,
773 comm, PERSISTENT | RECV);
775 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
776 double sleeptime= sg_cfg_get_double("smpi/iprobe");
777 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
778 static int nsleeps = 1;
780 simcall_process_sleep(sleeptime);
782 // behave like a receive, but don't do it
785 print_request("New iprobe", request);
786 // We have to test both mailboxes as we don't know if we will receive one one or another
787 if (sg_cfg_get_int("smpi/async_small_thres")>0){
788 mailbox = smpi_process_mailbox_small();
789 XBT_DEBUG("trying to probe the perm recv mailbox");
790 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
792 if (request->action==NULL){
793 mailbox = smpi_process_mailbox();
794 XBT_DEBUG("trying to probe the other mailbox");
795 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
799 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
801 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
802 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
803 status->MPI_TAG = req->tag;
804 status->MPI_ERROR = MPI_SUCCESS;
805 status->count = req->real_size;
807 nsleeps=1;//reset the number of sleeps we will do next time
813 smpi_mpi_request_free(&request);
818 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
820 print_request("Waiting", *request);
821 if ((*request)->flags & PREPARED) {
822 smpi_empty_status(status);
826 if ((*request)->action != NULL) { // this is not a detached send
827 simcall_comm_wait((*request)->action, -1.0);
829 if(MC_is_active() && (*request)->action)
830 (*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
834 finish_wait(request, status);
835 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
836 *request = MPI_REQUEST_NULL;
837 // FIXME for a detached send, finish_wait is not called:
840 int smpi_mpi_waitany(int count, MPI_Request requests[],
847 index = MPI_UNDEFINED;
849 // Wait for a request to complete
850 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
851 map = xbt_new(int, count);
853 XBT_DEBUG("Wait for one of %d", count);
854 for(i = 0; i < count; i++) {
855 if (requests[i] != MPI_REQUEST_NULL
856 && !(requests[i]->flags & PREPARED)
857 && !(requests[i]->flags & FINISHED)) {
858 if (requests[i]->action != NULL) {
859 XBT_DEBUG("Waiting any %p ", requests[i]);
860 xbt_dynar_push(comms, &requests[i]->action);
864 //This is a finished detached request, let's return this one
865 size=0;//so we free the dynar but don't do the waitany call
867 finish_wait(&requests[i], status);//cleanup if refcount = 0
868 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
869 requests[i]=MPI_REQUEST_NULL;//set to null
875 i = simcall_comm_waitany(comms);
877 // not MPI_UNDEFINED, as this is a simix return code
880 finish_wait(&requests[index], status);
881 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
882 requests[index] = MPI_REQUEST_NULL;
886 xbt_dynar_free(&comms);
889 if (index==MPI_UNDEFINED)
890 smpi_empty_status(status);
895 int smpi_mpi_waitall(int count, MPI_Request requests[],
900 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
901 int retvalue = MPI_SUCCESS;
902 //tag invalid requests in the set
903 if (status != MPI_STATUSES_IGNORE) {
904 for (c = 0; c < count; c++) {
905 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
906 (requests[c]->flags & PREPARED)) {
907 smpi_empty_status(&status[c]);
908 } else if (requests[c]->src == MPI_PROC_NULL) {
909 smpi_empty_status(&status[c]);
910 status[c].MPI_SOURCE = MPI_PROC_NULL;
914 for(c = 0; c < count; c++) {
916 if (MC_is_active()) {
917 smpi_mpi_wait(&requests[c], pstat);
920 index = smpi_mpi_waitany(count, requests, pstat);
921 if (index == MPI_UNDEFINED)
923 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
924 requests[index]=MPI_REQUEST_NULL;
926 if (status != MPI_STATUSES_IGNORE) {
927 status[index] = *pstat;
928 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
929 retvalue = MPI_ERR_IN_STATUS;
936 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
941 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
944 for(i = 0; i < incount; i++)
946 index=smpi_mpi_waitany(incount, requests, pstat);
947 if(index!=MPI_UNDEFINED){
948 indices[count] = index;
950 if(status != MPI_STATUSES_IGNORE) {
951 status[index] = *pstat;
953 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
954 requests[index]=MPI_REQUEST_NULL;
956 return MPI_UNDEFINED;
962 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
965 int i, count, count_dead;
967 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
971 for(i = 0; i < incount; i++) {
972 if((requests[i] != MPI_REQUEST_NULL)) {
973 if(smpi_mpi_test(&requests[i], pstat)) {
976 if(status != MPI_STATUSES_IGNORE) {
979 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
980 requests[i]=MPI_REQUEST_NULL;
986 if(count_dead==incount)return MPI_UNDEFINED;
990 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
993 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
994 nary_tree_bcast(buf, count, datatype, root, comm, 4);
997 void smpi_mpi_barrier(MPI_Comm comm)
999 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1000 nary_tree_barrier(comm, 4);
1003 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1004 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1005 int root, MPI_Comm comm)
1007 int system_tag = COLL_TAG_GATHER;
1008 int rank, size, src, index;
1009 MPI_Aint lb = 0, recvext = 0;
1010 MPI_Request *requests;
1012 rank = smpi_comm_rank(comm);
1013 size = smpi_comm_size(comm);
1015 // Send buffer to root
1016 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1018 // FIXME: check for errors
1019 smpi_datatype_extent(recvtype, &lb, &recvext);
1020 // Local copy from root
1021 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1022 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1023 // Receive buffers from senders
1024 requests = xbt_new(MPI_Request, size - 1);
1026 for(src = 0; src < size; src++) {
1028 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1029 recvcount, recvtype,
1030 src, system_tag, comm);
1034 // Wait for completion of irecv's.
1035 smpi_mpi_startall(size - 1, requests);
1036 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1037 for(src = 0; src < size-1; src++) {
1038 smpi_mpi_request_free(&requests[src]);
1045 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1046 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1050 int rank = smpi_process_index();
1053 /* arbitrarily choose root as rank 0 */
1054 size = smpi_comm_size(comm);
1056 displs = xbt_new(int, size);
1057 for (i = 0; i < size; i++) {
1059 count += recvcounts[i];
1061 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1062 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1063 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1064 recvcounts[rank], datatype, 0, comm);
1069 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1070 void *recvbuf, int *recvcounts, int *displs,
1071 MPI_Datatype recvtype, int root, MPI_Comm comm)
1073 int system_tag = COLL_TAG_GATHERV;
1074 int rank, size, src, index;
1075 MPI_Aint lb = 0, recvext = 0;
1076 MPI_Request *requests;
1078 rank = smpi_comm_rank(comm);
1079 size = smpi_comm_size(comm);
1081 // Send buffer to root
1082 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1084 // FIXME: check for errors
1085 smpi_datatype_extent(recvtype, &lb, &recvext);
1086 // Local copy from root
1087 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1088 (char *)recvbuf + displs[root] * recvext,
1089 recvcounts[root], recvtype);
1090 // Receive buffers from senders
1091 requests = xbt_new(MPI_Request, size - 1);
1093 for(src = 0; src < size; src++) {
1096 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1097 recvcounts[src], recvtype, src, system_tag, comm);
1101 // Wait for completion of irecv's.
1102 smpi_mpi_startall(size - 1, requests);
1103 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1104 for(src = 0; src < size-1; src++) {
1105 smpi_mpi_request_free(&requests[src]);
1111 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1112 MPI_Datatype sendtype, void *recvbuf,
1113 int recvcount, MPI_Datatype recvtype,
1116 int system_tag = COLL_TAG_ALLGATHER;
1117 int rank, size, other, index;
1118 MPI_Aint lb = 0, recvext = 0;
1119 MPI_Request *requests;
1121 rank = smpi_comm_rank(comm);
1122 size = smpi_comm_size(comm);
1123 // FIXME: check for errors
1124 smpi_datatype_extent(recvtype, &lb, &recvext);
1125 // Local copy from self
1126 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1127 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1129 // Send/Recv buffers to/from others;
1130 requests = xbt_new(MPI_Request, 2 * (size - 1));
1132 for(other = 0; other < size; other++) {
1135 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1138 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1139 recvcount, recvtype, other,
1144 // Wait for completion of all comms.
1145 smpi_mpi_startall(2 * (size - 1), requests);
1146 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1147 for(other = 0; other < 2*(size-1); other++) {
1148 smpi_mpi_request_free(&requests[other]);
1153 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1154 MPI_Datatype sendtype, void *recvbuf,
1155 int *recvcounts, int *displs,
1156 MPI_Datatype recvtype, MPI_Comm comm)
1158 int system_tag = COLL_TAG_ALLGATHERV;
1159 int rank, size, other, index;
1160 MPI_Aint lb = 0, recvext = 0;
1161 MPI_Request *requests;
1163 rank = smpi_comm_rank(comm);
1164 size = smpi_comm_size(comm);
1165 // FIXME: check for errors
1166 smpi_datatype_extent(recvtype, &lb, &recvext);
1167 // Local copy from self
1168 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1169 (char *)recvbuf + displs[rank] * recvext,
1170 recvcounts[rank], recvtype);
1171 // Send buffers to others;
1172 requests = xbt_new(MPI_Request, 2 * (size - 1));
1174 for(other = 0; other < size; other++) {
1177 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1181 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1182 recvtype, other, system_tag, comm);
1186 // Wait for completion of all comms.
1187 smpi_mpi_startall(2 * (size - 1), requests);
1188 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1189 for(other = 0; other < 2*(size-1); other++) {
1190 smpi_mpi_request_free(&requests[other]);
1195 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1196 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1197 int root, MPI_Comm comm)
1199 int system_tag = COLL_TAG_SCATTER;
1200 int rank, size, dst, index;
1201 MPI_Aint lb = 0, sendext = 0;
1202 MPI_Request *requests;
1204 rank = smpi_comm_rank(comm);
1205 size = smpi_comm_size(comm);
1207 // Recv buffer from root
1208 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1211 // FIXME: check for errors
1212 smpi_datatype_extent(sendtype, &lb, &sendext);
1213 // Local copy from root
1214 if(recvbuf!=MPI_IN_PLACE){
1215 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1216 sendcount, sendtype, recvbuf, recvcount, recvtype);
1218 // Send buffers to receivers
1219 requests = xbt_new(MPI_Request, size - 1);
1221 for(dst = 0; dst < size; dst++) {
1223 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1224 sendcount, sendtype, dst,
1229 // Wait for completion of isend's.
1230 smpi_mpi_startall(size - 1, requests);
1231 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1232 for(dst = 0; dst < size-1; dst++) {
1233 smpi_mpi_request_free(&requests[dst]);
1239 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1240 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1241 MPI_Datatype recvtype, int root, MPI_Comm comm)
1243 int system_tag = COLL_TAG_SCATTERV;
1244 int rank, size, dst, index;
1245 MPI_Aint lb = 0, sendext = 0;
1246 MPI_Request *requests;
1248 rank = smpi_comm_rank(comm);
1249 size = smpi_comm_size(comm);
1251 // Recv buffer from root
1252 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1255 // FIXME: check for errors
1256 smpi_datatype_extent(sendtype, &lb, &sendext);
1257 // Local copy from root
1258 if(recvbuf!=MPI_IN_PLACE){
1259 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1260 sendtype, recvbuf, recvcount, recvtype);
1262 // Send buffers to receivers
1263 requests = xbt_new(MPI_Request, size - 1);
1265 for(dst = 0; dst < size; dst++) {
1268 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1269 sendtype, dst, system_tag, comm);
1273 // Wait for completion of isend's.
1274 smpi_mpi_startall(size - 1, requests);
1275 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1276 for(dst = 0; dst < size-1; dst++) {
1277 smpi_mpi_request_free(&requests[dst]);
1283 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1284 MPI_Datatype datatype, MPI_Op op, int root,
1287 int system_tag = COLL_TAG_REDUCE;
1288 int rank, size, src, index;
1289 MPI_Aint lb = 0, dataext = 0;
1290 MPI_Request *requests;
1294 char* sendtmpbuf = (char*) sendbuf;
1295 if( sendbuf == MPI_IN_PLACE ) {
1296 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1297 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1300 rank = smpi_comm_rank(comm);
1301 size = smpi_comm_size(comm);
1302 //non commutative case, use a working algo from openmpi
1303 if(!smpi_op_is_commute(op)){
1304 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1305 datatype, op, root, comm);
1310 // Send buffer to root
1311 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1313 // FIXME: check for errors
1314 smpi_datatype_extent(datatype, &lb, &dataext);
1315 // Local copy from root
1316 if (sendtmpbuf && recvbuf)
1317 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1318 // Receive buffers from senders
1319 //TODO: make a MPI_barrier here ?
1320 requests = xbt_new(MPI_Request, size - 1);
1321 tmpbufs = xbt_new(void *, size - 1);
1323 for(src = 0; src < size; src++) {
1325 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1327 tmpbufs[index] = xbt_malloc(count * dataext);
1329 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1334 // Wait for completion of irecv's.
1335 smpi_mpi_startall(size - 1, requests);
1336 for(src = 0; src < size - 1; src++) {
1337 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1338 XBT_DEBUG("finished waiting any request with index %d", index);
1339 if(index == MPI_UNDEFINED) {
1342 smpi_mpi_request_free(&requests[index]);
1344 if(op) /* op can be MPI_OP_NULL that does nothing */
1345 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1347 for(index = 0; index < size - 1; index++) {
1348 xbt_free(tmpbufs[index]);
1353 if( sendbuf == MPI_IN_PLACE ) {
1354 xbt_free(sendtmpbuf);
1359 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1360 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1362 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1363 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1366 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1367 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1369 int system_tag = -888;
1370 int rank, size, other, index;
1371 MPI_Aint lb = 0, dataext = 0;
1372 MPI_Request *requests;
1375 rank = smpi_comm_rank(comm);
1376 size = smpi_comm_size(comm);
1378 // FIXME: check for errors
1379 smpi_datatype_extent(datatype, &lb, &dataext);
1381 // Local copy from self
1382 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1384 // Send/Recv buffers to/from others;
1385 requests = xbt_new(MPI_Request, size - 1);
1386 tmpbufs = xbt_new(void *, rank);
1388 for(other = 0; other < rank; other++) {
1389 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1391 tmpbufs[index] = xbt_malloc(count * dataext);
1393 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1397 for(other = rank + 1; other < size; other++) {
1399 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1402 // Wait for completion of all comms.
1403 smpi_mpi_startall(size - 1, requests);
1405 if(smpi_op_is_commute(op)){
1406 for(other = 0; other < size - 1; other++) {
1407 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1408 if(index == MPI_UNDEFINED) {
1412 // #Request is below rank: it's a irecv
1413 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1417 //non commutative case, wait in order
1418 for(other = 0; other < size - 1; other++) {
1419 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1421 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1425 for(index = 0; index < rank; index++) {
1426 xbt_free(tmpbufs[index]);
1428 for(index = 0; index < size-1; index++) {
1429 smpi_mpi_request_free(&requests[index]);
1435 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1436 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1438 int system_tag = -888;
1439 int rank, size, other, index;
1440 MPI_Aint lb = 0, dataext = 0;
1441 MPI_Request *requests;
1443 int recvbuf_is_empty=1;
1444 rank = smpi_comm_rank(comm);
1445 size = smpi_comm_size(comm);
1447 // FIXME: check for errors
1448 smpi_datatype_extent(datatype, &lb, &dataext);
1450 // Send/Recv buffers to/from others;
1451 requests = xbt_new(MPI_Request, size - 1);
1452 tmpbufs = xbt_new(void *, rank);
1454 for(other = 0; other < rank; other++) {
1455 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1457 tmpbufs[index] = xbt_malloc(count * dataext);
1459 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1463 for(other = rank + 1; other < size; other++) {
1465 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1468 // Wait for completion of all comms.
1469 smpi_mpi_startall(size - 1, requests);
1470 if(smpi_op_is_commute(op)){
1471 for(other = 0; other < size - 1; other++) {
1472 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1473 if(index == MPI_UNDEFINED) {
1477 if(recvbuf_is_empty){
1478 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1481 // #Request is below rank: it's a irecv
1482 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1486 //non commutative case, wait in order
1487 for(other = 0; other < size - 1; other++) {
1488 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1490 if(recvbuf_is_empty){
1491 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1493 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1497 for(index = 0; index < rank; index++) {
1498 xbt_free(tmpbufs[index]);
1500 for(index = 0; index < size-1; index++) {
1501 smpi_mpi_request_free(&requests[index]);