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 "mc/mc_replay.h"
11 #include "xbt/replay.h"
13 #include "simix/smx_private.h"
14 #include "surf/surf.h"
15 #include "simgrid/sg_config.h"
16 #include "colls/colls.h"
18 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
21 static int match_recv(void* a, void* b, smx_synchro_t ignored) {
22 MPI_Request ref = (MPI_Request)a;
23 MPI_Request req = (MPI_Request)b;
24 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
26 xbt_assert(ref, "Cannot match recv against null reference");
27 xbt_assert(req, "Cannot match recv against null request");
28 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
29 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
30 //we match, we can transfer some values
31 // FIXME : move this to the copy function ?
32 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
33 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
34 if(ref->real_size < req->real_size) ref->truncated = 1;
36 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
38 XBT_DEBUG("match succeeded");
43 static int match_send(void* a, void* b,smx_synchro_t ignored) {
44 MPI_Request ref = (MPI_Request)a;
45 MPI_Request req = (MPI_Request)b;
46 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
47 xbt_assert(ref, "Cannot match send against null reference");
48 xbt_assert(req, "Cannot match send against null request");
50 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
51 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
53 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
54 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
55 if(req->real_size < ref->real_size) req->truncated = 1;
57 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
59 XBT_DEBUG("match succeeded");
65 typedef struct s_smpi_factor *smpi_factor_t;
66 typedef struct s_smpi_factor {
69 double values[4];//arbitrary set to 4
71 xbt_dynar_t smpi_os_values = NULL;
72 xbt_dynar_t smpi_or_values = NULL;
73 xbt_dynar_t smpi_ois_values = NULL;
75 double smpi_wtime_sleep = 0.0;
76 double smpi_iprobe_sleep = 1e-4;
77 double smpi_test_sleep = 1e-4;
80 // Methods used to parse and store the values for timing injections in smpi
81 // These are taken from surf/network.c and generalized to have more factors
82 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
84 static int factor_cmp(const void *pa, const void *pb)
86 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
87 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
91 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
94 unsigned int iter = 0;
98 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
100 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
101 radical_elements = xbt_str_split(smpi_coef_string, ";");
102 xbt_dynar_foreach(radical_elements, iter, value) {
103 memset(&fact, 0, sizeof(s_smpi_factor_t));
104 radical_elements2 = xbt_str_split(value, ":");
105 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
106 xbt_die("Malformed radical for smpi factor!");
107 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
109 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
111 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
116 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
117 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
118 xbt_dynar_free(&radical_elements2);
120 xbt_dynar_free(&radical_elements);
122 xbt_dynar_sort(smpi_factor, &factor_cmp);
123 xbt_dynar_foreach(smpi_factor, iter, fact) {
124 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
129 static double smpi_os(double size)
131 if (!smpi_os_values) {
132 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
133 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
135 unsigned int iter = 0;
136 s_smpi_factor_t fact;
138 // Iterate over all the sections that were specified and find the right
139 // value. (fact.factor represents the interval sizes; we want to find the
140 // section that has fact.factor <= size and no other such fact.factor <= size)
141 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
142 xbt_dynar_foreach(smpi_os_values, iter, fact) {
143 if (size <= fact.factor) { // Values already too large, use the previously
144 // computed value of current!
145 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
148 // If the next section is too large, the current section must be used.
149 // Hence, save the cost, as we might have to use it.
150 current = fact.values[0]+fact.values[1]*size;
153 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
158 static double smpi_ois(double size)
160 if (!smpi_ois_values) {
161 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
162 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
164 unsigned int iter = 0;
165 s_smpi_factor_t fact;
167 // Iterate over all the sections that were specified and find the right
168 // value. (fact.factor represents the interval sizes; we want to find the
169 // section that has fact.factor <= size and no other such fact.factor <= size)
170 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
171 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
172 if (size <= fact.factor) { // Values already too large, use the previously
173 // computed value of current!
174 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
177 // If the next section is too large, the current section must be used.
178 // Hence, save the cost, as we might have to use it.
179 current = fact.values[0]+fact.values[1]*size;
182 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
187 static double smpi_or(double size)
189 if (!smpi_or_values) {
190 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
191 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
193 unsigned int iter = 0;
194 s_smpi_factor_t fact;
196 // Iterate over all the sections that were specified and find the right
197 // value. (fact.factor represents the interval sizes; we want to find the
198 // section that has fact.factor <= size and no other such fact.factor <= size)
199 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
200 xbt_dynar_foreach(smpi_or_values, iter, fact) {
201 if (size <= fact.factor) { // Values already too large, use the previously
202 // computed value of current!
203 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
206 // If the next section is too large, the current section must be used.
207 // Hence, save the cost, as we might have to use it.
208 current=fact.values[0]+fact.values[1]*size;
211 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
216 double smpi_mpi_wtime(){
218 if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
220 time = SIMIX_get_clock();
221 // to avoid deadlocks if used as a break condition, such as
222 // while (MPI_Wtime(...) < time_limit) {
225 // because the time will not normally advance when only calls to MPI_Wtime
226 // are made -> deadlock (MPI_Wtime never reaches the time limit)
227 if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep);
230 time = SIMIX_get_clock();
235 static MPI_Request build_request(void *buf, int count,
236 MPI_Datatype datatype, int src, int dst,
237 int tag, MPI_Comm comm, unsigned flags)
239 MPI_Request request = NULL;
241 void *old_buf = NULL;
243 request = xbt_new(s_smpi_mpi_request_t, 1);
245 s_smpi_subtype_t *subtype = datatype->substruct;
247 if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
248 // This part handles the problem of non-contiguous memory
250 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
251 if ((datatype->has_subtype == 1) && (flags & SEND)) {
252 subtype->serialize(old_buf, buf, count, datatype->substruct);
257 // This part handles the problem of non-contiguous memory (for the
258 // unserialisation at the reception)
259 request->old_buf = old_buf;
260 request->old_type = datatype;
262 request->size = smpi_datatype_size(datatype) * count;
266 request->comm = comm;
267 request->action = NULL;
268 request->flags = flags;
269 request->detached = 0;
270 request->detached_sender = NULL;
271 request->real_src = 0;
273 request->truncated = 0;
274 request->real_size = 0;
275 request->real_tag = 0;
276 if(flags & PERSISTENT)
277 request->refcount = 1;
279 request->refcount = 0;
280 request->op = MPI_REPLACE;
283 if (flags & SEND) smpi_datatype_unuse(datatype);
289 void smpi_empty_status(MPI_Status * status)
291 if(status != MPI_STATUS_IGNORE) {
292 status->MPI_SOURCE = MPI_ANY_SOURCE;
293 status->MPI_TAG = MPI_ANY_TAG;
294 status->MPI_ERROR = MPI_SUCCESS;
299 void smpi_action_trace_run(char *path)
303 xbt_dict_cursor_t cursor;
307 action_fp = fopen(path, "r");
308 if (action_fp == NULL)
309 xbt_die("Cannot open %s: %s", path, strerror(errno));
312 if (!xbt_dict_is_empty(action_queues)) {
314 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
317 xbt_dict_foreach(action_queues, cursor, name, todo) {
318 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
324 xbt_dict_free(&action_queues);
325 action_queues = xbt_dict_new_homogeneous(NULL);
328 static void smpi_mpi_request_free_voidp(void* request)
330 MPI_Request req = request;
331 smpi_mpi_request_free(&req);
334 /* MPI Low level calls */
335 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
336 int dst, int tag, MPI_Comm comm)
338 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
339 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
340 comm, PERSISTENT | SEND | PREPARED);
344 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
345 int dst, int tag, MPI_Comm comm)
347 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
348 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
349 comm, PERSISTENT | SSEND | SEND | PREPARED);
353 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
354 int src, int tag, MPI_Comm comm)
356 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
357 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,
358 comm, PERSISTENT | RECV | PREPARED);
362 void smpi_mpi_start(MPI_Request request)
366 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
367 request->flags &= ~PREPARED;
368 request->flags &= ~FINISHED;
371 if (request->flags & RECV) {
372 print_request("New recv", request);
374 xbt_mutex_t mut=smpi_process_mailboxes_mutex();
375 xbt_mutex_acquire(mut);
377 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
378 //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
379 mailbox = smpi_process_mailbox_small();
380 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
381 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
384 mailbox = smpi_process_mailbox();
385 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
386 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
388 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
389 mailbox = smpi_process_mailbox_small();
392 XBT_DEBUG("yes there was something for us in the large mailbox");
395 mailbox = smpi_process_mailbox_small();
396 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
397 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
400 XBT_DEBUG("No, nothing in the permanent receive mailbox");
401 mailbox = smpi_process_mailbox();
403 XBT_DEBUG("yes there was something for us in the small mailbox");
407 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
408 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
410 simcall_process_sleep(sleeptime);
411 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
414 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
415 request->real_size=request->size;
416 smpi_datatype_use(request->old_type);
417 smpi_comm_use(request->comm);
418 request->action = simcall_comm_irecv(mailbox, request->buf,
419 &request->real_size, &match_recv,
420 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
421 : &smpi_comm_null_copy_buffer_callback,
423 XBT_DEBUG("recv simcall posted");
425 xbt_mutex_release(mut);
429 int receiver = request->dst;
431 int rank = request->src;
432 if (TRACE_smpi_view_internals()) {
433 TRACE_smpi_send(rank, rank, receiver,request->size);
435 print_request("New send", request);
437 //if we are giving back the control to the user without waiting for completion, we have to inject timings
438 double sleeptime = 0.0;
439 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
440 //isend and send timings may be different
441 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
444 if(sleeptime != 0.0){
445 simcall_process_sleep(sleeptime);
446 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
449 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
450 xbt_mutex_acquire(mut);
452 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
453 mailbox = smpi_process_remote_mailbox(receiver);
454 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
455 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
457 if (! (request->flags & SSEND)){
458 mailbox = smpi_process_remote_mailbox_small(receiver);
459 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
461 mailbox = smpi_process_remote_mailbox_small(receiver);
462 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
463 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
465 XBT_DEBUG("No, we are first, send to large mailbox");
466 mailbox = smpi_process_remote_mailbox(receiver);
470 XBT_DEBUG("Yes there was something for us in the large mailbox");
473 mailbox = smpi_process_remote_mailbox(receiver);
474 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
477 void* buf = request->buf;
478 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
480 request->detached = 1;
481 XBT_DEBUG("Send request %p is detached", request);
483 if(request->old_type->has_subtype == 0){
484 oldbuf = request->buf;
485 if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
486 if((smpi_privatize_global_variables)
487 && ((char*) request->buf >= smpi_start_data_exe)
488 && ((char*)request->buf < smpi_start_data_exe + smpi_size_data_exe )){
489 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
490 smpi_switch_data_segment(request->src);
492 buf = xbt_malloc(request->size);
493 memcpy(buf,oldbuf,request->size);
494 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
499 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
500 request->real_size=request->size;
501 smpi_datatype_use(request->old_type);
502 smpi_comm_use(request->comm);
504 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
505 buf, request->real_size,
507 &xbt_free_f, // how to free the userdata if a detached send fails
508 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
509 : &smpi_comm_null_copy_buffer_callback,
511 // detach if msg size < eager/rdv switch limit
513 XBT_DEBUG("send simcall posted");
517 /* FIXME: detached sends are not traceable (request->action == NULL) */
519 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
521 xbt_mutex_release(mut);
526 void smpi_mpi_startall(int count, MPI_Request * requests)
529 if(requests==NULL) return;
531 for(i = 0; i < count; i++) {
532 smpi_mpi_start(requests[i]);
536 void smpi_mpi_request_free(MPI_Request * request)
538 if((*request) != MPI_REQUEST_NULL){
539 (*request)->refcount--;
540 if((*request)->refcount<0) xbt_die("wrong refcount");
542 if((*request)->refcount==0){
543 print_request("Destroying", (*request));
545 *request = MPI_REQUEST_NULL;
547 print_request("Decrementing", (*request));
550 xbt_die("freeing an already free request");
555 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
556 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
558 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
560 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
561 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
563 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
564 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
570 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
571 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
573 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
575 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
576 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
578 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
579 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
586 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
587 int dst, int tag, MPI_Comm comm)
589 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
590 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
591 comm, PERSISTENT | ISEND | SEND | PREPARED);
595 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
596 int dst, int tag, MPI_Comm comm)
598 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
599 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
600 comm, NON_PERSISTENT | ISEND | SEND);
601 smpi_mpi_start(request);
605 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
606 int dst, int tag, MPI_Comm comm)
608 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
609 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
610 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
611 smpi_mpi_start(request);
615 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
616 int src, int tag, MPI_Comm comm)
618 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
619 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,
620 comm, PERSISTENT | RECV | PREPARED);
624 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
625 int src, int tag, MPI_Comm comm)
627 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
628 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,
629 comm, NON_PERSISTENT | RECV);
630 smpi_mpi_start(request);
634 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
635 int tag, MPI_Comm comm, MPI_Status * status)
637 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
638 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
639 smpi_mpi_wait(&request, status);
645 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
646 int tag, MPI_Comm comm)
648 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
649 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
650 comm, NON_PERSISTENT | SEND);
652 smpi_mpi_start(request);
653 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
657 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
658 int dst, int tag, MPI_Comm comm)
660 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
661 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
662 comm, NON_PERSISTENT | SSEND | SEND);
664 smpi_mpi_start(request);
665 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
669 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
670 int dst, int sendtag, void *recvbuf, int recvcount,
671 MPI_Datatype recvtype, int src, int recvtag,
672 MPI_Comm comm, MPI_Status * status)
674 MPI_Request requests[2];
676 int myid=smpi_process_index();
677 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
678 smpi_datatype_copy(sendbuf, sendcount, sendtype,
679 recvbuf, recvcount, recvtype);
683 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
685 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
686 smpi_mpi_startall(2, requests);
687 smpi_mpi_waitall(2, requests, stats);
688 smpi_mpi_request_free(&requests[0]);
689 smpi_mpi_request_free(&requests[1]);
690 if(status != MPI_STATUS_IGNORE) {
691 // Copy receive status
696 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
698 return status->count / smpi_datatype_size(datatype);
701 static void finish_wait(MPI_Request * request, MPI_Status * status)
703 MPI_Request req = *request;
704 smpi_empty_status(status);
706 if(!(req->detached && req->flags & SEND)
707 && !(req->flags & PREPARED)){
708 if(status != MPI_STATUS_IGNORE) {
709 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
710 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
711 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
712 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
713 // this handles the case were size in receive differs from size in send
714 // FIXME: really this should just contain the count of receive-type blocks,
716 status->count = req->real_size;
719 print_request("Finishing", req);
720 MPI_Datatype datatype = req->old_type;
722 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
723 if (!smpi_process_get_replaying()){
724 if( smpi_privatize_global_variables
725 && ((char*)req->old_buf >= smpi_start_data_exe)
726 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )
728 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
729 smpi_switch_data_segment(smpi_process_index());
733 if(datatype->has_subtype == 1){
734 // This part handles the problem of non-contignous memory
735 // the unserialization at the reception
736 s_smpi_subtype_t *subtype = datatype->substruct;
737 if(req->flags & RECV)
738 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
739 if(req->detached == 0) free(req->buf);
740 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
741 int n =req->real_size/smpi_datatype_size(datatype);
742 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
745 smpi_comm_unuse(req->comm);
746 smpi_datatype_unuse(datatype);
750 if (TRACE_smpi_view_internals()) {
751 if(req->flags & RECV){
752 int rank = smpi_process_index();
753 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
754 TRACE_smpi_recv(rank, src_traced, rank);
758 if(req->detached_sender!=NULL){
759 smpi_mpi_request_free(&(req->detached_sender));
761 if(req->flags & PERSISTENT)
763 req->flags |= FINISHED;
765 smpi_mpi_request_free(request);
769 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
772 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
774 // to avoid deadlocks if used as a break condition, such as
775 // while (MPI_Test(request, flag, status) && flag) {
777 // because the time will not normally advance when only calls to MPI_Test
778 // are made -> deadlock
779 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
780 static int nsleeps = 1;
781 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
783 smpi_empty_status(status);
785 if (!((*request)->flags & PREPARED)) {
786 if ((*request)->action != NULL)
787 flag = simcall_comm_test((*request)->action);
789 finish_wait(request, status);
790 nsleeps=1;//reset the number of sleeps we will do next time
791 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
792 *request = MPI_REQUEST_NULL;
800 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
807 *index = MPI_UNDEFINED;
809 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
810 map = xbt_new(int, count);
812 for(i = 0; i < count; i++) {
813 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
814 !(requests[i]->flags & PREPARED)) {
815 xbt_dynar_push(comms, &requests[i]->action);
821 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
822 static int nsleeps = 1;
823 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
825 i = simcall_comm_testany(comms);
826 // not MPI_UNDEFINED, as this is a simix return code
829 finish_wait(&requests[*index], status);
830 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
831 requests[*index] = MPI_REQUEST_NULL;
838 //all requests are null or inactive, return true
840 smpi_empty_status(status);
843 xbt_dynar_free(&comms);
849 int smpi_mpi_testall(int count, MPI_Request requests[],
853 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
856 for(i=0; i<count; i++){
857 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
858 if (smpi_mpi_test(&requests[i], pstat)!=1){
861 requests[i]=MPI_REQUEST_NULL;
864 smpi_empty_status(pstat);
866 if(status != MPI_STATUSES_IGNORE) {
873 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
875 //FIXME find another wait to avoid busy waiting ?
876 // the issue here is that we have to wait on a nonexistent comm
878 smpi_mpi_iprobe(source, tag, comm, &flag, status);
879 XBT_DEBUG("Busy Waiting on probing : %d", flag);
883 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
885 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,
886 comm, PERSISTENT | RECV);
888 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
889 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
890 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
891 static int nsleeps = 1;
892 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
893 // behave like a receive, but don't do it
896 print_request("New iprobe", request);
897 // We have to test both mailboxes as we don't know if we will receive one one or another
898 if (sg_cfg_get_int("smpi/async_small_thres")>0){
899 mailbox = smpi_process_mailbox_small();
900 XBT_DEBUG("trying to probe the perm recv mailbox");
901 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
903 if (request->action==NULL){
904 mailbox = smpi_process_mailbox();
905 XBT_DEBUG("trying to probe the other mailbox");
906 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
909 if (request->action){
910 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
912 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
913 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
914 status->MPI_TAG = req->tag;
915 status->MPI_ERROR = MPI_SUCCESS;
916 status->count = req->real_size;
918 nsleeps=1;//reset the number of sleeps we will do next time
924 smpi_mpi_request_free(&request);
929 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
931 print_request("Waiting", *request);
932 if ((*request)->flags & PREPARED) {
933 smpi_empty_status(status);
937 if ((*request)->action != NULL) { // this is not a detached send
938 simcall_comm_wait((*request)->action, -1.0);
940 if((MC_is_active() || MC_record_replay_is_active()) && (*request)->action)
941 (*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
944 finish_wait(request, status);
945 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
946 *request = MPI_REQUEST_NULL;
947 // FIXME for a detached send, finish_wait is not called:
950 int smpi_mpi_waitany(int count, MPI_Request requests[],
957 index = MPI_UNDEFINED;
959 // Wait for a request to complete
960 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
961 map = xbt_new(int, count);
963 XBT_DEBUG("Wait for one of %d", count);
964 for(i = 0; i < count; i++) {
965 if (requests[i] != MPI_REQUEST_NULL
966 && !(requests[i]->flags & PREPARED)
967 && !(requests[i]->flags & FINISHED)) {
968 if (requests[i]->action != NULL) {
969 XBT_DEBUG("Waiting any %p ", requests[i]);
970 xbt_dynar_push(comms, &requests[i]->action);
974 //This is a finished detached request, let's return this one
975 size=0;//so we free the dynar but don't do the waitany call
977 finish_wait(&requests[i], status);//cleanup if refcount = 0
978 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
979 requests[i]=MPI_REQUEST_NULL;//set to null
985 i = simcall_comm_waitany(comms);
987 // not MPI_UNDEFINED, as this is a simix return code
990 finish_wait(&requests[index], status);
991 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
992 requests[index] = MPI_REQUEST_NULL;
996 xbt_dynar_free(&comms);
999 if (index==MPI_UNDEFINED)
1000 smpi_empty_status(status);
1005 int smpi_mpi_waitall(int count, MPI_Request requests[],
1006 MPI_Status status[])
1010 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1011 int retvalue = MPI_SUCCESS;
1012 //tag invalid requests in the set
1013 if (status != MPI_STATUSES_IGNORE) {
1014 for (c = 0; c < count; c++) {
1015 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
1016 (requests[c]->flags & PREPARED)) {
1017 smpi_empty_status(&status[c]);
1018 } else if (requests[c]->src == MPI_PROC_NULL) {
1019 smpi_empty_status(&status[c]);
1020 status[c].MPI_SOURCE = MPI_PROC_NULL;
1024 for(c = 0; c < count; c++) {
1026 if (MC_is_active() || MC_record_replay_is_active()) {
1027 smpi_mpi_wait(&requests[c], pstat);
1030 index = smpi_mpi_waitany(count, requests, pstat);
1031 if (index == MPI_UNDEFINED)
1033 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1034 requests[index]=MPI_REQUEST_NULL;
1036 if (status != MPI_STATUSES_IGNORE) {
1037 status[index] = *pstat;
1038 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1039 retvalue = MPI_ERR_IN_STATUS;
1046 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
1047 MPI_Status status[])
1049 int i, count, index;
1051 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1054 for(i = 0; i < incount; i++)
1056 index=smpi_mpi_waitany(incount, requests, pstat);
1057 if(index!=MPI_UNDEFINED){
1058 indices[count] = index;
1060 if(status != MPI_STATUSES_IGNORE) {
1061 status[index] = *pstat;
1063 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1064 requests[index]=MPI_REQUEST_NULL;
1066 return MPI_UNDEFINED;
1072 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
1073 MPI_Status status[])
1075 int i, count, count_dead;
1077 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1081 for(i = 0; i < incount; i++) {
1082 if((requests[i] != MPI_REQUEST_NULL)) {
1083 if(smpi_mpi_test(&requests[i], pstat)) {
1086 if(status != MPI_STATUSES_IGNORE) {
1089 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1090 requests[i]=MPI_REQUEST_NULL;
1096 if(count_dead==incount)return MPI_UNDEFINED;
1100 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1103 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1104 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1107 void smpi_mpi_barrier(MPI_Comm comm)
1109 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1110 nary_tree_barrier(comm, 4);
1113 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1114 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1115 int root, MPI_Comm comm)
1117 int system_tag = COLL_TAG_GATHER;
1118 int rank, size, src, index;
1119 MPI_Aint lb = 0, recvext = 0;
1120 MPI_Request *requests;
1122 rank = smpi_comm_rank(comm);
1123 size = smpi_comm_size(comm);
1125 // Send buffer to root
1126 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1128 // FIXME: check for errors
1129 smpi_datatype_extent(recvtype, &lb, &recvext);
1130 // Local copy from root
1131 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1132 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1133 // Receive buffers from senders
1134 requests = xbt_new(MPI_Request, size - 1);
1136 for(src = 0; src < size; src++) {
1138 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1139 recvcount, recvtype,
1140 src, system_tag, comm);
1144 // Wait for completion of irecv's.
1145 smpi_mpi_startall(size - 1, requests);
1146 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1147 for(src = 0; src < size-1; src++) {
1148 smpi_mpi_request_free(&requests[src]);
1155 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1156 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1160 int rank = smpi_process_index();
1163 /* arbitrarily choose root as rank 0 */
1164 size = smpi_comm_size(comm);
1166 displs = xbt_new(int, size);
1167 for (i = 0; i < size; i++) {
1169 count += recvcounts[i];
1171 tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1173 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1174 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1175 recvcounts[rank], datatype, 0, comm);
1177 smpi_free_tmp_buffer(tmpbuf);
1180 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1181 void *recvbuf, int *recvcounts, int *displs,
1182 MPI_Datatype recvtype, int root, MPI_Comm comm)
1184 int system_tag = COLL_TAG_GATHERV;
1185 int rank, size, src, index;
1186 MPI_Aint lb = 0, recvext = 0;
1187 MPI_Request *requests;
1189 rank = smpi_comm_rank(comm);
1190 size = smpi_comm_size(comm);
1192 // Send buffer to root
1193 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1195 // FIXME: check for errors
1196 smpi_datatype_extent(recvtype, &lb, &recvext);
1197 // Local copy from root
1198 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1199 (char *)recvbuf + displs[root] * recvext,
1200 recvcounts[root], recvtype);
1201 // Receive buffers from senders
1202 requests = xbt_new(MPI_Request, size - 1);
1204 for(src = 0; src < size; src++) {
1207 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1208 recvcounts[src], recvtype, src, system_tag, comm);
1212 // Wait for completion of irecv's.
1213 smpi_mpi_startall(size - 1, requests);
1214 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1215 for(src = 0; src < size-1; src++) {
1216 smpi_mpi_request_free(&requests[src]);
1222 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1223 MPI_Datatype sendtype, void *recvbuf,
1224 int recvcount, MPI_Datatype recvtype,
1227 int system_tag = COLL_TAG_ALLGATHER;
1228 int rank, size, other, index;
1229 MPI_Aint lb = 0, recvext = 0;
1230 MPI_Request *requests;
1232 rank = smpi_comm_rank(comm);
1233 size = smpi_comm_size(comm);
1234 // FIXME: check for errors
1235 smpi_datatype_extent(recvtype, &lb, &recvext);
1236 // Local copy from self
1237 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1238 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1240 // Send/Recv buffers to/from others;
1241 requests = xbt_new(MPI_Request, 2 * (size - 1));
1243 for(other = 0; other < size; other++) {
1246 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1249 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1250 recvcount, recvtype, other,
1255 // Wait for completion of all comms.
1256 smpi_mpi_startall(2 * (size - 1), requests);
1257 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1258 for(other = 0; other < 2*(size-1); other++) {
1259 smpi_mpi_request_free(&requests[other]);
1264 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1265 MPI_Datatype sendtype, void *recvbuf,
1266 int *recvcounts, int *displs,
1267 MPI_Datatype recvtype, MPI_Comm comm)
1269 int system_tag = COLL_TAG_ALLGATHERV;
1270 int rank, size, other, index;
1271 MPI_Aint lb = 0, recvext = 0;
1272 MPI_Request *requests;
1274 rank = smpi_comm_rank(comm);
1275 size = smpi_comm_size(comm);
1276 // FIXME: check for errors
1277 smpi_datatype_extent(recvtype, &lb, &recvext);
1278 // Local copy from self
1279 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1280 (char *)recvbuf + displs[rank] * recvext,
1281 recvcounts[rank], recvtype);
1282 // Send buffers to others;
1283 requests = xbt_new(MPI_Request, 2 * (size - 1));
1285 for(other = 0; other < size; other++) {
1288 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1292 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1293 recvtype, other, system_tag, comm);
1297 // Wait for completion of all comms.
1298 smpi_mpi_startall(2 * (size - 1), requests);
1299 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1300 for(other = 0; other < 2*(size-1); other++) {
1301 smpi_mpi_request_free(&requests[other]);
1306 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1307 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1308 int root, MPI_Comm comm)
1310 int system_tag = COLL_TAG_SCATTER;
1311 int rank, size, dst, index;
1312 MPI_Aint lb = 0, sendext = 0;
1313 MPI_Request *requests;
1315 rank = smpi_comm_rank(comm);
1316 size = smpi_comm_size(comm);
1318 // Recv buffer from root
1319 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1322 // FIXME: check for errors
1323 smpi_datatype_extent(sendtype, &lb, &sendext);
1324 // Local copy from root
1325 if(recvbuf!=MPI_IN_PLACE){
1326 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1327 sendcount, sendtype, recvbuf, recvcount, recvtype);
1329 // Send buffers to receivers
1330 requests = xbt_new(MPI_Request, size - 1);
1332 for(dst = 0; dst < size; dst++) {
1334 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1335 sendcount, sendtype, dst,
1340 // Wait for completion of isend's.
1341 smpi_mpi_startall(size - 1, requests);
1342 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1343 for(dst = 0; dst < size-1; dst++) {
1344 smpi_mpi_request_free(&requests[dst]);
1350 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1351 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1352 MPI_Datatype recvtype, int root, MPI_Comm comm)
1354 int system_tag = COLL_TAG_SCATTERV;
1355 int rank, size, dst, index;
1356 MPI_Aint lb = 0, sendext = 0;
1357 MPI_Request *requests;
1359 rank = smpi_comm_rank(comm);
1360 size = smpi_comm_size(comm);
1362 // Recv buffer from root
1363 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1366 // FIXME: check for errors
1367 smpi_datatype_extent(sendtype, &lb, &sendext);
1368 // Local copy from root
1369 if(recvbuf!=MPI_IN_PLACE){
1370 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1371 sendtype, recvbuf, recvcount, recvtype);
1373 // Send buffers to receivers
1374 requests = xbt_new(MPI_Request, size - 1);
1376 for(dst = 0; dst < size; dst++) {
1379 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1380 sendtype, dst, system_tag, comm);
1384 // Wait for completion of isend's.
1385 smpi_mpi_startall(size - 1, requests);
1386 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1387 for(dst = 0; dst < size-1; dst++) {
1388 smpi_mpi_request_free(&requests[dst]);
1394 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1395 MPI_Datatype datatype, MPI_Op op, int root,
1398 int system_tag = COLL_TAG_REDUCE;
1399 int rank, size, src, index;
1400 MPI_Aint lb = 0, dataext = 0;
1401 MPI_Request *requests;
1405 char* sendtmpbuf = (char*) sendbuf;
1406 if( sendbuf == MPI_IN_PLACE ) {
1407 sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1408 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1411 rank = smpi_comm_rank(comm);
1412 size = smpi_comm_size(comm);
1413 //non commutative case, use a working algo from openmpi
1414 if(!smpi_op_is_commute(op)){
1415 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1416 datatype, op, root, comm);
1421 // Send buffer to root
1422 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1424 // FIXME: check for errors
1425 smpi_datatype_extent(datatype, &lb, &dataext);
1426 // Local copy from root
1427 if (sendtmpbuf && recvbuf)
1428 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1429 // Receive buffers from senders
1430 //TODO: make a MPI_barrier here ?
1431 requests = xbt_new(MPI_Request, size - 1);
1432 tmpbufs = xbt_new(void *, size - 1);
1434 for(src = 0; src < size; src++) {
1436 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1438 if (!smpi_process_get_replaying())
1439 tmpbufs[index] = xbt_malloc(count * dataext);
1441 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1443 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1448 // Wait for completion of irecv's.
1449 smpi_mpi_startall(size - 1, requests);
1450 for(src = 0; src < size - 1; src++) {
1451 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1452 XBT_DEBUG("finished waiting any request with index %d", index);
1453 if(index == MPI_UNDEFINED) {
1456 smpi_mpi_request_free(&requests[index]);
1458 if(op) /* op can be MPI_OP_NULL that does nothing */
1459 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1461 for(index = 0; index < size - 1; index++) {
1462 smpi_free_tmp_buffer(tmpbufs[index]);
1467 if( sendbuf == MPI_IN_PLACE ) {
1468 smpi_free_tmp_buffer(sendtmpbuf);
1473 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1474 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1476 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1477 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1480 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1481 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1483 int system_tag = -888;
1484 int rank, size, other, index;
1485 MPI_Aint lb = 0, dataext = 0;
1486 MPI_Request *requests;
1489 rank = smpi_comm_rank(comm);
1490 size = smpi_comm_size(comm);
1492 // FIXME: check for errors
1493 smpi_datatype_extent(datatype, &lb, &dataext);
1495 // Local copy from self
1496 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1498 // Send/Recv buffers to/from others;
1499 requests = xbt_new(MPI_Request, size - 1);
1500 tmpbufs = xbt_new(void *, rank);
1502 for(other = 0; other < rank; other++) {
1503 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1505 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1507 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1511 for(other = rank + 1; other < size; other++) {
1513 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1516 // Wait for completion of all comms.
1517 smpi_mpi_startall(size - 1, requests);
1519 if(smpi_op_is_commute(op)){
1520 for(other = 0; other < size - 1; other++) {
1521 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1522 if(index == MPI_UNDEFINED) {
1526 // #Request is below rank: it's a irecv
1527 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1531 //non commutative case, wait in order
1532 for(other = 0; other < size - 1; other++) {
1533 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1535 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1539 for(index = 0; index < rank; index++) {
1540 smpi_free_tmp_buffer(tmpbufs[index]);
1542 for(index = 0; index < size-1; index++) {
1543 smpi_mpi_request_free(&requests[index]);
1549 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1550 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1552 int system_tag = -888;
1553 int rank, size, other, index;
1554 MPI_Aint lb = 0, dataext = 0;
1555 MPI_Request *requests;
1557 int recvbuf_is_empty=1;
1558 rank = smpi_comm_rank(comm);
1559 size = smpi_comm_size(comm);
1561 // FIXME: check for errors
1562 smpi_datatype_extent(datatype, &lb, &dataext);
1564 // Send/Recv buffers to/from others;
1565 requests = xbt_new(MPI_Request, size - 1);
1566 tmpbufs = xbt_new(void *, rank);
1568 for(other = 0; other < rank; other++) {
1569 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1571 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1573 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1577 for(other = rank + 1; other < size; other++) {
1579 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1582 // Wait for completion of all comms.
1583 smpi_mpi_startall(size - 1, requests);
1584 if(smpi_op_is_commute(op)){
1585 for(other = 0; other < size - 1; other++) {
1586 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1587 if(index == MPI_UNDEFINED) {
1591 if(recvbuf_is_empty){
1592 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1595 // #Request is below rank: it's a irecv
1596 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1600 //non commutative case, wait in order
1601 for(other = 0; other < size - 1; other++) {
1602 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1604 if(recvbuf_is_empty){
1605 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1607 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1611 for(index = 0; index < rank; index++) {
1612 smpi_free_tmp_buffer(tmpbufs[index]);
1614 for(index = 0; index < size-1; index++) {
1615 smpi_mpi_request_free(&requests[index]);