1 /* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "smpi_request.hpp"
10 #include "simgrid/Exception.hpp"
11 #include "simgrid/s4u/Exec.hpp"
12 #include "simgrid/s4u/Mutex.hpp"
13 #include "simgrid/s4u/ConditionVariable.hpp"
14 #include "smpi_comm.hpp"
15 #include "smpi_datatype.hpp"
16 #include "smpi_host.hpp"
17 #include "smpi_op.hpp"
18 #include "src/kernel/activity/CommImpl.hpp"
19 #include "src/mc/mc_replay.hpp"
20 #include "src/smpi/include/smpi_actor.hpp"
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
27 static simgrid::config::Flag<double> smpi_iprobe_sleep(
28 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
29 static simgrid::config::Flag<double> smpi_test_sleep(
30 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
32 std::vector<s_smpi_factor_t> smpi_ois_values;
34 extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
39 Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
40 unsigned flags, MPI_Op op)
41 : buf_(const_cast<void*>(buf))
44 , size_(datatype->size() * count)
54 if(op != MPI_REPLACE && op != MPI_OP_NULL)
58 detached_sender_ = nullptr;
61 unmatched_types_ = false;
64 if (flags & MPI_REQ_PERSISTENT)
76 void Request::unref(MPI_Request* request)
78 xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
80 (*request)->refcount_--;
81 if ((*request)->refcount_ < 0) {
82 (*request)->print_request("wrong refcount");
83 xbt_die("Whoops, wrong refcount");
85 if ((*request)->refcount_ == 0) {
86 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
87 ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
89 Comm::unref((*request)->comm_);
90 Datatype::unref((*request)->type_);
92 if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
93 Op::unref(&(*request)->op_);
95 (*request)->print_request("Destroying");
96 F2C::free_f((*request)->f2c_id());
98 *request = MPI_REQUEST_NULL;
100 (*request)->print_request("Decrementing");
104 bool Request::match_types(MPI_Datatype stype, MPI_Datatype rtype){
106 if ((stype == rtype) ||
107 //byte and packed always match with anything
108 (stype == MPI_PACKED || rtype == MPI_PACKED || stype == MPI_BYTE || rtype == MPI_BYTE) ||
109 //complex datatypes - we don't properly match these yet, as it would mean checking each subtype recursively.
110 (stype->flags() & DT_FLAG_DERIVED || rtype->flags() & DT_FLAG_DERIVED) ||
111 //duplicated datatypes, check if underlying is ok
112 (stype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype->duplicated_datatype(), rtype)) ||
113 (rtype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype, rtype->duplicated_datatype())))
116 XBT_WARN("Mismatched datatypes : sending %s and receiving %s", stype->name().c_str(), rtype->name().c_str());
121 bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
123 xbt_assert(sender, "Cannot match against null sender");
124 xbt_assert(receiver, "Cannot match against null receiver");
125 XBT_DEBUG("Trying to match %s of sender src %ld against %ld, tag %d against %d, id %d against %d",
126 (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
127 sender->comm_->id(), receiver->comm_->id());
129 if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED ||
130 receiver->comm_->id() == sender->comm_->id()) &&
131 ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) ||
132 receiver->src_ == sender->src_) &&
133 ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) {
134 // we match, we can transfer some values
135 if (receiver->src_ == MPI_ANY_SOURCE)
136 receiver->real_src_ = sender->src_;
137 if (receiver->tag_ == MPI_ANY_TAG)
138 receiver->real_tag_ = sender->tag_;
139 if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){
140 if (receiver->real_size_ < sender->real_size_){
141 XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
142 receiver->truncated_ = true;
143 } else if (receiver->real_size_ > sender->real_size_){
144 receiver->real_size_=sender->real_size_;
147 //0-sized datatypes/counts should not interfere and match
148 if ( sender->real_size_ != 0 && receiver->real_size_ != 0 &&
149 !match_types(sender->type_, receiver->type_))
150 receiver->unmatched_types_ = true;
151 if (sender->detached_)
152 receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
154 req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
155 XBT_DEBUG("match succeeded");
161 void Request::init_buffer(int count){
162 // FIXME Handle the case of a partial shared malloc.
163 // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
164 if (not smpi_process()->replaying() &&
165 ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED))) {
166 // This part handles the problem of non-contiguous memory
171 buf_ = xbt_malloc(count*type_->size());
172 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
173 type_->serialize(old_buf_, buf_, count);
179 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
181 auto ref = static_cast<MPI_Request>(a);
182 auto req = static_cast<MPI_Request>(b);
183 return match_common(req, req, ref);
186 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
188 auto ref = static_cast<MPI_Request>(a);
189 auto req = static_cast<MPI_Request>(b);
190 return match_common(req, ref, req);
193 void Request::print_request(const char* message) const
195 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
196 size_, src_, dst_, tag_, flags_);
199 /* factories, to hide the internal flags from the caller */
200 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
202 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
203 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
204 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
207 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
209 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
210 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
211 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
214 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
216 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
217 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
218 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
221 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
223 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
224 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
225 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
228 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
233 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
234 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
235 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
237 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
238 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
239 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
246 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
248 aid_t source = MPI_PROC_NULL;
249 if (src == MPI_ANY_SOURCE)
250 source = MPI_ANY_SOURCE;
251 else if (src != MPI_PROC_NULL)
252 source = comm->group()->actor(src);
253 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
255 simgrid::s4u::this_actor::get_pid(), tag, comm,
256 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
259 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
262 aid_t source = MPI_PROC_NULL;
263 if (src == MPI_ANY_SOURCE)
264 source = MPI_ANY_SOURCE;
265 else if (src != MPI_PROC_NULL)
266 source = comm->group()->actor(src);
269 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
270 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
271 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
274 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
275 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
276 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
281 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
283 aid_t source = MPI_PROC_NULL;
284 if (src == MPI_ANY_SOURCE)
285 source = MPI_ANY_SOURCE;
286 else if (src != MPI_PROC_NULL)
287 source = comm->group()->actor(src);
288 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
289 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
290 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
293 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
295 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
296 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
297 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
298 if(dst != MPI_PROC_NULL)
303 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
305 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
306 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
307 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
308 if(dst != MPI_PROC_NULL)
313 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
315 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
316 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
317 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
318 if(dst != MPI_PROC_NULL)
323 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
325 aid_t source = MPI_PROC_NULL;
326 if (src == MPI_ANY_SOURCE)
327 source = MPI_ANY_SOURCE;
328 else if (src != MPI_PROC_NULL)
329 source = comm->group()->actor(src);
330 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
331 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
332 if(src != MPI_PROC_NULL)
337 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
339 MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
340 int retval = wait(&request,status);
345 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
347 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
348 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
349 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
351 if(dst != MPI_PROC_NULL)
353 wait(&request, MPI_STATUS_IGNORE);
357 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
359 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
360 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
361 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
362 if(dst != MPI_PROC_NULL)
364 wait(&request, MPI_STATUS_IGNORE);
368 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
370 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
371 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
372 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
374 if(dst != MPI_PROC_NULL)
376 wait(&request,MPI_STATUS_IGNORE);
380 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
381 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
382 MPI_Comm comm, MPI_Status * status)
384 aid_t source = MPI_PROC_NULL;
385 if (src == MPI_ANY_SOURCE)
386 source = MPI_ANY_SOURCE;
387 else if (src != MPI_PROC_NULL)
388 source = comm->group()->actor(src);
389 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
391 std::array<MPI_Request, 2> requests;
392 std::array<MPI_Status, 2> stats;
393 aid_t myid = simgrid::s4u::this_actor::get_pid();
394 if ((destination == myid) && (source == myid)) {
395 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
396 if (status != MPI_STATUS_IGNORE) {
397 status->MPI_SOURCE = source;
398 status->MPI_TAG = recvtag;
399 status->MPI_ERROR = MPI_SUCCESS;
400 status->count = sendcount * sendtype->size();
404 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
405 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
406 startall(2, requests.data());
407 waitall(2, requests.data(), stats.data());
410 if(status != MPI_STATUS_IGNORE) {
411 // Copy receive status
416 void Request::start()
418 s4u::Mailbox* mailbox;
420 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
421 //reinitialize temporary buffer for persistent requests
422 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
424 init_buffer(real_size_/type_->size());
426 flags_ &= ~MPI_REQ_PREPARED;
427 flags_ &= ~MPI_REQ_FINISHED;
430 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
432 if ((flags_ & MPI_REQ_RECV) != 0) {
433 this->print_request("New recv");
435 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
437 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
438 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
441 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
442 mailbox = process->mailbox();
443 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
444 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
445 //begin with the more appropriate one : the small one.
446 mailbox = process->mailbox_small();
447 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
448 mailbox->get_cname());
449 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
451 if (action == nullptr) {
452 mailbox = process->mailbox();
453 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
454 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
455 if (action == nullptr) {
456 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
457 mailbox = process->mailbox_small();
460 XBT_DEBUG("yes there was something for us in the large mailbox");
463 mailbox = process->mailbox_small();
464 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
465 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
467 if (action == nullptr) {
468 XBT_DEBUG("No, nothing in the permanent receive mailbox");
469 mailbox = process->mailbox();
471 XBT_DEBUG("yes there was something for us in the small mailbox");
475 action_ = simcall_comm_irecv(
476 process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
477 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
478 XBT_DEBUG("recv simcall posted");
480 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
482 } else { /* the RECV flag was not set, so this is a send */
483 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
484 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
485 if (TRACE_smpi_view_internals())
486 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
487 this->print_request("New send");
490 if ((flags_ & MPI_REQ_SSEND) == 0 &&
491 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
492 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
493 void *oldbuf = nullptr;
495 XBT_DEBUG("Send request %p is detached", this);
497 if (not(type_->flags() & DT_FLAG_DERIVED)) {
499 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
500 if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
501 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
503 //we need this temporary buffer even for bsend, as it will be released in the copy callback and we don't have a way to differentiate it
504 //so actually ... don't use manually attached buffer space.
505 buf = xbt_malloc(size_);
506 memcpy(buf,oldbuf,size_);
507 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
512 //if we are giving back the control to the user without waiting for completion, we have to inject timings
513 double sleeptime = 0.0;
514 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
515 // isend and send timings may be different
516 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
517 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
518 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
522 simgrid::s4u::this_actor::sleep_for(sleeptime);
523 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
526 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
528 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
531 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
532 mailbox = process->mailbox();
533 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
534 mailbox = process->mailbox();
535 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
536 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
537 if (action == nullptr) {
538 if ((flags_ & MPI_REQ_SSEND) == 0) {
539 mailbox = process->mailbox_small();
540 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
541 mailbox->get_cname());
543 mailbox = process->mailbox_small();
544 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
545 mailbox->get_cname());
546 action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
547 if (action == nullptr) {
548 XBT_DEBUG("No, we are first, send to large mailbox");
549 mailbox = process->mailbox();
553 XBT_DEBUG("Yes there was something for us in the large mailbox");
556 mailbox = process->mailbox();
557 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
560 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
561 action_ = simcall_comm_isend(
562 simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_,
564 &xbt_free_f, // how to free the userdata if a detached send fails
565 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
566 // detach if msg size < eager/rdv switch limit
568 XBT_DEBUG("send simcall posted");
570 /* FIXME: detached sends are not traceable (action_ == nullptr) */
571 if (action_ != nullptr) {
572 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
573 smpi_process()->get_tracing_category());
576 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
581 void Request::startall(int count, MPI_Request * requests)
583 if(requests== nullptr)
586 for(int i = 0; i < count; i++) {
587 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
588 requests[i]->start();
592 void Request::cancel()
594 this->flags_ |= MPI_REQ_CANCELLED;
595 if (this->action_ != nullptr)
596 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
599 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
600 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
601 // to avoid deadlocks if used as a break condition, such as
602 // while (MPI_Test(request, flag, status) && flag) dostuff...
603 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
604 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
605 xbt_assert(*request != MPI_REQUEST_NULL);
607 static int nsleeps = 1;
608 int ret = MPI_SUCCESS;
610 if(smpi_test_sleep > 0)
611 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
613 Status::empty(status);
616 if ((*request)->flags_ & MPI_REQ_NBC){
617 *flag = finish_nbc_requests(request, 1);
620 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
621 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
623 *flag = simcall_comm_test((*request)->action_.get());
624 } catch (const Exception&) {
629 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
632 finish_wait(request, status); // may invalidate *request
633 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
634 MPI_Status tmp_status;
635 MPI_Status* mystatus;
636 if (status == MPI_STATUS_IGNORE) {
637 mystatus = &tmp_status;
638 Status::empty(mystatus);
642 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
644 nsleeps=1;//reset the number of sleeps we will do next time
645 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
646 *request = MPI_REQUEST_NULL;
647 } else if (smpi_cfg_grow_injected_times()) {
654 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
660 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
663 for (int i = 0; i < incount; i++) {
664 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
665 int ret = test(&requests[i], pstat, &flag);
670 if (status != MPI_STATUSES_IGNORE)
671 status[*count] = *pstat;
673 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
674 requests[i] = MPI_REQUEST_NULL;
680 if(count_dead==incount)*count=MPI_UNDEFINED;
682 return MPI_ERR_IN_STATUS;
687 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
689 std::vector<simgrid::kernel::activity::CommImpl*> comms;
690 comms.reserve(count);
693 int ret = MPI_SUCCESS;
694 *index = MPI_UNDEFINED;
696 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
697 for (int i = 0; i < count; i++) {
698 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
699 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
703 if (not map.empty()) {
704 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
705 static int nsleeps = 1;
706 if(smpi_test_sleep > 0)
707 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
710 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
711 } catch (const Exception&) {
712 XBT_DEBUG("Exception in testany");
716 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
718 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
719 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
722 finish_wait(&requests[*index],status);
723 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
724 MPI_Status tmp_status;
725 MPI_Status* mystatus;
726 if (status == MPI_STATUS_IGNORE) {
727 mystatus = &tmp_status;
728 Status::empty(mystatus);
732 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
735 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
736 *flag = finish_nbc_requests(&requests[*index] , 1);
739 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
740 requests[*index] = MPI_REQUEST_NULL;
741 XBT_DEBUG("Testany - returning with index %d", *index);
749 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
750 //all requests are null or inactive, return true
752 *index = MPI_UNDEFINED;
753 Status::empty(status);
759 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
762 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
766 for(int i=0; i<count; i++){
767 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
768 int ret = test(&requests[i], pstat, &flag);
774 if (ret != MPI_SUCCESS)
777 Status::empty(pstat);
779 if(status != MPI_STATUSES_IGNORE) {
784 return MPI_ERR_IN_STATUS;
789 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
791 //FIXME find another way to avoid busy waiting ?
792 // the issue here is that we have to wait on a nonexistent comm
794 iprobe(source, tag, comm, &flag, status);
795 XBT_DEBUG("Busy Waiting on probing : %d", flag);
799 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
800 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
801 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
802 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
803 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
804 static int nsleeps = 1;
805 double speed = s4u::this_actor::get_host()->get_speed();
806 double maxrate = smpi_cfg_iprobe_cpu_usage();
808 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
809 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
810 if (smpi_iprobe_sleep > 0) {
811 /** Compute the number of flops we will sleep **/
812 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
813 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
815 /* Not the entire CPU can be used when iprobing: This is important for
816 * the energy consumption caused by polling with iprobes.
817 * Note also that the number of flops that was
818 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
820 ->set_bound(maxrate*speed)
824 // behave like a receive, but don't do it
825 s4u::Mailbox* mailbox;
827 request->print_request("New iprobe");
828 // We have to test both mailboxes as we don't know if we will receive one or another
829 if (smpi_cfg_async_small_thresh() > 0) {
830 mailbox = smpi_process()->mailbox_small();
831 XBT_DEBUG("Trying to probe the perm recv mailbox");
832 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
835 if (request->action_ == nullptr){
836 mailbox = smpi_process()->mailbox();
837 XBT_DEBUG("trying to probe the other mailbox");
838 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
841 if (request->action_ != nullptr){
842 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
843 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
845 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
846 status->MPI_SOURCE = comm->group()->rank(req->src_);
847 status->MPI_TAG = req->tag_;
848 status->MPI_ERROR = MPI_SUCCESS;
849 status->count = req->real_size_;
851 nsleeps = 1;//reset the number of sleeps we will do next time
855 if (smpi_cfg_grow_injected_times())
859 xbt_assert(request == MPI_REQUEST_NULL);
862 int Request::finish_nbc_requests(MPI_Request* request, int test){
866 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
868 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
871 xbt_die("Failure when waiting on non blocking collective sub-requests");
873 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
874 for(auto& req: (*request)->nbc_requests_){
875 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
876 void * buf=req->buf_;
877 if((*request)->type_->flags() & DT_FLAG_DERIVED)
879 if(req->flags_ & MPI_REQ_RECV ){
880 if((*request)->op_!=MPI_OP_NULL){
881 int count=(*request)->size_/ (*request)->type_->size();
882 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
884 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
887 if(req!=MPI_REQUEST_NULL)
888 Request::unref(&req);
890 (*request)->nbc_requests_.clear();
895 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
897 MPI_Request req = *request;
898 Status::empty(status);
899 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
900 if (status!=MPI_STATUS_IGNORE)
902 if(req->detached_sender_ != nullptr)
903 unref(&(req->detached_sender_));
908 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
909 if (status != MPI_STATUS_IGNORE) {
910 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
911 Status::empty(status);
912 status->MPI_SOURCE = MPI_PROC_NULL;
914 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
915 status->MPI_SOURCE = req->comm_->group()->rank(src);
916 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
917 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
919 // this handles the case were size in receive differs from size in send
920 status->count = req->real_size_;
922 //detached send will be finished at the other end
923 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
924 req->print_request("Finishing");
925 MPI_Datatype datatype = req->type_;
927 // FIXME Handle the case of a partial shared malloc.
928 if (not smpi_process()->replaying() &&
929 (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
930 if (smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
931 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
933 if(datatype->flags() & DT_FLAG_DERIVED){
934 // This part handles the problem of non-contiguous memory the unserialization at the reception
935 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
936 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
939 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
940 if (datatype->size() != 0) {
941 int n = req->real_size_ / datatype->size();
942 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
951 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
952 aid_t rank = simgrid::s4u::this_actor::get_pid();
953 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
954 TRACE_smpi_recv(src_traced, rank,req->tag_);
956 if(req->detached_sender_ != nullptr){
957 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
959 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
960 if (sleeptime > 0.0) {
961 simgrid::s4u::this_actor::sleep_for(sleeptime);
962 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
964 unref(&(req->detached_sender_));
966 if (req->flags_ & MPI_REQ_PERSISTENT)
967 req->action_ = nullptr;
968 req->flags_ |= MPI_REQ_FINISHED;
970 if (req->truncated_ || req->unmatched_types_) {
971 char error_string[MPI_MAX_ERROR_STRING];
975 errkind = MPI_ERR_TRUNCATE;
977 errkind = MPI_ERR_TYPE;
978 PMPI_Error_string(errkind, error_string, &error_size);
979 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
980 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
981 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
982 else if (err == MPI_ERRORS_ARE_FATAL)
983 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
985 err->call((req->comm_), errkind);
986 if (err != MPI_ERRHANDLER_NULL)
987 simgrid::smpi::Errhandler::unref(err);
988 MC_assert(not MC_is_active()); /* Only fail in MC mode */
990 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
994 int Request::wait(MPI_Request * request, MPI_Status * status)
996 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
997 xbt_assert(*request != MPI_REQUEST_NULL);
1001 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
1002 if (status != MPI_STATUS_IGNORE) {
1003 Status::empty(status);
1004 status->MPI_SOURCE = MPI_PROC_NULL;
1006 (*request)=MPI_REQUEST_NULL;
1010 (*request)->print_request("Waiting");
1011 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1012 Status::empty(status);
1016 if ((*request)->action_ != nullptr){
1018 // this is not a detached send
1019 simcall_comm_wait((*request)->action_.get(), -1.0);
1020 } catch (const CancelException&) {
1021 XBT_VERB("Request cancelled");
1025 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1026 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1027 ((*request)->generalized_funcs)->mutex->lock();
1028 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1029 ((*request)->generalized_funcs)->mutex->unlock();
1031 MPI_Status tmp_status;
1032 MPI_Status* mystatus;
1033 if (status == MPI_STATUS_IGNORE) {
1034 mystatus = &tmp_status;
1035 Status::empty(mystatus);
1039 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1042 if ((*request)->truncated_)
1043 ret = MPI_ERR_TRUNCATE;
1045 if ((*request)->flags_ & MPI_REQ_NBC)
1046 finish_nbc_requests(request, 0);
1048 finish_wait(request, status); // may invalidate *request
1049 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1050 *request = MPI_REQUEST_NULL;
1054 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1056 int index = MPI_UNDEFINED;
1059 // Wait for a request to complete
1060 std::vector<simgrid::kernel::activity::CommImpl*> comms;
1061 std::vector<int> map;
1062 XBT_DEBUG("Wait for one of %d", count);
1063 for(int i = 0; i < count; i++) {
1064 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1065 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1066 if (requests[i]->action_ != nullptr) {
1067 XBT_DEBUG("Waiting any %p ", requests[i]);
1068 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
1071 // This is a finished detached request, let's return this one
1072 comms.clear(); // don't do the waitany call afterwards
1074 if (requests[index]->flags_ & MPI_REQ_NBC)
1075 finish_nbc_requests(&requests[index], 0);
1076 finish_wait(&requests[i], status); // cleanup if refcount = 0
1077 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1078 requests[i] = MPI_REQUEST_NULL; // set to null
1083 if (not comms.empty()) {
1084 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1087 i = simcall_comm_waitany(comms.data(), comms.size(), -1);
1088 } catch (const CancelException&) {
1089 XBT_INFO("request cancelled");
1093 // not MPI_UNDEFINED, as this is a simix return code
1096 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1097 if ((requests[index] == MPI_REQUEST_NULL) ||
1098 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1099 finish_wait(&requests[index],status);
1100 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1101 requests[index] = MPI_REQUEST_NULL;
1108 if (index==MPI_UNDEFINED)
1109 Status::empty(status);
1114 static int sort_accumulates(const Request* a, const Request* b)
1116 return (a->tag() > b->tag());
1119 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1121 std::vector<MPI_Request> accumulates;
1124 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1125 int retvalue = MPI_SUCCESS;
1126 //tag invalid requests in the set
1127 if (status != MPI_STATUSES_IGNORE) {
1128 for (int c = 0; c < count; c++) {
1129 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1130 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1131 Status::empty(&status[c]);
1132 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1133 Status::empty(&status[c]);
1134 status[c].MPI_SOURCE = MPI_PROC_NULL;
1138 for (int c = 0; c < count; c++) {
1139 if (MC_is_active() || MC_record_replay_is_active()) {
1140 wait(&requests[c],pstat);
1143 index = waitany(count, requests, pstat);
1145 if (index == MPI_UNDEFINED)
1148 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1149 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1150 accumulates.push_back(requests[index]);
1151 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1152 requests[index] = MPI_REQUEST_NULL;
1154 if (status != MPI_STATUSES_IGNORE) {
1155 status[index] = *pstat;
1156 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1157 retvalue = MPI_ERR_IN_STATUS;
1161 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1162 for (auto& req : accumulates)
1163 finish_wait(&req, status);
1168 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1174 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1175 index = waitany(incount, requests, pstat);
1176 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1177 if(status != MPI_STATUSES_IGNORE) {
1178 status[count] = *pstat;
1180 indices[count] = index;
1182 for (int i = 0; i < incount; i++) {
1183 if (i!=index && requests[i] != MPI_REQUEST_NULL
1184 && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1185 test(&requests[i], pstat,&flag);
1188 if(status != MPI_STATUSES_IGNORE) {
1189 status[count] = *pstat;
1191 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1192 requests[i]=MPI_REQUEST_NULL;
1200 MPI_Request Request::f2c(int id)
1202 if(id==MPI_FORTRAN_REQUEST_NULL)
1203 return MPI_REQUEST_NULL;
1204 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1207 void Request::free_f(int id)
1209 if (id != MPI_FORTRAN_REQUEST_NULL) {
1210 F2C::lookup()->erase(id);
1214 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1218 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1219 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1223 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1229 if(req != MPI_REQUEST_NULL &&
1230 status != MPI_STATUS_IGNORE) {
1231 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1232 status->MPI_SOURCE = req->comm_->group()->rank(src);
1233 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1234 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1235 status->count = req->real_size_;
1240 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1241 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1243 *request = new Request();
1244 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1245 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1246 (*request)->refcount_ = 1;
1247 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1248 ((*request)->generalized_funcs)->query_fn=query_fn;
1249 ((*request)->generalized_funcs)->free_fn=free_fn;
1250 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1251 ((*request)->generalized_funcs)->extra_state=extra_state;
1252 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1253 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1257 int Request::grequest_complete(MPI_Request request)
1259 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1260 return MPI_ERR_REQUEST;
1261 request->generalized_funcs->mutex->lock();
1262 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1263 request->generalized_funcs->cond->notify_one();
1264 request->generalized_funcs->mutex->unlock();
1268 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1269 if (not reqs.empty()) {
1270 nbc_requests_ = reqs;
1271 Request::startall(reqs.size(), reqs.data());
1275 std::vector<MPI_Request> Request::get_nbc_requests() const
1277 return nbc_requests_;