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 = nullptr;
80 xbt_dynar_t smpi_or_values = nullptr;
81 xbt_dynar_t smpi_ois_values = nullptr;
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_init_sleep(
86 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
87 static simgrid::config::Flag<double> smpi_iprobe_sleep(
88 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
89 static simgrid::config::Flag<double> smpi_test_sleep(
90 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
92 static int factor_cmp(const void *pa, const void *pb)
94 return ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor > (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? 1 :
95 ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor < (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? -1 : 0;
98 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
101 unsigned int iter = 0;
102 s_smpi_factor_multival_t fact;
105 xbt_dynar_t radical_elements2 = NULL;
107 xbt_dynar_t smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_multival_t), NULL);
108 xbt_dynar_t radical_elements = xbt_str_split(smpi_coef_string, ";");
109 xbt_dynar_foreach(radical_elements, iter, value) {
110 memset(&fact, 0, sizeof(s_smpi_factor_multival_t));
111 radical_elements2 = xbt_str_split(value, ":");
112 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
113 xbt_die("Malformed radical for smpi factor: '%s'", smpi_coef_string);
114 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
117 errmsg = bprintf("Invalid factor in chunk #%d: %%s", iter+1);
118 fact.factor = xbt_str_parse_int(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
120 errmsg = bprintf("Invalid factor value %d in chunk #%d: %%s", i, iter+1);
121 fact.values[fact.nb_values] = xbt_str_parse_double(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
127 xbt_dynar_push_as(smpi_factor, s_smpi_factor_multival_t, fact);
128 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
129 xbt_dynar_free(&radical_elements2);
131 xbt_dynar_free(&radical_elements);
132 xbt_dynar_sort(smpi_factor, &factor_cmp);
133 xbt_dynar_foreach(smpi_factor, iter, fact) {
134 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
139 static double smpi_os(double size)
141 if (smpi_os_values == nullptr) {
142 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
143 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
145 unsigned int iter = 0;
146 s_smpi_factor_multival_t fact;
148 // Iterate over all the sections that were specified and find the right
149 // value. (fact.factor represents the interval sizes; we want to find the
150 // section that has fact.factor <= size and no other such fact.factor <= size)
151 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
152 xbt_dynar_foreach(smpi_os_values, iter, fact) {
153 if (size <= fact.factor) { // Values already too large, use the previously
154 // computed value of current!
155 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
158 // If the next section is too large, the current section must be used.
159 // Hence, save the cost, as we might have to use it.
160 current = fact.values[0]+fact.values[1]*size;
163 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
168 static double smpi_ois(double size)
170 if (smpi_ois_values == nullptr) {
171 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
172 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
174 unsigned int iter = 0;
175 s_smpi_factor_multival_t fact;
177 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
178 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
179 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
180 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
181 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
182 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
185 // If the next section is too large, the current section must be used.
186 // Hence, save the cost, as we might have to use it.
187 current = fact.values[0]+fact.values[1]*size;
190 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
195 static double smpi_or(double size)
197 if (smpi_or_values == nullptr) {
198 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
199 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
201 unsigned int iter = 0;
202 s_smpi_factor_multival_t fact;
204 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
205 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
206 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
207 xbt_dynar_foreach(smpi_or_values, iter, fact) {
208 if (size <= fact.factor) { // Values already too large, use the previously
209 // computed value of current!
210 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
213 // If the next section is too large, the current section must be used.
214 // Hence, save the cost, as we might have to use it.
215 current=fact.values[0]+fact.values[1]*size;
218 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
223 void smpi_mpi_init() {
224 if(smpi_init_sleep > 0)
225 simcall_process_sleep(smpi_init_sleep);
228 double smpi_mpi_wtime(){
230 if (smpi_process_initialized() != 0 &&
231 smpi_process_finalized() == 0 &&
232 smpi_process_get_sampling() == 0) {
234 time = SIMIX_get_clock();
235 // to avoid deadlocks if used as a break condition, such as
236 // while (MPI_Wtime(...) < time_limit) {
239 // because the time will not normally advance when only calls to MPI_Wtime
240 // are made -> deadlock (MPI_Wtime never reaches the time limit)
241 if(smpi_wtime_sleep > 0)
242 simcall_process_sleep(smpi_wtime_sleep);
245 time = SIMIX_get_clock();
250 static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
253 MPI_Request request = NULL;
255 void *old_buf = NULL;
257 request = xbt_new(s_smpi_mpi_request_t, 1);
259 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
261 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
262 // This part handles the problem of non-contiguous memory
264 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
265 if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
266 subtype->serialize(old_buf, buf, count, datatype->substruct);
271 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
272 request->old_buf = old_buf;
273 request->old_type = datatype;
275 request->size = smpi_datatype_size(datatype) * count;
276 smpi_datatype_use(datatype);
280 request->comm = comm;
281 smpi_comm_use(request->comm);
282 request->action = nullptr;
283 request->flags = flags;
284 request->detached = 0;
285 request->detached_sender = nullptr;
286 request->real_src = 0;
287 request->truncated = 0;
288 request->real_size = 0;
289 request->real_tag = 0;
290 if (flags & PERSISTENT)
291 request->refcount = 1;
293 request->refcount = 0;
294 request->op = MPI_REPLACE;
301 void smpi_empty_status(MPI_Status * status)
303 if(status != MPI_STATUS_IGNORE) {
304 status->MPI_SOURCE = MPI_ANY_SOURCE;
305 status->MPI_TAG = MPI_ANY_TAG;
306 status->MPI_ERROR = MPI_SUCCESS;
311 static void smpi_mpi_request_free_voidp(void* request)
313 MPI_Request req = static_cast<MPI_Request>(request);
314 smpi_mpi_request_free(&req);
317 /* MPI Low level calls */
318 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
319 int dst, int tag, MPI_Comm comm)
321 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
322 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
323 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
327 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
328 int dst, int tag, MPI_Comm comm)
330 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
331 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
332 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
336 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
337 int src, int tag, MPI_Comm comm)
339 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
340 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype,
341 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
342 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
346 void smpi_mpi_start(MPI_Request request)
348 smx_mailbox_t mailbox;
350 xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
351 request->flags &= ~PREPARED;
352 request->flags &= ~FINISHED;
355 if ((request->flags & RECV) != 0) {
356 print_request("New recv", request);
358 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
360 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
361 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
362 xbt_mutex_acquire(mut);
364 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
365 mailbox = smpi_process_mailbox();
367 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) {
368 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
369 //begin with the more appropriate one : the small one.
370 mailbox = smpi_process_mailbox_small();
371 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
372 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
374 if (action == nullptr) {
375 mailbox = smpi_process_mailbox();
376 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
377 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
378 if (action == nullptr) {
379 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
380 mailbox = smpi_process_mailbox_small();
384 XBT_DEBUG("yes there was something for us in the large mailbox");
388 mailbox = smpi_process_mailbox_small();
389 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
390 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
392 if (action == nullptr) {
393 XBT_DEBUG("No, nothing in the permanent receive mailbox");
394 mailbox = smpi_process_mailbox();
397 XBT_DEBUG("yes there was something for us in the small mailbox");
401 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
402 double sleeptime = (request->detached != 0) ? smpi_or(request->size) : 0.0;
404 simcall_process_sleep(sleeptime);
405 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
408 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
409 request->real_size=request->size;
410 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
411 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
412 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
413 XBT_DEBUG("recv simcall posted");
415 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
416 xbt_mutex_release(mut);
418 int receiver = request->dst;
420 int rank = request->src;
421 if (TRACE_smpi_view_internals()) {
422 TRACE_smpi_send(rank, rank, receiver,request->size);
424 print_request("New send", request);
426 //if we are giving back the control to the user without waiting for completion, we have to inject timings
427 double sleeptime = 0.0;
428 if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
429 //isend and send timings may be different
430 sleeptime = ((request->flags & ISEND) != 0)? smpi_ois(request->size) : smpi_os(request->size);
434 simcall_process_sleep(sleeptime);
435 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
438 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
440 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
442 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
443 xbt_mutex_acquire(mut);
445 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
446 mailbox = smpi_process_remote_mailbox(receiver);
448 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
449 mailbox = smpi_process_remote_mailbox(receiver);
450 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
451 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
452 if (action == nullptr) {
453 if ((request->flags & SSEND) == 0){
454 mailbox = smpi_process_remote_mailbox_small(receiver);
455 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
458 mailbox = smpi_process_remote_mailbox_small(receiver);
459 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
460 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
461 if (action == nullptr) {
462 XBT_DEBUG("No, we are first, send to large mailbox");
463 mailbox = smpi_process_remote_mailbox(receiver);
468 XBT_DEBUG("Yes there was something for us in the large mailbox");
472 mailbox = smpi_process_remote_mailbox(receiver);
473 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
476 void* buf = request->buf;
477 if ( ((request->flags & SSEND) == 0) && (((request->flags & RMA) != 0) ||
478 (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
480 request->detached = 1;
481 XBT_DEBUG("Send request %p is detached", request);
483 if(request->old_type->sizeof_substruct == 0){
484 oldbuf = request->buf;
485 if (!smpi_process_get_replaying() && oldbuf != NULL && request->size!=0){
486 if((smpi_privatize_global_variables != 0)
487 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
488 && (static_cast<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 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
502 buf, request->real_size, &match_send,
503 &xbt_free_f, // how to free the userdata if a detached send fails
504 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
505 : &smpi_comm_null_copy_buffer_callback, request,
506 // detach if msg size < eager/rdv switch limit
508 XBT_DEBUG("send simcall posted");
510 /* FIXME: detached sends are not traceable (request->action == nullptr) */
511 if (request->action != nullptr)
512 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
514 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
515 xbt_mutex_release(mut);
519 void smpi_mpi_startall(int count, MPI_Request * requests)
521 if(requests== nullptr)
524 for(int i = 0; i < count; i++) {
525 smpi_mpi_start(requests[i]);
529 void smpi_mpi_request_free(MPI_Request * request)
531 if((*request) != MPI_REQUEST_NULL){
532 (*request)->refcount--;
533 if((*request)->refcount<0) xbt_die("wrong refcount");
535 if((*request)->refcount==0){
536 smpi_datatype_unuse((*request)->old_type);
537 smpi_comm_unuse((*request)->comm);
538 print_request("Destroying", (*request));
540 *request = MPI_REQUEST_NULL;
542 print_request("Decrementing", (*request));
545 xbt_die("freeing an already free request");
549 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
552 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
554 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, src, dst, tag,
555 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
557 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
558 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
564 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
567 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, src, dst, tag,
570 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
572 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
573 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
579 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
581 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
582 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, smpi_process_index(),
583 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
587 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, 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 ? NULL : buf, count, datatype, smpi_process_index(),
591 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
592 smpi_mpi_start(request);
596 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, 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 ? NULL : buf, count, datatype, smpi_process_index(),
600 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
601 smpi_mpi_start(request);
605 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
607 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
608 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
609 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
610 comm, PERSISTENT | RECV | PREPARED);
614 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
616 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
617 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
618 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
619 NON_PERSISTENT | RECV);
620 smpi_mpi_start(request);
624 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
626 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
627 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
628 smpi_mpi_wait(&request, status);
632 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
634 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
635 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
636 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
638 smpi_mpi_start(request);
639 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
643 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
645 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
646 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
647 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
649 smpi_mpi_start(request);
650 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
654 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
655 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
656 MPI_Comm comm, MPI_Status * status)
658 MPI_Request requests[2];
660 int myid=smpi_process_index();
661 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
662 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
665 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
666 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
667 smpi_mpi_startall(2, requests);
668 smpi_mpi_waitall(2, requests, stats);
669 smpi_mpi_request_free(&requests[0]);
670 smpi_mpi_request_free(&requests[1]);
671 if(status != MPI_STATUS_IGNORE) {
672 // Copy receive status
677 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
679 return status->count / smpi_datatype_size(datatype);
682 static void finish_wait(MPI_Request * request, MPI_Status * status)
684 MPI_Request req = *request;
685 smpi_empty_status(status);
687 if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
688 if(status != MPI_STATUS_IGNORE) {
689 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
690 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
691 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
692 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
693 // this handles the case were size in receive differs from size in send
694 status->count = req->real_size;
697 print_request("Finishing", req);
698 MPI_Datatype datatype = req->old_type;
700 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
701 if (!smpi_process_get_replaying()){
702 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
703 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
704 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
705 smpi_switch_data_segment(smpi_process_index());
709 if(datatype->sizeof_substruct != 0){
710 // This part handles the problem of non-contignous memory the unserialization at the reception
711 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
712 if(req->flags & RECV)
713 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
714 datatype->substruct, req->op);
716 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
717 int n =req->real_size/smpi_datatype_size(datatype);
718 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
724 if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
725 int rank = smpi_process_index();
726 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
727 TRACE_smpi_recv(rank, src_traced, rank);
730 if(req->detached_sender!=NULL){
731 smpi_mpi_request_free(&(req->detached_sender));
733 if(req->flags & PERSISTENT)
735 req->flags |= FINISHED;
737 smpi_mpi_request_free(request);
740 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
741 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
743 // to avoid deadlocks if used as a break condition, such as
744 // while (MPI_Test(request, flag, status) && flag) {
746 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
747 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
748 static int nsleeps = 1;
749 if(smpi_test_sleep > 0)
750 simcall_process_sleep(nsleeps*smpi_test_sleep);
752 smpi_empty_status(status);
754 if (((*request)->flags & PREPARED) == 0) {
755 if ((*request)->action != NULL)
756 flag = simcall_comm_test((*request)->action);
758 finish_wait(request, status);
759 nsleeps=1;//reset the number of sleeps we will do next time
760 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
761 *request = MPI_REQUEST_NULL;
769 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
777 *index = MPI_UNDEFINED;
778 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
779 map = xbt_new(int, count);
780 for(i = 0; i < count; i++) {
781 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
782 xbt_dynar_push(comms, &requests[i]->action);
788 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
789 static int nsleeps = 1;
790 if(smpi_test_sleep > 0)
791 simcall_process_sleep(nsleeps*smpi_test_sleep);
793 i = simcall_comm_testany(comms);
794 // not MPI_UNDEFINED, as this is a simix return code
797 finish_wait(&requests[*index], status);
798 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
799 requests[*index] = MPI_REQUEST_NULL;
806 //all requests are null or inactive, return true
808 smpi_empty_status(status);
811 xbt_dynar_free(&comms);
816 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
819 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
822 for(i=0; i<count; i++){
823 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
824 if (smpi_mpi_test(&requests[i], pstat)!=1){
827 requests[i]=MPI_REQUEST_NULL;
830 smpi_empty_status(pstat);
832 if(status != MPI_STATUSES_IGNORE) {
839 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
841 //FIXME find another way to avoid busy waiting ?
842 // the issue here is that we have to wait on a nonexistent comm
844 smpi_mpi_iprobe(source, tag, comm, &flag, status);
845 XBT_DEBUG("Busy Waiting on probing : %d", flag);
849 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
851 MPI_Request request = build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
852 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
854 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
855 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
856 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
857 static int nsleeps = 1;
858 if(smpi_iprobe_sleep > 0)
859 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
860 // behave like a receive, but don't do it
861 smx_mailbox_t mailbox;
863 print_request("New iprobe", request);
864 // We have to test both mailboxes as we don't know if we will receive one one or another
865 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
866 mailbox = smpi_process_mailbox_small();
867 XBT_DEBUG("Trying to probe the perm recv mailbox");
868 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
871 if (request->action == nullptr){
872 mailbox = smpi_process_mailbox();
873 XBT_DEBUG("trying to probe the other mailbox");
874 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
877 if (request->action != nullptr){
878 simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
879 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
881 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
882 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
883 status->MPI_TAG = req->tag;
884 status->MPI_ERROR = MPI_SUCCESS;
885 status->count = req->real_size;
887 nsleeps = 1;//reset the number of sleeps we will do next time
893 smpi_mpi_request_free(&request);
898 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
900 print_request("Waiting", *request);
901 if ((*request)->flags & PREPARED) {
902 smpi_empty_status(status);
906 if ((*request)->action != NULL)
907 // this is not a detached send
908 simcall_comm_wait((*request)->action, -1.0);
910 finish_wait(request, status);
911 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
912 *request = MPI_REQUEST_NULL;
915 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
920 int index = MPI_UNDEFINED;
924 // Wait for a request to complete
925 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
926 map = xbt_new(int, count);
927 XBT_DEBUG("Wait for one of %d", count);
928 for(i = 0; i < count; i++) {
929 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
930 if (requests[i]->action != NULL) {
931 XBT_DEBUG("Waiting any %p ", requests[i]);
932 xbt_dynar_push(comms, &requests[i]->action);
936 //This is a finished detached request, let's return this one
937 size=0;//so we free the dynar but don't do the waitany call
939 finish_wait(&requests[i], status);//cleanup if refcount = 0
940 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
941 requests[i]=MPI_REQUEST_NULL;//set to null
947 i = simcall_comm_waitany(comms);
949 // not MPI_UNDEFINED, as this is a simix return code
952 finish_wait(&requests[index], status);
953 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
954 requests[index] = MPI_REQUEST_NULL;
958 xbt_dynar_free(&comms);
961 if (index==MPI_UNDEFINED)
962 smpi_empty_status(status);
967 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
971 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
972 int retvalue = MPI_SUCCESS;
973 //tag invalid requests in the set
974 if (status != MPI_STATUSES_IGNORE) {
975 for (c = 0; c < count; c++) {
976 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
977 smpi_empty_status(&status[c]);
978 } else if (requests[c]->src == MPI_PROC_NULL) {
979 smpi_empty_status(&status[c]);
980 status[c].MPI_SOURCE = MPI_PROC_NULL;
984 for(c = 0; c < count; c++) {
986 if (MC_is_active() || MC_record_replay_is_active()) {
987 smpi_mpi_wait(&requests[c], pstat);
990 index = smpi_mpi_waitany(count, requests, pstat);
991 if (index == MPI_UNDEFINED)
993 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
994 requests[index]=MPI_REQUEST_NULL;
996 if (status != MPI_STATUSES_IGNORE) {
997 status[index] = *pstat;
998 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
999 retvalue = MPI_ERR_IN_STATUS;
1006 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1012 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1014 for(i = 0; i < incount; i++)
1016 index=smpi_mpi_waitany(incount, requests, pstat);
1017 if(index!=MPI_UNDEFINED){
1018 indices[count] = index;
1020 if(status != MPI_STATUSES_IGNORE) {
1021 status[index] = *pstat;
1023 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1024 requests[index]=MPI_REQUEST_NULL;
1026 return MPI_UNDEFINED;
1032 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1038 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1040 for(i = 0; i < incount; i++) {
1041 if((requests[i] != MPI_REQUEST_NULL)) {
1042 if(smpi_mpi_test(&requests[i], pstat)) {
1045 if(status != MPI_STATUSES_IGNORE) {
1048 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1049 requests[i]=MPI_REQUEST_NULL;
1055 if(count_dead==incount)
1056 return MPI_UNDEFINED;
1060 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
1062 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
1065 void smpi_mpi_barrier(MPI_Comm comm)
1067 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
1070 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1071 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1073 int system_tag = COLL_TAG_GATHER;
1074 int rank, size, src, index;
1075 MPI_Aint lb = 0, recvext = 0;
1076 MPI_Request *requests;
1078 rank = smpi_comm_rank(comm);
1079 size = smpi_comm_size(comm);
1081 // Send buffer to root
1082 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1084 smpi_datatype_extent(recvtype, &lb, &recvext);
1085 // Local copy from root
1086 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext, recvcount, recvtype);
1087 // Receive buffers from senders
1088 requests = xbt_new(MPI_Request, size - 1);
1090 for(src = 0; src < size; src++) {
1092 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1093 src, system_tag, comm);
1097 // Wait for completion of irecv's.
1098 smpi_mpi_startall(size - 1, requests);
1099 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1100 for(src = 0; src < size-1; src++) {
1101 smpi_mpi_request_free(&requests[src]);
1107 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1112 int rank = smpi_comm_rank(comm);
1115 /* arbitrarily choose root as rank 0 */
1116 size = smpi_comm_size(comm);
1118 displs = xbt_new(int, size);
1119 for (i = 0; i < size; i++) {
1121 count += recvcounts[i];
1123 tmpbuf=static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1125 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1126 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1128 smpi_free_tmp_buffer(tmpbuf);
1131 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1132 MPI_Datatype recvtype, int root, MPI_Comm comm)
1134 int system_tag = COLL_TAG_GATHERV;
1135 int rank, size, src, index;
1136 MPI_Aint lb = 0, recvext = 0;
1137 MPI_Request *requests;
1139 rank = smpi_comm_rank(comm);
1140 size = smpi_comm_size(comm);
1142 // Send buffer to root
1143 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1145 smpi_datatype_extent(recvtype, &lb, &recvext);
1146 // Local copy from root
1147 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1148 recvcounts[root], recvtype);
1149 // Receive buffers from senders
1150 requests = xbt_new(MPI_Request, size - 1);
1152 for(src = 0; src < size; src++) {
1154 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1155 recvcounts[src], recvtype, src, system_tag, comm);
1159 // Wait for completion of irecv's.
1160 smpi_mpi_startall(size - 1, requests);
1161 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1162 for(src = 0; src < size-1; src++) {
1163 smpi_mpi_request_free(&requests[src]);
1169 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1170 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1172 int system_tag = COLL_TAG_ALLGATHER;
1173 int rank, size, other, index;
1174 MPI_Aint lb = 0, recvext = 0;
1175 MPI_Request *requests;
1177 rank = smpi_comm_rank(comm);
1178 size = smpi_comm_size(comm);
1179 // FIXME: check for errors
1180 smpi_datatype_extent(recvtype, &lb, &recvext);
1181 // Local copy from self
1182 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
1183 // Send/Recv buffers to/from others;
1184 requests = xbt_new(MPI_Request, 2 * (size - 1));
1186 for(other = 0; other < size; other++) {
1188 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1190 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other,
1195 // Wait for completion of all comms.
1196 smpi_mpi_startall(2 * (size - 1), requests);
1197 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1198 for(other = 0; other < 2*(size-1); other++) {
1199 smpi_mpi_request_free(&requests[other]);
1204 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1205 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1207 int system_tag = COLL_TAG_ALLGATHERV;
1208 int rank, size, other, index;
1209 MPI_Aint lb = 0, recvext = 0;
1210 MPI_Request *requests;
1212 rank = smpi_comm_rank(comm);
1213 size = smpi_comm_size(comm);
1214 smpi_datatype_extent(recvtype, &lb, &recvext);
1215 // Local copy from self
1216 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1217 // Send buffers to others;
1218 requests = xbt_new(MPI_Request, 2 * (size - 1));
1220 for(other = 0; other < size; other++) {
1223 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1225 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1226 recvtype, other, system_tag, comm);
1230 // Wait for completion of all comms.
1231 smpi_mpi_startall(2 * (size - 1), requests);
1232 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1233 for(other = 0; other < 2*(size-1); other++) {
1234 smpi_mpi_request_free(&requests[other]);
1239 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1240 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1242 int system_tag = COLL_TAG_SCATTER;
1243 int rank, size, dst, index;
1244 MPI_Aint lb = 0, sendext = 0;
1245 MPI_Request *requests;
1247 rank = smpi_comm_rank(comm);
1248 size = smpi_comm_size(comm);
1250 // Recv buffer from root
1251 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1253 smpi_datatype_extent(sendtype, &lb, &sendext);
1254 // Local copy from root
1255 if(recvbuf!=MPI_IN_PLACE){
1256 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1257 sendcount, sendtype, recvbuf, recvcount, recvtype);
1259 // Send buffers to receivers
1260 requests = xbt_new(MPI_Request, size - 1);
1262 for(dst = 0; dst < size; dst++) {
1264 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst,
1269 // Wait for completion of isend's.
1270 smpi_mpi_startall(size - 1, requests);
1271 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1272 for(dst = 0; dst < size-1; dst++) {
1273 smpi_mpi_request_free(&requests[dst]);
1279 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1280 MPI_Datatype recvtype, int root, MPI_Comm comm)
1282 int system_tag = COLL_TAG_SCATTERV;
1283 int rank, size, dst, index;
1284 MPI_Aint lb = 0, sendext = 0;
1285 MPI_Request *requests;
1287 rank = smpi_comm_rank(comm);
1288 size = smpi_comm_size(comm);
1290 // Recv buffer from root
1291 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1293 smpi_datatype_extent(sendtype, &lb, &sendext);
1294 // Local copy from root
1295 if(recvbuf!=MPI_IN_PLACE){
1296 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1297 sendtype, recvbuf, recvcount, recvtype);
1299 // Send buffers to receivers
1300 requests = xbt_new(MPI_Request, size - 1);
1302 for(dst = 0; dst < size; dst++) {
1304 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1305 sendtype, dst, system_tag, comm);
1309 // Wait for completion of isend's.
1310 smpi_mpi_startall(size - 1, requests);
1311 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1312 for(dst = 0; dst < size-1; dst++) {
1313 smpi_mpi_request_free(&requests[dst]);
1319 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1322 int system_tag = COLL_TAG_REDUCE;
1323 int rank, size, src, index;
1324 MPI_Aint lb = 0, dataext = 0;
1325 MPI_Request *requests;
1328 char* sendtmpbuf = static_cast<char *>(sendbuf);
1331 rank = smpi_comm_rank(comm);
1332 size = smpi_comm_size(comm);
1333 //non commutative case, use a working algo from openmpi
1334 if(!smpi_op_is_commute(op)){
1335 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1339 if( sendbuf == MPI_IN_PLACE ) {
1340 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1341 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1345 // Send buffer to root
1346 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1348 smpi_datatype_extent(datatype, &lb, &dataext);
1349 // Local copy from root
1350 if (sendtmpbuf != NULL && recvbuf != NULL)
1351 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1352 // Receive buffers from senders
1353 requests = xbt_new(MPI_Request, size - 1);
1354 tmpbufs = xbt_new(void *, size - 1);
1356 for(src = 0; src < size; src++) {
1358 if (!smpi_process_get_replaying())
1359 tmpbufs[index] = xbt_malloc(count * dataext);
1361 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1363 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1367 // Wait for completion of irecv's.
1368 smpi_mpi_startall(size - 1, requests);
1369 for(src = 0; src < size - 1; src++) {
1370 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1371 XBT_DEBUG("finished waiting any request with index %d", index);
1372 if(index == MPI_UNDEFINED) {
1375 smpi_mpi_request_free(&requests[index]);
1377 if(op) /* op can be MPI_OP_NULL that does nothing */
1378 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1380 for(index = 0; index < size - 1; index++) {
1381 smpi_free_tmp_buffer(tmpbufs[index]);
1387 if( sendbuf == MPI_IN_PLACE ) {
1388 smpi_free_tmp_buffer(sendtmpbuf);
1392 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1394 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1395 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1398 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1400 int system_tag = -888;
1401 int rank, size, other, index;
1402 MPI_Aint lb = 0, dataext = 0;
1403 MPI_Request *requests;
1406 rank = smpi_comm_rank(comm);
1407 size = smpi_comm_size(comm);
1409 smpi_datatype_extent(datatype, &lb, &dataext);
1411 // Local copy from self
1412 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1414 // Send/Recv buffers to/from others;
1415 requests = xbt_new(MPI_Request, size - 1);
1416 tmpbufs = xbt_new(void *, rank);
1418 for(other = 0; other < rank; other++) {
1419 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1420 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1423 for(other = rank + 1; other < size; other++) {
1424 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1427 // Wait for completion of all comms.
1428 smpi_mpi_startall(size - 1, requests);
1430 if(smpi_op_is_commute(op)){
1431 for(other = 0; other < size - 1; other++) {
1432 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1433 if(index == MPI_UNDEFINED) {
1437 // #Request is below rank: it's a irecv
1438 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1442 //non commutative case, wait in order
1443 for(other = 0; other < size - 1; other++) {
1444 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1446 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1450 for(index = 0; index < rank; index++) {
1451 smpi_free_tmp_buffer(tmpbufs[index]);
1453 for(index = 0; index < size-1; index++) {
1454 smpi_mpi_request_free(&requests[index]);
1460 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1462 int system_tag = -888;
1463 int rank, size, other, index;
1464 MPI_Aint lb = 0, dataext = 0;
1465 MPI_Request *requests;
1467 int recvbuf_is_empty=1;
1468 rank = smpi_comm_rank(comm);
1469 size = smpi_comm_size(comm);
1471 smpi_datatype_extent(datatype, &lb, &dataext);
1473 // Send/Recv buffers to/from others;
1474 requests = xbt_new(MPI_Request, size - 1);
1475 tmpbufs = xbt_new(void *, rank);
1477 for(other = 0; other < rank; other++) {
1478 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1480 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1483 for(other = rank + 1; other < size; other++) {
1485 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1488 // Wait for completion of all comms.
1489 smpi_mpi_startall(size - 1, requests);
1490 if(smpi_op_is_commute(op)){
1491 for(other = 0; other < size - 1; other++) {
1492 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1493 if(index == MPI_UNDEFINED) {
1497 if(recvbuf_is_empty){
1498 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1501 // #Request is below rank: it's a irecv
1502 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1506 //non commutative case, wait in order
1507 for(other = 0; other < size - 1; other++) {
1508 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1510 if(recvbuf_is_empty){
1511 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1513 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1517 for(index = 0; index < rank; index++) {
1518 smpi_free_tmp_buffer(tmpbufs[index]);
1520 for(index = 0; index < size-1; index++) {
1521 smpi_mpi_request_free(&requests[index]);