1 /* Copyright (c) 2007-2015. 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. */
7 #include <xbt/config.hpp>
10 #include "xbt/virtu.h"
12 #include "src/mc/mc_replay.h"
13 #include "xbt/replay.h"
15 #include "src/simix/smx_private.h"
16 #include "surf/surf.h"
17 #include "simgrid/sg_config.h"
18 #include "colls/colls.h"
20 #include "src/simix/SynchroComm.hpp"
22 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
24 static int match_recv(void* a, void* b, smx_synchro_t ignored) {
25 MPI_Request ref = static_cast<MPI_Request>(a);
26 MPI_Request req = static_cast<MPI_Request>(b);
27 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
29 xbt_assert(ref, "Cannot match recv against null reference");
30 xbt_assert(req, "Cannot match recv against null request");
31 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
32 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
33 //we match, we can transfer some values
34 if(ref->src == MPI_ANY_SOURCE)
35 ref->real_src = req->src;
36 if(ref->tag == MPI_ANY_TAG)
37 ref->real_tag = req->tag;
38 if(ref->real_size < req->real_size)
41 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
42 XBT_DEBUG("match succeeded");
47 static int match_send(void* a, void* b,smx_synchro_t ignored) {
48 MPI_Request ref = static_cast<MPI_Request>(a);
49 MPI_Request req = static_cast<MPI_Request>(b);
50 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
51 xbt_assert(ref, "Cannot match send against null reference");
52 xbt_assert(req, "Cannot match send against null request");
54 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
55 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
57 if(req->src == MPI_ANY_SOURCE)
58 req->real_src = ref->src;
59 if(req->tag == MPI_ANY_TAG)
60 req->real_tag = ref->tag;
61 if(req->real_size < ref->real_size)
64 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
65 XBT_DEBUG("match succeeded");
70 // Methods used to parse and store the values for timing injections in smpi
71 // These are taken from surf/network.c and generalized to have more values for each factor
72 typedef struct s_smpi_factor_multival *smpi_os_factor_multival_t;
73 typedef struct s_smpi_factor_multival { // FIXME: this should be merged (deduplicated) with s_smpi_factor defined in network_smpi.c
76 double values[4];//arbitrary set to 4
77 } s_smpi_factor_multival_t;
79 xbt_dynar_t smpi_os_values = NULL;
80 xbt_dynar_t smpi_or_values = NULL;
81 xbt_dynar_t smpi_ois_values = NULL;
83 static simgrid::config::Flag<double> smpi_wtime_sleep(
84 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
85 static simgrid::config::Flag<double> smpi_iprobe_sleep(
86 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
87 static simgrid::config::Flag<double> smpi_test_sleep(
88 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
90 static int factor_cmp(const void *pa, const void *pb)
92 return ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor > (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? 1 :
93 ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor < (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? -1 : 0;
96 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
99 unsigned int iter = 0;
100 s_smpi_factor_multival_t fact;
103 xbt_dynar_t radical_elements2 = NULL;
105 xbt_dynar_t smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_multival_t), NULL);
106 xbt_dynar_t radical_elements = xbt_str_split(smpi_coef_string, ";");
107 xbt_dynar_foreach(radical_elements, iter, value) {
108 memset(&fact, 0, sizeof(s_smpi_factor_multival_t));
109 radical_elements2 = xbt_str_split(value, ":");
110 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
111 xbt_die("Malformed radical for smpi factor: '%s'", smpi_coef_string);
112 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
115 errmsg = bprintf("Invalid factor in chunk #%d: %%s", iter+1);
116 fact.factor = xbt_str_parse_int(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
118 errmsg = bprintf("Invalid factor value %d in chunk #%d: %%s", i, iter+1);
119 fact.values[fact.nb_values] = xbt_str_parse_double(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
125 xbt_dynar_push_as(smpi_factor, s_smpi_factor_multival_t, fact);
126 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
127 xbt_dynar_free(&radical_elements2);
129 xbt_dynar_free(&radical_elements);
130 xbt_dynar_sort(smpi_factor, &factor_cmp);
131 xbt_dynar_foreach(smpi_factor, iter, fact) {
132 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
137 static double smpi_os(double size)
139 if (smpi_os_values == NULL) {
140 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
141 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
143 unsigned int iter = 0;
144 s_smpi_factor_multival_t fact;
146 // Iterate over all the sections that were specified and find the right
147 // value. (fact.factor represents the interval sizes; we want to find the
148 // section that has fact.factor <= size and no other such fact.factor <= size)
149 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
150 xbt_dynar_foreach(smpi_os_values, iter, fact) {
151 if (size <= fact.factor) { // Values already too large, use the previously
152 // computed value of current!
153 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
156 // If the next section is too large, the current section must be used.
157 // Hence, save the cost, as we might have to use it.
158 current = fact.values[0]+fact.values[1]*size;
161 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
166 static double smpi_ois(double size)
168 if (smpi_ois_values == NULL) {
169 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
170 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
172 unsigned int iter = 0;
173 s_smpi_factor_multival_t fact;
175 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
176 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
177 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
178 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
179 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
180 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
183 // If the next section is too large, the current section must be used.
184 // Hence, save the cost, as we might have to use it.
185 current = fact.values[0]+fact.values[1]*size;
188 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
193 static double smpi_or(double size)
195 if (smpi_or_values == NULL) {
196 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
197 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
199 unsigned int iter = 0;
200 s_smpi_factor_multival_t fact;
202 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
203 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
204 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
205 xbt_dynar_foreach(smpi_or_values, iter, fact) {
206 if (size <= fact.factor) { // Values already too large, use the previously
207 // computed value of current!
208 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
211 // If the next section is too large, the current section must be used.
212 // Hence, save the cost, as we might have to use it.
213 current=fact.values[0]+fact.values[1]*size;
216 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
221 double smpi_mpi_wtime(){
223 if (smpi_process_initialized() != 0 &&
224 smpi_process_finalized() == 0 &&
225 smpi_process_get_sampling() == 0) {
227 time = SIMIX_get_clock();
228 // to avoid deadlocks if used as a break condition, such as
229 // while (MPI_Wtime(...) < time_limit) {
232 // because the time will not normally advance when only calls to MPI_Wtime
233 // are made -> deadlock (MPI_Wtime never reaches the time limit)
234 if(smpi_wtime_sleep > 0)
235 simcall_process_sleep(smpi_wtime_sleep);
238 time = SIMIX_get_clock();
243 static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
246 MPI_Request request = NULL;
248 void *old_buf = NULL;
250 request = xbt_new(s_smpi_mpi_request_t, 1);
252 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
254 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
255 // This part handles the problem of non-contiguous memory
257 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
258 if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
259 subtype->serialize(old_buf, buf, count, datatype->substruct);
264 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
265 request->old_buf = old_buf;
266 request->old_type = datatype;
268 request->size = smpi_datatype_size(datatype) * count;
269 smpi_datatype_use(datatype);
273 request->comm = comm;
274 smpi_comm_use(request->comm);
275 request->action = NULL;
276 request->flags = flags;
277 request->detached = 0;
278 request->detached_sender = NULL;
279 request->real_src = 0;
281 request->truncated = 0;
282 request->real_size = 0;
283 request->real_tag = 0;
284 if(flags & PERSISTENT)
285 request->refcount = 1;
287 request->refcount = 0;
288 request->op = MPI_REPLACE;
295 void smpi_empty_status(MPI_Status * status)
297 if(status != MPI_STATUS_IGNORE) {
298 status->MPI_SOURCE = MPI_ANY_SOURCE;
299 status->MPI_TAG = MPI_ANY_TAG;
300 status->MPI_ERROR = MPI_SUCCESS;
305 static void smpi_mpi_request_free_voidp(void* request)
307 MPI_Request req = static_cast<MPI_Request>(request);
308 smpi_mpi_request_free(&req);
311 /* MPI Low level calls */
312 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
313 int dst, int tag, MPI_Comm comm)
315 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
316 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
317 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
321 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
322 int dst, int tag, MPI_Comm comm)
324 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
325 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
326 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
330 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
331 int src, int tag, MPI_Comm comm)
333 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
334 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype,
335 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
336 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
340 void smpi_mpi_start(MPI_Request request)
342 smx_mailbox_t mailbox;
344 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
345 request->flags &= ~PREPARED;
346 request->flags &= ~FINISHED;
349 if ((request->flags & RECV) != 0) {
350 print_request("New recv", request);
352 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
354 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
355 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
356 xbt_mutex_acquire(mut);
358 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
359 mailbox = smpi_process_mailbox();
360 } else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh){
361 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
362 //begin with the more appropriate one : the small one.
363 mailbox = smpi_process_mailbox_small();
364 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
365 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
368 mailbox = smpi_process_mailbox();
369 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
370 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
372 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
373 mailbox = smpi_process_mailbox_small();
376 XBT_DEBUG("yes there was something for us in the large mailbox");
379 mailbox = smpi_process_mailbox_small();
380 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
381 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
384 XBT_DEBUG("No, nothing in the permanent receive mailbox");
385 mailbox = smpi_process_mailbox();
387 XBT_DEBUG("yes there was something for us in the small mailbox");
391 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
392 double sleeptime = (request->detached != 0) ? smpi_or(request->size) : 0.0;
394 simcall_process_sleep(sleeptime);
395 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
398 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
399 request->real_size=request->size;
400 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
401 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
402 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
403 XBT_DEBUG("recv simcall posted");
405 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
406 xbt_mutex_release(mut);
408 int receiver = request->dst;
410 int rank = request->src;
411 if (TRACE_smpi_view_internals()) {
412 TRACE_smpi_send(rank, rank, receiver,request->size);
414 print_request("New send", request);
416 //if we are giving back the control to the user without waiting for completion, we have to inject timings
417 double sleeptime = 0.0;
418 if(request->detached != 0 || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
419 //isend and send timings may be different
420 sleeptime = ((request->flags & ISEND) != 0)? smpi_ois(request->size) : smpi_os(request->size);
424 simcall_process_sleep(sleeptime);
425 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
428 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
430 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
432 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
433 xbt_mutex_acquire(mut);
435 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
436 mailbox = smpi_process_remote_mailbox(receiver);
438 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
439 mailbox = smpi_process_remote_mailbox(receiver);
440 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
441 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
443 if ((request->flags & SSEND) == 0){
444 mailbox = smpi_process_remote_mailbox_small(receiver);
445 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
447 mailbox = smpi_process_remote_mailbox_small(receiver);
448 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
449 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
451 XBT_DEBUG("No, we are first, send to large mailbox");
452 mailbox = smpi_process_remote_mailbox(receiver);
456 XBT_DEBUG("Yes there was something for us in the large mailbox");
459 mailbox = smpi_process_remote_mailbox(receiver);
460 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
463 void* buf = request->buf;
464 if ( ((request->flags & SSEND) == 0) && (((request->flags & RMA) != 0) ||
465 (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
467 request->detached = 1;
468 XBT_DEBUG("Send request %p is detached", request);
470 if(request->old_type->sizeof_substruct == 0){
471 oldbuf = request->buf;
472 if (!smpi_process_get_replaying() && oldbuf != NULL && request->size!=0){
473 if((smpi_privatize_global_variables != 0)
474 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
475 && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
476 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
477 smpi_switch_data_segment(request->src);
479 buf = xbt_malloc(request->size);
480 memcpy(buf,oldbuf,request->size);
481 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
486 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
487 request->real_size=request->size;
488 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
489 buf, request->real_size, &match_send,
490 &xbt_free_f, // how to free the userdata if a detached send fails
491 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
492 : &smpi_comm_null_copy_buffer_callback, request,
493 // detach if msg size < eager/rdv switch limit
495 XBT_DEBUG("send simcall posted");
497 /* FIXME: detached sends are not traceable (request->action == NULL) */
499 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
501 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
502 xbt_mutex_release(mut);
506 void smpi_mpi_startall(int count, MPI_Request * requests)
511 for(int i = 0; i < count; i++) {
512 smpi_mpi_start(requests[i]);
516 void smpi_mpi_request_free(MPI_Request * request)
518 if((*request) != MPI_REQUEST_NULL){
519 (*request)->refcount--;
520 if((*request)->refcount<0) xbt_die("wrong refcount");
522 if((*request)->refcount==0){
523 smpi_datatype_unuse((*request)->old_type);
524 smpi_comm_unuse((*request)->comm);
525 print_request("Destroying", (*request));
527 *request = MPI_REQUEST_NULL;
529 print_request("Decrementing", (*request));
532 xbt_die("freeing an already free request");
536 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
539 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
541 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, src, dst, tag,
542 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
544 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
545 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
551 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
554 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
556 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
557 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
559 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
560 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
566 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
568 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
569 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, smpi_process_index(),
570 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
574 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
576 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
577 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
578 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
579 smpi_mpi_start(request);
583 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
585 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
586 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
587 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
588 smpi_mpi_start(request);
592 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
594 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
595 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
596 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
597 comm, PERSISTENT | RECV | PREPARED);
601 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
603 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
604 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
605 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
606 NON_PERSISTENT | RECV);
607 smpi_mpi_start(request);
611 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
613 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
614 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
615 smpi_mpi_wait(&request, status);
619 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
621 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
622 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
623 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
625 smpi_mpi_start(request);
626 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
630 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
632 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
633 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
634 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
636 smpi_mpi_start(request);
637 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
641 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
642 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
643 MPI_Comm comm, MPI_Status * status)
645 MPI_Request requests[2];
647 int myid=smpi_process_index();
648 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
649 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
652 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
653 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
654 smpi_mpi_startall(2, requests);
655 smpi_mpi_waitall(2, requests, stats);
656 smpi_mpi_request_free(&requests[0]);
657 smpi_mpi_request_free(&requests[1]);
658 if(status != MPI_STATUS_IGNORE) {
659 // Copy receive status
664 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
666 return status->count / smpi_datatype_size(datatype);
669 static void finish_wait(MPI_Request * request, MPI_Status * status)
671 MPI_Request req = *request;
672 smpi_empty_status(status);
674 if(!((req->detached != 0) && req->flags & SEND) && ((req->flags & PREPARED) == 0)){
675 if(status != MPI_STATUS_IGNORE) {
676 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
677 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
678 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
679 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
680 // this handles the case were size in receive differs from size in send
681 status->count = req->real_size;
684 print_request("Finishing", req);
685 MPI_Datatype datatype = req->old_type;
687 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
688 if (!smpi_process_get_replaying()){
689 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
690 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
691 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
692 smpi_switch_data_segment(smpi_process_index());
696 if(datatype->sizeof_substruct != 0){
697 // This part handles the problem of non-contignous memory the unserialization at the reception
698 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
699 if(req->flags & RECV)
700 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
701 datatype->substruct, req->op);
703 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
704 int n =req->real_size/smpi_datatype_size(datatype);
705 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
711 if (TRACE_smpi_view_internals() && req->flags & RECV){
712 int rank = smpi_process_index();
713 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
714 TRACE_smpi_recv(rank, src_traced, rank);
717 if(req->detached_sender!=NULL){
718 smpi_mpi_request_free(&(req->detached_sender));
720 if(req->flags & PERSISTENT)
722 req->flags |= FINISHED;
724 smpi_mpi_request_free(request);
727 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
728 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
730 // to avoid deadlocks if used as a break condition, such as
731 // while (MPI_Test(request, flag, status) && flag) {
733 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
734 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
735 static int nsleeps = 1;
736 if(smpi_test_sleep > 0)
737 simcall_process_sleep(nsleeps*smpi_test_sleep);
739 smpi_empty_status(status);
741 if (((*request)->flags & PREPARED) == 0) {
742 if ((*request)->action != NULL)
743 flag = simcall_comm_test((*request)->action);
745 finish_wait(request, status);
746 nsleeps=1;//reset the number of sleeps we will do next time
747 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
748 *request = MPI_REQUEST_NULL;
756 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
764 *index = MPI_UNDEFINED;
765 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
766 map = xbt_new(int, count);
767 for(i = 0; i < count; i++) {
768 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
769 xbt_dynar_push(comms, &requests[i]->action);
775 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
776 static int nsleeps = 1;
777 if(smpi_test_sleep > 0)
778 simcall_process_sleep(nsleeps*smpi_test_sleep);
780 i = simcall_comm_testany(comms);
781 // not MPI_UNDEFINED, as this is a simix return code
784 finish_wait(&requests[*index], status);
785 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
786 requests[*index] = MPI_REQUEST_NULL;
793 //all requests are null or inactive, return true
795 smpi_empty_status(status);
798 xbt_dynar_free(&comms);
803 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
806 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
809 for(i=0; i<count; i++){
810 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
811 if (smpi_mpi_test(&requests[i], pstat)!=1){
814 requests[i]=MPI_REQUEST_NULL;
817 smpi_empty_status(pstat);
819 if(status != MPI_STATUSES_IGNORE) {
826 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
828 //FIXME find another way to avoid busy waiting ?
829 // the issue here is that we have to wait on a nonexistent comm
831 smpi_mpi_iprobe(source, tag, comm, &flag, status);
832 XBT_DEBUG("Busy Waiting on probing : %d", flag);
836 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
838 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
839 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
841 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
842 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
843 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
844 static int nsleeps = 1;
845 if(smpi_iprobe_sleep > 0)
846 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
847 // behave like a receive, but don't do it
848 smx_mailbox_t mailbox;
850 print_request("New iprobe", request);
851 // We have to test both mailboxes as we don't know if we will receive one one or another
852 if (xbt_cfg_get_int("smpi/async-small-thresh")>0){
853 mailbox = smpi_process_mailbox_small();
854 XBT_DEBUG("trying to probe the perm recv mailbox");
855 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
857 if (request->action==NULL){
858 mailbox = smpi_process_mailbox();
859 XBT_DEBUG("trying to probe the other mailbox");
860 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
863 if (request->action){
864 simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
865 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
867 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED)==0) {
868 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
869 status->MPI_TAG = req->tag;
870 status->MPI_ERROR = MPI_SUCCESS;
871 status->count = req->real_size;
873 nsleeps=1;//reset the number of sleeps we will do next time
879 smpi_mpi_request_free(&request);
884 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
886 print_request("Waiting", *request);
887 if ((*request)->flags & PREPARED) {
888 smpi_empty_status(status);
892 if ((*request)->action != NULL)
893 // this is not a detached send
894 simcall_comm_wait((*request)->action, -1.0);
896 finish_wait(request, status);
897 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
898 *request = MPI_REQUEST_NULL;
901 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
906 int index = MPI_UNDEFINED;
910 // Wait for a request to complete
911 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
912 map = xbt_new(int, count);
913 XBT_DEBUG("Wait for one of %d", count);
914 for(i = 0; i < count; i++) {
915 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
916 if (requests[i]->action != NULL) {
917 XBT_DEBUG("Waiting any %p ", requests[i]);
918 xbt_dynar_push(comms, &requests[i]->action);
922 //This is a finished detached request, let's return this one
923 size=0;//so we free the dynar but don't do the waitany call
925 finish_wait(&requests[i], status);//cleanup if refcount = 0
926 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
927 requests[i]=MPI_REQUEST_NULL;//set to null
933 i = simcall_comm_waitany(comms);
935 // not MPI_UNDEFINED, as this is a simix return code
938 finish_wait(&requests[index], status);
939 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
940 requests[index] = MPI_REQUEST_NULL;
944 xbt_dynar_free(&comms);
947 if (index==MPI_UNDEFINED)
948 smpi_empty_status(status);
953 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
957 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
958 int retvalue = MPI_SUCCESS;
959 //tag invalid requests in the set
960 if (status != MPI_STATUSES_IGNORE) {
961 for (c = 0; c < count; c++) {
962 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
963 smpi_empty_status(&status[c]);
964 } else if (requests[c]->src == MPI_PROC_NULL) {
965 smpi_empty_status(&status[c]);
966 status[c].MPI_SOURCE = MPI_PROC_NULL;
970 for(c = 0; c < count; c++) {
972 if (MC_is_active() || MC_record_replay_is_active()) {
973 smpi_mpi_wait(&requests[c], pstat);
976 index = smpi_mpi_waitany(count, requests, pstat);
977 if (index == MPI_UNDEFINED)
979 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
980 requests[index]=MPI_REQUEST_NULL;
982 if (status != MPI_STATUSES_IGNORE) {
983 status[index] = *pstat;
984 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
985 retvalue = MPI_ERR_IN_STATUS;
992 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
998 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1000 for(i = 0; i < incount; i++)
1002 index=smpi_mpi_waitany(incount, requests, pstat);
1003 if(index!=MPI_UNDEFINED){
1004 indices[count] = index;
1006 if(status != MPI_STATUSES_IGNORE) {
1007 status[index] = *pstat;
1009 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1010 requests[index]=MPI_REQUEST_NULL;
1012 return MPI_UNDEFINED;
1018 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1024 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1026 for(i = 0; i < incount; i++) {
1027 if((requests[i] != MPI_REQUEST_NULL)) {
1028 if(smpi_mpi_test(&requests[i], pstat)) {
1031 if(status != MPI_STATUSES_IGNORE) {
1034 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1035 requests[i]=MPI_REQUEST_NULL;
1041 if(count_dead==incount)
1042 return MPI_UNDEFINED;
1046 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
1048 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
1051 void smpi_mpi_barrier(MPI_Comm comm)
1053 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
1056 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1057 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1059 int system_tag = COLL_TAG_GATHER;
1060 int rank, size, src, index;
1061 MPI_Aint lb = 0, recvext = 0;
1062 MPI_Request *requests;
1064 rank = smpi_comm_rank(comm);
1065 size = smpi_comm_size(comm);
1067 // Send buffer to root
1068 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1070 smpi_datatype_extent(recvtype, &lb, &recvext);
1071 // Local copy from root
1072 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext, recvcount, recvtype);
1073 // Receive buffers from senders
1074 requests = xbt_new(MPI_Request, size - 1);
1076 for(src = 0; src < size; src++) {
1078 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1079 src, system_tag, comm);
1083 // Wait for completion of irecv's.
1084 smpi_mpi_startall(size - 1, requests);
1085 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1086 for(src = 0; src < size-1; src++) {
1087 smpi_mpi_request_free(&requests[src]);
1093 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1098 int rank = smpi_comm_rank(comm);
1101 /* arbitrarily choose root as rank 0 */
1102 size = smpi_comm_size(comm);
1104 displs = xbt_new(int, size);
1105 for (i = 0; i < size; i++) {
1107 count += recvcounts[i];
1109 tmpbuf=static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1111 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1112 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1114 smpi_free_tmp_buffer(tmpbuf);
1117 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1118 MPI_Datatype recvtype, int root, MPI_Comm comm)
1120 int system_tag = COLL_TAG_GATHERV;
1121 int rank, size, src, index;
1122 MPI_Aint lb = 0, recvext = 0;
1123 MPI_Request *requests;
1125 rank = smpi_comm_rank(comm);
1126 size = smpi_comm_size(comm);
1128 // Send buffer to root
1129 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1131 smpi_datatype_extent(recvtype, &lb, &recvext);
1132 // Local copy from root
1133 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1134 recvcounts[root], recvtype);
1135 // Receive buffers from senders
1136 requests = xbt_new(MPI_Request, size - 1);
1138 for(src = 0; src < size; src++) {
1140 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1141 recvcounts[src], recvtype, src, system_tag, comm);
1145 // Wait for completion of irecv's.
1146 smpi_mpi_startall(size - 1, requests);
1147 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1148 for(src = 0; src < size-1; src++) {
1149 smpi_mpi_request_free(&requests[src]);
1155 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1156 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1158 int system_tag = COLL_TAG_ALLGATHER;
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, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
1169 // Send/Recv buffers to/from others;
1170 requests = xbt_new(MPI_Request, 2 * (size - 1));
1172 for(other = 0; other < size; other++) {
1174 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1176 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other,
1181 // Wait for completion of all comms.
1182 smpi_mpi_startall(2 * (size - 1), requests);
1183 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1184 for(other = 0; other < 2*(size-1); other++) {
1185 smpi_mpi_request_free(&requests[other]);
1190 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1191 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1193 int system_tag = COLL_TAG_ALLGATHERV;
1194 int rank, size, other, index;
1195 MPI_Aint lb = 0, recvext = 0;
1196 MPI_Request *requests;
1198 rank = smpi_comm_rank(comm);
1199 size = smpi_comm_size(comm);
1200 smpi_datatype_extent(recvtype, &lb, &recvext);
1201 // Local copy from self
1202 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1203 // Send buffers to others;
1204 requests = xbt_new(MPI_Request, 2 * (size - 1));
1206 for(other = 0; other < size; other++) {
1209 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1211 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1212 recvtype, other, system_tag, comm);
1216 // Wait for completion of all comms.
1217 smpi_mpi_startall(2 * (size - 1), requests);
1218 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1219 for(other = 0; other < 2*(size-1); other++) {
1220 smpi_mpi_request_free(&requests[other]);
1225 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1226 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1228 int system_tag = COLL_TAG_SCATTER;
1229 int rank, size, dst, index;
1230 MPI_Aint lb = 0, sendext = 0;
1231 MPI_Request *requests;
1233 rank = smpi_comm_rank(comm);
1234 size = smpi_comm_size(comm);
1236 // Recv buffer from root
1237 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1239 smpi_datatype_extent(sendtype, &lb, &sendext);
1240 // Local copy from root
1241 if(recvbuf!=MPI_IN_PLACE){
1242 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1243 sendcount, sendtype, recvbuf, recvcount, recvtype);
1245 // Send buffers to receivers
1246 requests = xbt_new(MPI_Request, size - 1);
1248 for(dst = 0; dst < size; dst++) {
1250 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst,
1255 // Wait for completion of isend's.
1256 smpi_mpi_startall(size - 1, requests);
1257 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1258 for(dst = 0; dst < size-1; dst++) {
1259 smpi_mpi_request_free(&requests[dst]);
1265 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1266 MPI_Datatype recvtype, int root, MPI_Comm comm)
1268 int system_tag = COLL_TAG_SCATTERV;
1269 int rank, size, dst, index;
1270 MPI_Aint lb = 0, sendext = 0;
1271 MPI_Request *requests;
1273 rank = smpi_comm_rank(comm);
1274 size = smpi_comm_size(comm);
1276 // Recv buffer from root
1277 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1279 smpi_datatype_extent(sendtype, &lb, &sendext);
1280 // Local copy from root
1281 if(recvbuf!=MPI_IN_PLACE){
1282 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1283 sendtype, recvbuf, recvcount, recvtype);
1285 // Send buffers to receivers
1286 requests = xbt_new(MPI_Request, size - 1);
1288 for(dst = 0; dst < size; dst++) {
1290 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1291 sendtype, dst, system_tag, comm);
1295 // Wait for completion of isend's.
1296 smpi_mpi_startall(size - 1, requests);
1297 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1298 for(dst = 0; dst < size-1; dst++) {
1299 smpi_mpi_request_free(&requests[dst]);
1305 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1308 int system_tag = COLL_TAG_REDUCE;
1309 int rank, size, src, index;
1310 MPI_Aint lb = 0, dataext = 0;
1311 MPI_Request *requests;
1314 char* sendtmpbuf = static_cast<char *>(sendbuf);
1317 rank = smpi_comm_rank(comm);
1318 size = smpi_comm_size(comm);
1319 //non commutative case, use a working algo from openmpi
1320 if(!smpi_op_is_commute(op)){
1321 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1325 if( sendbuf == MPI_IN_PLACE ) {
1326 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1327 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1331 // Send buffer to root
1332 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1334 smpi_datatype_extent(datatype, &lb, &dataext);
1335 // Local copy from root
1336 if (sendtmpbuf != NULL && recvbuf != NULL)
1337 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1338 // Receive buffers from senders
1339 requests = xbt_new(MPI_Request, size - 1);
1340 tmpbufs = xbt_new(void *, size - 1);
1342 for(src = 0; src < size; src++) {
1344 if (!smpi_process_get_replaying())
1345 tmpbufs[index] = xbt_malloc(count * dataext);
1347 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1349 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1353 // Wait for completion of irecv's.
1354 smpi_mpi_startall(size - 1, requests);
1355 for(src = 0; src < size - 1; src++) {
1356 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1357 XBT_DEBUG("finished waiting any request with index %d", index);
1358 if(index == MPI_UNDEFINED) {
1361 smpi_mpi_request_free(&requests[index]);
1363 if(op) /* op can be MPI_OP_NULL that does nothing */
1364 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1366 for(index = 0; index < size - 1; index++) {
1367 smpi_free_tmp_buffer(tmpbufs[index]);
1373 if( sendbuf == MPI_IN_PLACE ) {
1374 smpi_free_tmp_buffer(sendtmpbuf);
1378 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1380 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1381 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1384 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1386 int system_tag = -888;
1387 int rank, size, other, index;
1388 MPI_Aint lb = 0, dataext = 0;
1389 MPI_Request *requests;
1392 rank = smpi_comm_rank(comm);
1393 size = smpi_comm_size(comm);
1395 smpi_datatype_extent(datatype, &lb, &dataext);
1397 // Local copy from self
1398 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1400 // Send/Recv buffers to/from others;
1401 requests = xbt_new(MPI_Request, size - 1);
1402 tmpbufs = xbt_new(void *, rank);
1404 for(other = 0; other < rank; other++) {
1405 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1406 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1409 for(other = rank + 1; other < size; other++) {
1410 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1413 // Wait for completion of all comms.
1414 smpi_mpi_startall(size - 1, requests);
1416 if(smpi_op_is_commute(op)){
1417 for(other = 0; other < size - 1; other++) {
1418 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1419 if(index == MPI_UNDEFINED) {
1423 // #Request is below rank: it's a irecv
1424 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1428 //non commutative case, wait in order
1429 for(other = 0; other < size - 1; other++) {
1430 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1432 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1436 for(index = 0; index < rank; index++) {
1437 smpi_free_tmp_buffer(tmpbufs[index]);
1439 for(index = 0; index < size-1; index++) {
1440 smpi_mpi_request_free(&requests[index]);
1446 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1448 int system_tag = -888;
1449 int rank, size, other, index;
1450 MPI_Aint lb = 0, dataext = 0;
1451 MPI_Request *requests;
1453 int recvbuf_is_empty=1;
1454 rank = smpi_comm_rank(comm);
1455 size = smpi_comm_size(comm);
1457 smpi_datatype_extent(datatype, &lb, &dataext);
1459 // Send/Recv buffers to/from others;
1460 requests = xbt_new(MPI_Request, size - 1);
1461 tmpbufs = xbt_new(void *, rank);
1463 for(other = 0; other < rank; other++) {
1464 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1466 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1469 for(other = rank + 1; other < size; other++) {
1471 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1474 // Wait for completion of all comms.
1475 smpi_mpi_startall(size - 1, requests);
1476 if(smpi_op_is_commute(op)){
1477 for(other = 0; other < size - 1; other++) {
1478 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1479 if(index == MPI_UNDEFINED) {
1483 if(recvbuf_is_empty){
1484 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1487 // #Request is below rank: it's a irecv
1488 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1492 //non commutative case, wait in order
1493 for(other = 0; other < size - 1; other++) {
1494 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1496 if(recvbuf_is_empty){
1497 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1499 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1503 for(index = 0; index < rank; index++) {
1504 smpi_free_tmp_buffer(tmpbufs[index]);
1506 for(index = 0; index < size-1; index++) {
1507 smpi_mpi_request_free(&requests[index]);