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>
8 #include <boost/tokenizer.hpp>
12 #include "xbt/virtu.h"
14 #include "src/mc/mc_replay.h"
15 #include "xbt/replay.h"
17 #include "src/simix/smx_private.h"
18 #include "surf/surf.h"
19 #include "simgrid/sg_config.h"
20 #include "colls/colls.h"
22 #include "src/simix/SynchroComm.hpp"
24 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
26 static int match_recv(void* a, void* b, smx_synchro_t ignored) {
27 MPI_Request ref = static_cast<MPI_Request>(a);
28 MPI_Request req = static_cast<MPI_Request>(b);
29 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
31 xbt_assert(ref, "Cannot match recv against null reference");
32 xbt_assert(req, "Cannot match recv against null request");
33 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
34 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
35 //we match, we can transfer some values
36 if(ref->src == MPI_ANY_SOURCE)
37 ref->real_src = req->src;
38 if(ref->tag == MPI_ANY_TAG)
39 ref->real_tag = req->tag;
40 if(ref->real_size < req->real_size)
43 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
44 XBT_DEBUG("match succeeded");
49 static int match_send(void* a, void* b,smx_synchro_t ignored) {
50 MPI_Request ref = static_cast<MPI_Request>(a);
51 MPI_Request req = static_cast<MPI_Request>(b);
52 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
53 xbt_assert(ref, "Cannot match send against null reference");
54 xbt_assert(req, "Cannot match send against null request");
56 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
57 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
59 if(req->src == MPI_ANY_SOURCE)
60 req->real_src = ref->src;
61 if(req->tag == MPI_ANY_TAG)
62 req->real_tag = ref->tag;
63 if(req->real_size < ref->real_size)
66 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
67 XBT_DEBUG("match succeeded");
72 // Methods used to parse and store the values for timing injections in smpi
73 // These are taken from surf/network.c and generalized to have more values for each factor
74 typedef struct s_smpi_factor_multival *smpi_os_factor_multival_t;
75 typedef struct s_smpi_factor_multival { // FIXME: this should be merged (deduplicated) with s_smpi_factor defined in network_smpi.c
77 std::vector<double> values; /** We allocate arbitrarily 4 elements **/
78 } s_smpi_factor_multival_t;
80 std::vector<s_smpi_factor_multival_t> smpi_os_values;
81 std::vector<s_smpi_factor_multival_t> smpi_or_values;
82 std::vector<s_smpi_factor_multival_t> smpi_ois_values;
84 static simgrid::config::Flag<double> smpi_wtime_sleep(
85 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
86 static simgrid::config::Flag<double> smpi_init_sleep(
87 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
88 static simgrid::config::Flag<double> smpi_iprobe_sleep(
89 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
90 static simgrid::config::Flag<double> smpi_test_sleep(
91 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
93 static int factor_cmp(const s_smpi_factor_multival_t& pa, const s_smpi_factor_multival_t& pb)
95 return (pa.factor > pb.factor) ? 1 :
96 (pa.factor < pb.factor) ? -1 : 0;
99 static std::vector<s_smpi_factor_multival_t> parse_factor(const char *smpi_coef_string)
101 std::vector<s_smpi_factor_multival_t> smpi_factor;
102 s_smpi_factor_multival_t fact;
104 /** Setup the tokenizer that parses the string **/
105 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
106 boost::char_separator<char> sep(";");
107 boost::char_separator<char> factor_separator(":");
108 std::string tmp_string(smpi_coef_string);
109 Tokenizer tokens(tmp_string, sep);
112 * Iterate over patterns like A:B:C:D;E:F;G:H
113 * These will be broken down into:
118 for (Tokenizer::iterator token_iter = tokens.begin();
119 token_iter != tokens.end(); token_iter++) {
120 Tokenizer factor_values(*token_iter, factor_separator);
122 if (factor_values.begin() == factor_values.end()) {
123 xbt_die("Malformed radical for smpi factor: '%s'", smpi_coef_string);
125 unsigned int iteration = 0;
126 for (Tokenizer::iterator factor_iter = factor_values.begin();
127 factor_iter != factor_values.end(); factor_iter++, iteration++) {
130 if (factor_iter == factor_values.begin()) { /* first element */
131 errmsg = bprintf("Invalid factor in chunk #%zu: %%s", smpi_factor.size()+1);
132 fact.factor = xbt_str_parse_int(factor_iter->c_str(), errmsg);
135 errmsg = bprintf("Invalid factor value %d in chunk #%zu: %%s", iteration, smpi_factor.size()+1);
136 fact.values.push_back(xbt_str_parse_double((*factor_iter).c_str(), errmsg));
141 smpi_factor.push_back(fact);
142 XBT_DEBUG("smpi_factor:\t%ld : %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]);
144 std::sort(smpi_factor.begin(), smpi_factor.end(), &factor_cmp);
145 for (auto& fact : smpi_factor) {
146 XBT_DEBUG("smpi_factor:\t%ld : %zu values, first: %f", fact.factor, smpi_factor.size() ,fact.values[0]);
152 static double smpi_os(double size)
154 if (smpi_os_values.empty()) {
155 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
158 // Iterate over all the sections that were specified and find the right
159 // value. (fact.factor represents the interval sizes; we want to find the
160 // section that has fact.factor <= size and no other such fact.factor <= size)
161 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
162 for (auto& fact : smpi_os_values) {
163 if (size <= fact.factor) { // Values already too large, use the previously
164 // computed value of current!
165 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
168 // If the next section is too large, the current section must be used.
169 // Hence, save the cost, as we might have to use it.
170 current = fact.values[0]+fact.values[1]*size;
173 XBT_DEBUG("Searching for smpi/os: %f is larger than the largest boundary, return %f", size, current);
178 static double smpi_ois(double size)
180 if (smpi_ois_values.empty()) {
181 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
184 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
185 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
186 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
187 for (auto& fact : smpi_ois_values) {
188 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
189 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
192 // If the next section is too large, the current section must be used.
193 // Hence, save the cost, as we might have to use it.
194 current = fact.values[0]+fact.values[1]*size;
197 XBT_DEBUG("Searching for smpi/ois: %f is larger than the largest boundary, return %f", size, current);
202 static double smpi_or(double size)
204 if (smpi_or_values.empty()) {
205 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
208 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
209 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
210 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
211 for (auto fact : smpi_or_values) {
212 if (size <= fact.factor) { // Values already too large, use the previously
213 // computed value of current!
214 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
217 // If the next section is too large, the current section must be used.
218 // Hence, save the cost, as we might have to use it.
219 current=fact.values[0]+fact.values[1]*size;
222 XBT_DEBUG("smpi_or: %f is larger than largest boundary, return %f", size, current);
227 void smpi_mpi_init() {
228 if(smpi_init_sleep > 0)
229 simcall_process_sleep(smpi_init_sleep);
232 double smpi_mpi_wtime(){
234 if (smpi_process_initialized() != 0 &&
235 smpi_process_finalized() == 0 &&
236 smpi_process_get_sampling() == 0) {
238 time = SIMIX_get_clock();
239 // to avoid deadlocks if used as a break condition, such as
240 // while (MPI_Wtime(...) < time_limit) {
243 // because the time will not normally advance when only calls to MPI_Wtime
244 // are made -> deadlock (MPI_Wtime never reaches the time limit)
245 if(smpi_wtime_sleep > 0)
246 simcall_process_sleep(smpi_wtime_sleep);
249 time = SIMIX_get_clock();
254 static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
257 MPI_Request request = nullptr;
259 void *old_buf = nullptr;
261 request = xbt_new(s_smpi_mpi_request_t, 1);
263 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
265 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
266 // This part handles the problem of non-contiguous memory
268 buf = count==0 ? nullptr : xbt_malloc(count*smpi_datatype_size(datatype));
269 if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
270 subtype->serialize(old_buf, buf, count, datatype->substruct);
275 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
276 request->old_buf = old_buf;
277 request->old_type = datatype;
279 request->size = smpi_datatype_size(datatype) * count;
280 smpi_datatype_use(datatype);
284 request->comm = comm;
285 smpi_comm_use(request->comm);
286 request->action = nullptr;
287 request->flags = flags;
288 request->detached = 0;
289 request->detached_sender = nullptr;
290 request->real_src = 0;
291 request->truncated = 0;
292 request->real_size = 0;
293 request->real_tag = 0;
294 if (flags & PERSISTENT)
295 request->refcount = 1;
297 request->refcount = 0;
298 request->op = MPI_REPLACE;
305 void smpi_empty_status(MPI_Status * status)
307 if(status != MPI_STATUS_IGNORE) {
308 status->MPI_SOURCE = MPI_ANY_SOURCE;
309 status->MPI_TAG = MPI_ANY_TAG;
310 status->MPI_ERROR = MPI_SUCCESS;
315 static void smpi_mpi_request_free_voidp(void* request)
317 MPI_Request req = static_cast<MPI_Request>(request);
318 smpi_mpi_request_free(&req);
321 /* MPI Low level calls */
322 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
323 int dst, int tag, MPI_Comm comm)
325 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
326 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
327 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
331 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
332 int dst, int tag, MPI_Comm comm)
334 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
335 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
336 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
340 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
341 int src, int tag, MPI_Comm comm)
343 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
344 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
345 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
346 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
350 void smpi_mpi_start(MPI_Request request)
352 smx_mailbox_t mailbox;
354 xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
355 request->flags &= ~PREPARED;
356 request->flags &= ~FINISHED;
359 if ((request->flags & RECV) != 0) {
360 print_request("New recv", request);
362 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
364 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
365 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
366 xbt_mutex_acquire(mut);
368 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
369 mailbox = smpi_process_mailbox();
371 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) {
372 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
373 //begin with the more appropriate one : the small one.
374 mailbox = smpi_process_mailbox_small();
375 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
376 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
378 if (action == nullptr) {
379 mailbox = smpi_process_mailbox();
380 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
381 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
382 if (action == nullptr) {
383 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
384 mailbox = smpi_process_mailbox_small();
388 XBT_DEBUG("yes there was something for us in the large mailbox");
392 mailbox = smpi_process_mailbox_small();
393 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
394 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
396 if (action == nullptr) {
397 XBT_DEBUG("No, nothing in the permanent receive mailbox");
398 mailbox = smpi_process_mailbox();
401 XBT_DEBUG("yes there was something for us in the small mailbox");
405 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
406 request->real_size=request->size;
407 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
408 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
409 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
410 XBT_DEBUG("recv simcall posted");
412 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
413 xbt_mutex_release(mut);
415 else { /* the RECV flag was not set, so this is a send */
416 int receiver = request->dst;
418 int rank = request->src;
419 if (TRACE_smpi_view_internals()) {
420 TRACE_smpi_send(rank, rank, receiver,request->size);
422 print_request("New send", request);
424 void* buf = request->buf;
425 if ( (request->flags & SSEND) == 0
426 && ( (request->flags & RMA) != 0 || static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
427 void *oldbuf = nullptr;
428 request->detached = 1;
429 XBT_DEBUG("Send request %p is detached", request);
431 if(request->old_type->sizeof_substruct == 0){
432 oldbuf = request->buf;
433 if (!smpi_process_get_replaying() && oldbuf != nullptr && request->size!=0){
434 if((smpi_privatize_global_variables != 0)
435 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
436 && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
437 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
438 smpi_switch_data_segment(request->src);
440 buf = xbt_malloc(request->size);
441 memcpy(buf,oldbuf,request->size);
442 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
447 //if we are giving back the control to the user without waiting for completion, we have to inject timings
448 double sleeptime = 0.0;
449 if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
450 //isend and send timings may be different
451 sleeptime = ((request->flags & ISEND) != 0) ? smpi_ois(request->size) : smpi_os(request->size);
455 simcall_process_sleep(sleeptime);
456 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, sleeptime);
459 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
461 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
463 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
464 xbt_mutex_acquire(mut);
466 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
467 mailbox = smpi_process_remote_mailbox(receiver);
469 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
470 mailbox = smpi_process_remote_mailbox(receiver);
471 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
472 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
473 if (action == nullptr) {
474 if ((request->flags & SSEND) == 0){
475 mailbox = smpi_process_remote_mailbox_small(receiver);
476 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
479 mailbox = smpi_process_remote_mailbox_small(receiver);
480 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
481 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
482 if (action == nullptr) {
483 XBT_DEBUG("No, we are first, send to large mailbox");
484 mailbox = smpi_process_remote_mailbox(receiver);
489 XBT_DEBUG("Yes there was something for us in the large mailbox");
493 mailbox = smpi_process_remote_mailbox(receiver);
494 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
497 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
498 request->real_size=request->size;
499 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
500 buf, request->real_size, &match_send,
501 &xbt_free_f, // how to free the userdata if a detached send fails
502 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
503 : &smpi_comm_null_copy_buffer_callback, request,
504 // detach if msg size < eager/rdv switch limit
506 XBT_DEBUG("send simcall posted");
508 /* FIXME: detached sends are not traceable (request->action == nullptr) */
509 if (request->action != nullptr)
510 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
512 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
513 xbt_mutex_release(mut);
517 void smpi_mpi_startall(int count, MPI_Request * requests)
519 if(requests== nullptr)
522 for(int i = 0; i < count; i++) {
523 smpi_mpi_start(requests[i]);
527 void smpi_mpi_request_free(MPI_Request * request)
529 if((*request) != MPI_REQUEST_NULL){
530 (*request)->refcount--;
531 if((*request)->refcount<0) xbt_die("wrong refcount");
533 if((*request)->refcount==0){
534 smpi_datatype_unuse((*request)->old_type);
535 smpi_comm_unuse((*request)->comm);
536 print_request("Destroying", (*request));
538 *request = MPI_REQUEST_NULL;
540 print_request("Decrementing", (*request));
543 xbt_die("freeing an already free request");
547 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
550 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
552 request = build_request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
553 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
555 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
556 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
562 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
565 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
567 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
568 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
570 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
571 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
577 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
579 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
580 request = build_request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process_index(),
581 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
585 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
587 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
588 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
589 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
590 smpi_mpi_start(request);
594 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
596 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
597 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
598 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
599 smpi_mpi_start(request);
603 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
605 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
606 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
607 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
608 comm, PERSISTENT | RECV | PREPARED);
612 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
614 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
615 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
616 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
617 NON_PERSISTENT | RECV);
618 smpi_mpi_start(request);
622 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
624 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
625 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
626 smpi_mpi_wait(&request, status);
630 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
632 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
633 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
634 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
636 smpi_mpi_start(request);
637 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
641 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
643 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
644 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
645 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
647 smpi_mpi_start(request);
648 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
652 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
653 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
654 MPI_Comm comm, MPI_Status * status)
656 MPI_Request requests[2];
658 int myid=smpi_process_index();
659 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
660 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
663 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
664 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
665 smpi_mpi_startall(2, requests);
666 smpi_mpi_waitall(2, requests, stats);
667 smpi_mpi_request_free(&requests[0]);
668 smpi_mpi_request_free(&requests[1]);
669 if(status != MPI_STATUS_IGNORE) {
670 // Copy receive status
675 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
677 return status->count / smpi_datatype_size(datatype);
680 static void finish_wait(MPI_Request * request, MPI_Status * status)
682 MPI_Request req = *request;
683 smpi_empty_status(status);
685 if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
686 if(status != MPI_STATUS_IGNORE) {
687 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
688 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
689 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
690 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
691 // this handles the case were size in receive differs from size in send
692 status->count = req->real_size;
695 print_request("Finishing", req);
696 MPI_Datatype datatype = req->old_type;
698 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
699 if (!smpi_process_get_replaying()){
700 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
701 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
702 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
703 smpi_switch_data_segment(smpi_process_index());
707 if(datatype->sizeof_substruct != 0){
708 // This part handles the problem of non-contignous memory the unserialization at the reception
709 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
710 if(req->flags & RECV)
711 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
712 datatype->substruct, req->op);
714 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
715 int n =req->real_size/smpi_datatype_size(datatype);
716 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
722 if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
723 int rank = smpi_process_index();
724 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
725 TRACE_smpi_recv(rank, src_traced, rank);
728 if(req->detached_sender != nullptr){
730 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
731 double sleeptime = smpi_or(req->real_size);
733 simcall_process_sleep(sleeptime);
734 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size, sleeptime);
736 smpi_mpi_request_free(&(req->detached_sender));
738 if(req->flags & PERSISTENT)
739 req->action = nullptr;
740 req->flags |= FINISHED;
742 smpi_mpi_request_free(request);
745 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
746 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
748 // to avoid deadlocks if used as a break condition, such as
749 // while (MPI_Test(request, flag, status) && flag) {
751 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
752 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
753 static int nsleeps = 1;
754 if(smpi_test_sleep > 0)
755 simcall_process_sleep(nsleeps*smpi_test_sleep);
757 smpi_empty_status(status);
759 if (((*request)->flags & PREPARED) == 0) {
760 if ((*request)->action != nullptr)
761 flag = simcall_comm_test((*request)->action);
763 finish_wait(request, status);
764 nsleeps=1;//reset the number of sleeps we will do next time
765 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
766 *request = MPI_REQUEST_NULL;
774 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
782 *index = MPI_UNDEFINED;
783 comms = xbt_dynar_new(sizeof(smx_synchro_t), nullptr);
784 map = xbt_new(int, count);
785 for(i = 0; i < count; i++) {
786 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
787 xbt_dynar_push(comms, &requests[i]->action);
793 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
794 static int nsleeps = 1;
795 if(smpi_test_sleep > 0)
796 simcall_process_sleep(nsleeps*smpi_test_sleep);
798 i = simcall_comm_testany(comms);
799 // not MPI_UNDEFINED, as this is a simix return code
802 finish_wait(&requests[*index], status);
803 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
804 requests[*index] = MPI_REQUEST_NULL;
811 //all requests are null or inactive, return true
813 smpi_empty_status(status);
816 xbt_dynar_free(&comms);
821 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
824 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
827 for(i=0; i<count; i++){
828 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
829 if (smpi_mpi_test(&requests[i], pstat)!=1){
832 requests[i]=MPI_REQUEST_NULL;
835 smpi_empty_status(pstat);
837 if(status != MPI_STATUSES_IGNORE) {
844 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
846 //FIXME find another way to avoid busy waiting ?
847 // the issue here is that we have to wait on a nonexistent comm
849 smpi_mpi_iprobe(source, tag, comm, &flag, status);
850 XBT_DEBUG("Busy Waiting on probing : %d", flag);
854 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
856 MPI_Request request = build_request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
857 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
859 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
860 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
861 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
862 static int nsleeps = 1;
863 if(smpi_iprobe_sleep > 0)
864 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
865 // behave like a receive, but don't do it
866 smx_mailbox_t mailbox;
868 print_request("New iprobe", request);
869 // We have to test both mailboxes as we don't know if we will receive one one or another
870 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
871 mailbox = smpi_process_mailbox_small();
872 XBT_DEBUG("Trying to probe the perm recv mailbox");
873 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
876 if (request->action == nullptr){
877 mailbox = smpi_process_mailbox();
878 XBT_DEBUG("trying to probe the other mailbox");
879 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
882 if (request->action != nullptr){
883 simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
884 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
886 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
887 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
888 status->MPI_TAG = req->tag;
889 status->MPI_ERROR = MPI_SUCCESS;
890 status->count = req->real_size;
892 nsleeps = 1;//reset the number of sleeps we will do next time
898 smpi_mpi_request_free(&request);
903 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
905 print_request("Waiting", *request);
906 if ((*request)->flags & PREPARED) {
907 smpi_empty_status(status);
911 if ((*request)->action != nullptr)
912 // this is not a detached send
913 simcall_comm_wait((*request)->action, -1.0);
915 finish_wait(request, status);
916 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
917 *request = MPI_REQUEST_NULL;
920 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
925 int index = MPI_UNDEFINED;
929 // Wait for a request to complete
930 comms = xbt_dynar_new(sizeof(smx_synchro_t), nullptr);
931 map = xbt_new(int, count);
932 XBT_DEBUG("Wait for one of %d", count);
933 for(i = 0; i < count; i++) {
934 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
935 if (requests[i]->action != nullptr) {
936 XBT_DEBUG("Waiting any %p ", requests[i]);
937 xbt_dynar_push(comms, &requests[i]->action);
941 //This is a finished detached request, let's return this one
942 size=0;//so we free the dynar but don't do the waitany call
944 finish_wait(&requests[i], status);//cleanup if refcount = 0
945 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
946 requests[i]=MPI_REQUEST_NULL;//set to null
952 i = simcall_comm_waitany(comms);
954 // not MPI_UNDEFINED, as this is a simix return code
957 finish_wait(&requests[index], status);
958 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
959 requests[index] = MPI_REQUEST_NULL;
963 xbt_dynar_free(&comms);
966 if (index==MPI_UNDEFINED)
967 smpi_empty_status(status);
972 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
976 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
977 int retvalue = MPI_SUCCESS;
978 //tag invalid requests in the set
979 if (status != MPI_STATUSES_IGNORE) {
980 for (c = 0; c < count; c++) {
981 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
982 smpi_empty_status(&status[c]);
983 } else if (requests[c]->src == MPI_PROC_NULL) {
984 smpi_empty_status(&status[c]);
985 status[c].MPI_SOURCE = MPI_PROC_NULL;
989 for(c = 0; c < count; c++) {
991 if (MC_is_active() || MC_record_replay_is_active()) {
992 smpi_mpi_wait(&requests[c], pstat);
995 index = smpi_mpi_waitany(count, requests, pstat);
996 if (index == MPI_UNDEFINED)
998 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
999 requests[index]=MPI_REQUEST_NULL;
1001 if (status != MPI_STATUSES_IGNORE) {
1002 status[index] = *pstat;
1003 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1004 retvalue = MPI_ERR_IN_STATUS;
1011 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1017 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1019 for(i = 0; i < incount; i++)
1021 index=smpi_mpi_waitany(incount, requests, pstat);
1022 if(index!=MPI_UNDEFINED){
1023 indices[count] = index;
1025 if(status != MPI_STATUSES_IGNORE) {
1026 status[index] = *pstat;
1028 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1029 requests[index]=MPI_REQUEST_NULL;
1031 return MPI_UNDEFINED;
1037 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1043 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1045 for(i = 0; i < incount; i++) {
1046 if((requests[i] != MPI_REQUEST_NULL)) {
1047 if(smpi_mpi_test(&requests[i], pstat)) {
1050 if(status != MPI_STATUSES_IGNORE) {
1053 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1054 requests[i]=MPI_REQUEST_NULL;
1060 if(count_dead==incount)
1061 return MPI_UNDEFINED;
1065 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
1067 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
1070 void smpi_mpi_barrier(MPI_Comm comm)
1072 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
1075 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1076 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1078 int system_tag = COLL_TAG_GATHER;
1079 int rank, size, src, index;
1080 MPI_Aint lb = 0, recvext = 0;
1081 MPI_Request *requests;
1083 rank = smpi_comm_rank(comm);
1084 size = smpi_comm_size(comm);
1086 // Send buffer to root
1087 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1089 smpi_datatype_extent(recvtype, &lb, &recvext);
1090 // Local copy from root
1091 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext, recvcount, recvtype);
1092 // Receive buffers from senders
1093 requests = xbt_new(MPI_Request, size - 1);
1095 for(src = 0; src < size; src++) {
1097 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1098 src, system_tag, comm);
1102 // Wait for completion of irecv's.
1103 smpi_mpi_startall(size - 1, requests);
1104 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1105 for(src = 0; src < size-1; src++) {
1106 smpi_mpi_request_free(&requests[src]);
1112 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1117 int rank = smpi_comm_rank(comm);
1120 /* arbitrarily choose root as rank 0 */
1121 size = smpi_comm_size(comm);
1123 displs = xbt_new(int, size);
1124 for (i = 0; i < size; i++) {
1126 count += recvcounts[i];
1128 tmpbuf=static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1130 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1131 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1133 smpi_free_tmp_buffer(tmpbuf);
1136 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1137 MPI_Datatype recvtype, int root, MPI_Comm comm)
1139 int system_tag = COLL_TAG_GATHERV;
1140 int rank, size, src, index;
1141 MPI_Aint lb = 0, recvext = 0;
1142 MPI_Request *requests;
1144 rank = smpi_comm_rank(comm);
1145 size = smpi_comm_size(comm);
1147 // Send buffer to root
1148 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1150 smpi_datatype_extent(recvtype, &lb, &recvext);
1151 // Local copy from root
1152 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1153 recvcounts[root], recvtype);
1154 // Receive buffers from senders
1155 requests = xbt_new(MPI_Request, size - 1);
1157 for(src = 0; src < size; src++) {
1159 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1160 recvcounts[src], recvtype, src, system_tag, comm);
1164 // Wait for completion of irecv's.
1165 smpi_mpi_startall(size - 1, requests);
1166 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1167 for(src = 0; src < size-1; src++) {
1168 smpi_mpi_request_free(&requests[src]);
1174 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1175 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1177 int system_tag = COLL_TAG_ALLGATHER;
1178 int rank, size, other, index;
1179 MPI_Aint lb = 0, recvext = 0;
1180 MPI_Request *requests;
1182 rank = smpi_comm_rank(comm);
1183 size = smpi_comm_size(comm);
1184 // FIXME: check for errors
1185 smpi_datatype_extent(recvtype, &lb, &recvext);
1186 // Local copy from self
1187 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
1188 // Send/Recv buffers to/from others;
1189 requests = xbt_new(MPI_Request, 2 * (size - 1));
1191 for(other = 0; other < size; other++) {
1193 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1195 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other,
1200 // Wait for completion of all comms.
1201 smpi_mpi_startall(2 * (size - 1), requests);
1202 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1203 for(other = 0; other < 2*(size-1); other++) {
1204 smpi_mpi_request_free(&requests[other]);
1209 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1210 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1212 int system_tag = COLL_TAG_ALLGATHERV;
1213 int rank, size, other, index;
1214 MPI_Aint lb = 0, recvext = 0;
1215 MPI_Request *requests;
1217 rank = smpi_comm_rank(comm);
1218 size = smpi_comm_size(comm);
1219 smpi_datatype_extent(recvtype, &lb, &recvext);
1220 // Local copy from self
1221 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1222 // Send buffers to others;
1223 requests = xbt_new(MPI_Request, 2 * (size - 1));
1225 for(other = 0; other < size; other++) {
1228 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1230 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1231 recvtype, other, system_tag, comm);
1235 // Wait for completion of all comms.
1236 smpi_mpi_startall(2 * (size - 1), requests);
1237 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1238 for(other = 0; other < 2*(size-1); other++) {
1239 smpi_mpi_request_free(&requests[other]);
1244 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1245 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1247 int system_tag = COLL_TAG_SCATTER;
1248 int rank, size, dst, index;
1249 MPI_Aint lb = 0, sendext = 0;
1250 MPI_Request *requests;
1252 rank = smpi_comm_rank(comm);
1253 size = smpi_comm_size(comm);
1255 // Recv buffer from root
1256 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1258 smpi_datatype_extent(sendtype, &lb, &sendext);
1259 // Local copy from root
1260 if(recvbuf!=MPI_IN_PLACE){
1261 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1262 sendcount, sendtype, recvbuf, recvcount, recvtype);
1264 // Send buffers to receivers
1265 requests = xbt_new(MPI_Request, size - 1);
1267 for(dst = 0; dst < size; dst++) {
1269 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst,
1274 // Wait for completion of isend's.
1275 smpi_mpi_startall(size - 1, requests);
1276 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1277 for(dst = 0; dst < size-1; dst++) {
1278 smpi_mpi_request_free(&requests[dst]);
1284 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1285 MPI_Datatype recvtype, int root, MPI_Comm comm)
1287 int system_tag = COLL_TAG_SCATTERV;
1288 int rank, size, dst, index;
1289 MPI_Aint lb = 0, sendext = 0;
1290 MPI_Request *requests;
1292 rank = smpi_comm_rank(comm);
1293 size = smpi_comm_size(comm);
1295 // Recv buffer from root
1296 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1298 smpi_datatype_extent(sendtype, &lb, &sendext);
1299 // Local copy from root
1300 if(recvbuf!=MPI_IN_PLACE){
1301 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1302 sendtype, recvbuf, recvcount, recvtype);
1304 // Send buffers to receivers
1305 requests = xbt_new(MPI_Request, size - 1);
1307 for(dst = 0; dst < size; dst++) {
1309 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1310 sendtype, dst, system_tag, comm);
1314 // Wait for completion of isend's.
1315 smpi_mpi_startall(size - 1, requests);
1316 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1317 for(dst = 0; dst < size-1; dst++) {
1318 smpi_mpi_request_free(&requests[dst]);
1324 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1327 int system_tag = COLL_TAG_REDUCE;
1328 int rank, size, src, index;
1329 MPI_Aint lb = 0, dataext = 0;
1330 MPI_Request *requests;
1333 char* sendtmpbuf = static_cast<char *>(sendbuf);
1336 rank = smpi_comm_rank(comm);
1337 size = smpi_comm_size(comm);
1338 //non commutative case, use a working algo from openmpi
1339 if(!smpi_op_is_commute(op)){
1340 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1344 if( sendbuf == MPI_IN_PLACE ) {
1345 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1346 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1350 // Send buffer to root
1351 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1353 smpi_datatype_extent(datatype, &lb, &dataext);
1354 // Local copy from root
1355 if (sendtmpbuf != nullptr && recvbuf != nullptr)
1356 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1357 // Receive buffers from senders
1358 requests = xbt_new(MPI_Request, size - 1);
1359 tmpbufs = xbt_new(void *, size - 1);
1361 for(src = 0; src < size; src++) {
1363 if (!smpi_process_get_replaying())
1364 tmpbufs[index] = xbt_malloc(count * dataext);
1366 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1368 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1372 // Wait for completion of irecv's.
1373 smpi_mpi_startall(size - 1, requests);
1374 for(src = 0; src < size - 1; src++) {
1375 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1376 XBT_DEBUG("finished waiting any request with index %d", index);
1377 if(index == MPI_UNDEFINED) {
1380 smpi_mpi_request_free(&requests[index]);
1382 if(op) /* op can be MPI_OP_NULL that does nothing */
1383 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1385 for(index = 0; index < size - 1; index++) {
1386 smpi_free_tmp_buffer(tmpbufs[index]);
1392 if( sendbuf == MPI_IN_PLACE ) {
1393 smpi_free_tmp_buffer(sendtmpbuf);
1397 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1399 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1400 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1403 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1405 int system_tag = -888;
1406 int rank, size, other, index;
1407 MPI_Aint lb = 0, dataext = 0;
1408 MPI_Request *requests;
1411 rank = smpi_comm_rank(comm);
1412 size = smpi_comm_size(comm);
1414 smpi_datatype_extent(datatype, &lb, &dataext);
1416 // Local copy from self
1417 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1419 // Send/Recv buffers to/from others;
1420 requests = xbt_new(MPI_Request, size - 1);
1421 tmpbufs = xbt_new(void *, rank);
1423 for(other = 0; other < rank; other++) {
1424 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1425 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1428 for(other = rank + 1; other < size; other++) {
1429 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1432 // Wait for completion of all comms.
1433 smpi_mpi_startall(size - 1, requests);
1435 if(smpi_op_is_commute(op)){
1436 for(other = 0; other < size - 1; other++) {
1437 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1438 if(index == MPI_UNDEFINED) {
1442 // #Request is below rank: it's a irecv
1443 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1447 //non commutative case, wait in order
1448 for(other = 0; other < size - 1; other++) {
1449 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1451 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1455 for(index = 0; index < rank; index++) {
1456 smpi_free_tmp_buffer(tmpbufs[index]);
1458 for(index = 0; index < size-1; index++) {
1459 smpi_mpi_request_free(&requests[index]);
1465 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1467 int system_tag = -888;
1468 int rank, size, other, index;
1469 MPI_Aint lb = 0, dataext = 0;
1470 MPI_Request *requests;
1472 int recvbuf_is_empty=1;
1473 rank = smpi_comm_rank(comm);
1474 size = smpi_comm_size(comm);
1476 smpi_datatype_extent(datatype, &lb, &dataext);
1478 // Send/Recv buffers to/from others;
1479 requests = xbt_new(MPI_Request, size - 1);
1480 tmpbufs = xbt_new(void *, rank);
1482 for(other = 0; other < rank; other++) {
1483 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1485 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1488 for(other = rank + 1; other < size; other++) {
1490 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1493 // Wait for completion of all comms.
1494 smpi_mpi_startall(size - 1, requests);
1495 if(smpi_op_is_commute(op)){
1496 for(other = 0; other < size - 1; other++) {
1497 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1498 if(index == MPI_UNDEFINED) {
1502 if(recvbuf_is_empty){
1503 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1506 // #Request is below rank: it's a irecv
1507 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1511 //non commutative case, wait in order
1512 for(other = 0; other < size - 1; other++) {
1513 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1515 if(recvbuf_is_empty){
1516 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1518 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1522 for(index = 0; index < rank; index++) {
1523 smpi_free_tmp_buffer(tmpbufs[index]);
1525 for(index = 0; index < size-1; index++) {
1526 smpi_mpi_request_free(&requests[index]);