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 ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED)) {
165 // This part handles the problem of non-contiguous memory
170 buf_ = xbt_malloc(count*type_->size());
171 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
172 type_->serialize(old_buf_, buf_, count);
178 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
180 auto ref = static_cast<MPI_Request>(a);
181 auto req = static_cast<MPI_Request>(b);
182 return match_common(req, req, ref);
185 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
187 auto ref = static_cast<MPI_Request>(a);
188 auto req = static_cast<MPI_Request>(b);
189 return match_common(req, ref, req);
192 void Request::print_request(const char* message) const
194 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
195 size_, src_, dst_, tag_, flags_);
198 /* factories, to hide the internal flags from the caller */
199 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
201 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
202 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
203 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
206 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
208 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
209 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
210 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
213 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
215 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
216 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
217 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
220 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
222 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
223 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
224 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
227 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
232 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
233 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
234 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
236 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
237 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
238 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
245 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
247 aid_t source = MPI_PROC_NULL;
248 if (src == MPI_ANY_SOURCE)
249 source = MPI_ANY_SOURCE;
250 else if (src != MPI_PROC_NULL)
251 source = comm->group()->actor(src);
252 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
254 simgrid::s4u::this_actor::get_pid(), tag, comm,
255 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
258 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
261 aid_t source = MPI_PROC_NULL;
262 if (src == MPI_ANY_SOURCE)
263 source = MPI_ANY_SOURCE;
264 else if (src != MPI_PROC_NULL)
265 source = comm->group()->actor(src);
268 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
269 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
270 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
273 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
274 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
275 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
280 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
282 aid_t source = MPI_PROC_NULL;
283 if (src == MPI_ANY_SOURCE)
284 source = MPI_ANY_SOURCE;
285 else if (src != MPI_PROC_NULL)
286 source = comm->group()->actor(src);
287 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
288 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
289 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
292 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
294 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
295 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
296 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
297 if(dst != MPI_PROC_NULL)
302 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
304 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
305 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
306 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
307 if(dst != MPI_PROC_NULL)
312 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
314 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
315 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
316 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
317 if(dst != MPI_PROC_NULL)
322 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
324 aid_t source = MPI_PROC_NULL;
325 if (src == MPI_ANY_SOURCE)
326 source = MPI_ANY_SOURCE;
327 else if (src != MPI_PROC_NULL)
328 source = comm->group()->actor(src);
329 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
330 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
331 if(src != MPI_PROC_NULL)
336 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
338 MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
339 int retval = wait(&request,status);
344 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
346 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
347 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
348 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
350 if(dst != MPI_PROC_NULL)
352 wait(&request, MPI_STATUS_IGNORE);
356 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
358 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
359 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
360 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
361 if(dst != MPI_PROC_NULL)
363 wait(&request, MPI_STATUS_IGNORE);
367 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
369 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
370 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
371 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
373 if(dst != MPI_PROC_NULL)
375 wait(&request,MPI_STATUS_IGNORE);
379 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
380 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
381 MPI_Comm comm, MPI_Status * status)
383 aid_t source = MPI_PROC_NULL;
384 if (src == MPI_ANY_SOURCE)
385 source = MPI_ANY_SOURCE;
386 else if (src != MPI_PROC_NULL)
387 source = comm->group()->actor(src);
388 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
390 std::array<MPI_Request, 2> requests;
391 std::array<MPI_Status, 2> stats;
392 aid_t myid = simgrid::s4u::this_actor::get_pid();
393 if ((destination == myid) && (source == myid)) {
394 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
395 if (status != MPI_STATUS_IGNORE) {
396 status->MPI_SOURCE = source;
397 status->MPI_TAG = recvtag;
398 status->MPI_ERROR = MPI_SUCCESS;
399 status->count = sendcount * sendtype->size();
403 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
404 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
405 startall(2, requests.data());
406 waitall(2, requests.data(), stats.data());
409 if(status != MPI_STATUS_IGNORE) {
410 // Copy receive status
415 void Request::start()
417 s4u::Mailbox* mailbox;
419 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
420 //reinitialize temporary buffer for persistent requests
421 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
423 init_buffer(real_size_/type_->size());
425 flags_ &= ~MPI_REQ_PREPARED;
426 flags_ &= ~MPI_REQ_FINISHED;
429 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
431 if ((flags_ & MPI_REQ_RECV) != 0) {
432 this->print_request("New recv");
434 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
436 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
437 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
440 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
441 mailbox = process->mailbox();
442 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
443 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
444 //begin with the more appropriate one : the small one.
445 mailbox = process->mailbox_small();
446 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
447 mailbox->get_cname());
448 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
450 if (action == nullptr) {
451 mailbox = process->mailbox();
452 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
453 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
454 if (action == nullptr) {
455 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
456 mailbox = process->mailbox_small();
459 XBT_DEBUG("yes there was something for us in the large mailbox");
462 mailbox = process->mailbox_small();
463 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
464 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
466 if (action == nullptr) {
467 XBT_DEBUG("No, nothing in the permanent receive mailbox");
468 mailbox = process->mailbox();
470 XBT_DEBUG("yes there was something for us in the small mailbox");
474 action_ = simcall_comm_irecv(
475 process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
476 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
477 XBT_DEBUG("recv simcall posted");
479 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
481 } else { /* the RECV flag was not set, so this is a send */
482 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
483 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
484 if (TRACE_smpi_view_internals())
485 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
486 this->print_request("New send");
489 if ((flags_ & MPI_REQ_SSEND) == 0 &&
490 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
491 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
492 void *oldbuf = nullptr;
494 XBT_DEBUG("Send request %p is detached", this);
496 if (not(type_->flags() & DT_FLAG_DERIVED)) {
498 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
499 if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
500 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
502 //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
503 //so actually ... don't use manually attached buffer space.
504 buf = xbt_malloc(size_);
505 memcpy(buf,oldbuf,size_);
506 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
511 //if we are giving back the control to the user without waiting for completion, we have to inject timings
512 double sleeptime = 0.0;
513 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
514 // isend and send timings may be different
515 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
516 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
517 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
521 simgrid::s4u::this_actor::sleep_for(sleeptime);
522 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
525 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
527 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
530 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
531 mailbox = process->mailbox();
532 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
533 mailbox = process->mailbox();
534 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
535 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
536 if (action == nullptr) {
537 if ((flags_ & MPI_REQ_SSEND) == 0) {
538 mailbox = process->mailbox_small();
539 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
540 mailbox->get_cname());
542 mailbox = process->mailbox_small();
543 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
544 mailbox->get_cname());
545 action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
546 if (action == nullptr) {
547 XBT_DEBUG("No, we are first, send to large mailbox");
548 mailbox = process->mailbox();
552 XBT_DEBUG("Yes there was something for us in the large mailbox");
555 mailbox = process->mailbox();
556 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
559 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
560 action_ = simcall_comm_isend(
561 simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_,
563 &xbt_free_f, // how to free the userdata if a detached send fails
564 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
565 // detach if msg size < eager/rdv switch limit
567 XBT_DEBUG("send simcall posted");
569 /* FIXME: detached sends are not traceable (action_ == nullptr) */
570 if (action_ != nullptr) {
571 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
572 smpi_process()->get_tracing_category());
575 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
580 void Request::startall(int count, MPI_Request * requests)
582 if(requests== nullptr)
585 for(int i = 0; i < count; i++) {
586 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
587 requests[i]->start();
591 void Request::cancel()
593 this->flags_ |= MPI_REQ_CANCELLED;
594 if (this->action_ != nullptr)
595 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
598 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
599 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
600 // to avoid deadlocks if used as a break condition, such as
601 // while (MPI_Test(request, flag, status) && flag) dostuff...
602 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
603 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
604 xbt_assert(*request != MPI_REQUEST_NULL);
606 static int nsleeps = 1;
607 int ret = MPI_SUCCESS;
609 if(smpi_test_sleep > 0)
610 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
612 Status::empty(status);
615 if ((*request)->flags_ & MPI_REQ_NBC){
616 *flag = finish_nbc_requests(request, 1);
619 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
620 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
622 *flag = simcall_comm_test((*request)->action_.get());
623 } catch (const Exception&) {
628 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
631 finish_wait(request, status); // may invalidate *request
632 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
633 MPI_Status tmp_status;
634 MPI_Status* mystatus;
635 if (status == MPI_STATUS_IGNORE) {
636 mystatus = &tmp_status;
637 Status::empty(mystatus);
641 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
643 nsleeps=1;//reset the number of sleeps we will do next time
644 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
645 *request = MPI_REQUEST_NULL;
646 } else if (smpi_cfg_grow_injected_times()) {
653 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
659 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
662 for (int i = 0; i < incount; i++) {
663 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
664 int ret = test(&requests[i], pstat, &flag);
669 if (status != MPI_STATUSES_IGNORE)
670 status[*count] = *pstat;
672 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
673 requests[i] = MPI_REQUEST_NULL;
679 if(count_dead==incount)*count=MPI_UNDEFINED;
681 return MPI_ERR_IN_STATUS;
686 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
688 std::vector<simgrid::kernel::activity::CommImpl*> comms;
689 comms.reserve(count);
692 int ret = MPI_SUCCESS;
693 *index = MPI_UNDEFINED;
695 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
696 for (int i = 0; i < count; i++) {
697 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
698 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
702 if (not map.empty()) {
703 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
704 static int nsleeps = 1;
705 if(smpi_test_sleep > 0)
706 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
709 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
710 } catch (const Exception&) {
711 XBT_DEBUG("Exception in testany");
715 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
717 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
718 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
721 finish_wait(&requests[*index],status);
722 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
723 MPI_Status tmp_status;
724 MPI_Status* mystatus;
725 if (status == MPI_STATUS_IGNORE) {
726 mystatus = &tmp_status;
727 Status::empty(mystatus);
731 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
734 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
735 *flag = finish_nbc_requests(&requests[*index] , 1);
738 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
739 requests[*index] = MPI_REQUEST_NULL;
740 XBT_DEBUG("Testany - returning with index %d", *index);
748 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
749 //all requests are null or inactive, return true
751 *index = MPI_UNDEFINED;
752 Status::empty(status);
758 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
761 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
765 for(int i=0; i<count; i++){
766 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
767 int ret = test(&requests[i], pstat, &flag);
773 if (ret != MPI_SUCCESS)
776 Status::empty(pstat);
778 if(status != MPI_STATUSES_IGNORE) {
783 return MPI_ERR_IN_STATUS;
788 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
790 //FIXME find another way to avoid busy waiting ?
791 // the issue here is that we have to wait on a nonexistent comm
793 iprobe(source, tag, comm, &flag, status);
794 XBT_DEBUG("Busy Waiting on probing : %d", flag);
798 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
799 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
800 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
801 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
802 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
803 static int nsleeps = 1;
804 double speed = s4u::this_actor::get_host()->get_speed();
805 double maxrate = smpi_cfg_iprobe_cpu_usage();
807 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
808 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
809 if (smpi_iprobe_sleep > 0) {
810 /** Compute the number of flops we will sleep **/
811 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
812 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
814 /* Not the entire CPU can be used when iprobing: This is important for
815 * the energy consumption caused by polling with iprobes.
816 * Note also that the number of flops that was
817 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
819 ->set_bound(maxrate*speed)
823 // behave like a receive, but don't do it
824 s4u::Mailbox* mailbox;
826 request->print_request("New iprobe");
827 // We have to test both mailboxes as we don't know if we will receive one or another
828 if (smpi_cfg_async_small_thresh() > 0) {
829 mailbox = smpi_process()->mailbox_small();
830 XBT_DEBUG("Trying to probe the perm recv mailbox");
831 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
834 if (request->action_ == nullptr){
835 mailbox = smpi_process()->mailbox();
836 XBT_DEBUG("trying to probe the other mailbox");
837 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
840 if (request->action_ != nullptr){
841 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
842 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
844 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
845 status->MPI_SOURCE = comm->group()->rank(req->src_);
846 status->MPI_TAG = req->tag_;
847 status->MPI_ERROR = MPI_SUCCESS;
848 status->count = req->real_size_;
850 nsleeps = 1;//reset the number of sleeps we will do next time
854 if (smpi_cfg_grow_injected_times())
858 xbt_assert(request == MPI_REQUEST_NULL);
861 int Request::finish_nbc_requests(MPI_Request* request, int test){
865 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
867 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
870 xbt_die("Failure when waiting on non blocking collective sub-requests");
872 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
873 for(auto& req: (*request)->nbc_requests_){
874 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
875 void * buf=req->buf_;
876 if((*request)->type_->flags() & DT_FLAG_DERIVED)
878 if(req->flags_ & MPI_REQ_RECV ){
879 if((*request)->op_!=MPI_OP_NULL){
880 int count=(*request)->size_/ (*request)->type_->size();
881 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
883 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
886 if(req!=MPI_REQUEST_NULL)
887 Request::unref(&req);
889 (*request)->nbc_requests_.clear();
894 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
896 MPI_Request req = *request;
897 Status::empty(status);
898 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
899 if (status!=MPI_STATUS_IGNORE)
901 if(req->detached_sender_ != nullptr)
902 unref(&(req->detached_sender_));
907 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
908 if (status != MPI_STATUS_IGNORE) {
909 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
910 Status::empty(status);
911 status->MPI_SOURCE = MPI_PROC_NULL;
913 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
914 status->MPI_SOURCE = req->comm_->group()->rank(src);
915 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
916 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
918 // this handles the case were size in receive differs from size in send
919 status->count = req->real_size_;
921 //detached send will be finished at the other end
922 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
923 req->print_request("Finishing");
924 MPI_Datatype datatype = req->type_;
926 // FIXME Handle the case of a partial shared malloc.
927 if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
928 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
929 if (not smpi_process()->replaying() && smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
930 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
932 if(datatype->flags() & DT_FLAG_DERIVED){
933 // This part handles the problem of non-contiguous memory the unserialization at the reception
934 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
935 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
938 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
939 if (datatype->size() != 0) {
940 int n = req->real_size_ / datatype->size();
941 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
950 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
951 aid_t rank = simgrid::s4u::this_actor::get_pid();
952 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
953 TRACE_smpi_recv(src_traced, rank,req->tag_);
955 if(req->detached_sender_ != nullptr){
956 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
958 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
959 if (sleeptime > 0.0) {
960 simgrid::s4u::this_actor::sleep_for(sleeptime);
961 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
963 unref(&(req->detached_sender_));
965 if (req->flags_ & MPI_REQ_PERSISTENT)
966 req->action_ = nullptr;
967 req->flags_ |= MPI_REQ_FINISHED;
969 if (req->truncated_ || req->unmatched_types_) {
970 char error_string[MPI_MAX_ERROR_STRING];
974 errkind = MPI_ERR_TRUNCATE;
976 errkind = MPI_ERR_TYPE;
977 PMPI_Error_string(errkind, error_string, &error_size);
978 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
979 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
980 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
981 else if (err == MPI_ERRORS_ARE_FATAL)
982 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
984 err->call((req->comm_), errkind);
985 if (err != MPI_ERRHANDLER_NULL)
986 simgrid::smpi::Errhandler::unref(err);
987 MC_assert(not MC_is_active()); /* Only fail in MC mode */
989 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
993 int Request::wait(MPI_Request * request, MPI_Status * status)
995 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
996 xbt_assert(*request != MPI_REQUEST_NULL);
1000 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
1001 if (status != MPI_STATUS_IGNORE) {
1002 Status::empty(status);
1003 status->MPI_SOURCE = MPI_PROC_NULL;
1005 (*request)=MPI_REQUEST_NULL;
1009 (*request)->print_request("Waiting");
1010 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1011 Status::empty(status);
1015 if ((*request)->action_ != nullptr){
1017 // this is not a detached send
1018 simcall_comm_wait((*request)->action_.get(), -1.0);
1019 } catch (const CancelException&) {
1020 XBT_VERB("Request cancelled");
1024 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1025 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1026 ((*request)->generalized_funcs)->mutex->lock();
1027 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1028 ((*request)->generalized_funcs)->mutex->unlock();
1030 MPI_Status tmp_status;
1031 MPI_Status* mystatus;
1032 if (status == MPI_STATUS_IGNORE) {
1033 mystatus = &tmp_status;
1034 Status::empty(mystatus);
1038 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1041 if ((*request)->truncated_)
1042 ret = MPI_ERR_TRUNCATE;
1044 if ((*request)->flags_ & MPI_REQ_NBC)
1045 finish_nbc_requests(request, 0);
1047 finish_wait(request, status); // may invalidate *request
1048 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1049 *request = MPI_REQUEST_NULL;
1053 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1055 int index = MPI_UNDEFINED;
1058 // Wait for a request to complete
1059 std::vector<simgrid::kernel::activity::CommImpl*> comms;
1060 std::vector<int> map;
1061 XBT_DEBUG("Wait for one of %d", count);
1062 for(int i = 0; i < count; i++) {
1063 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1064 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1065 if (requests[i]->action_ != nullptr) {
1066 XBT_DEBUG("Waiting any %p ", requests[i]);
1067 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
1070 // This is a finished detached request, let's return this one
1071 comms.clear(); // don't do the waitany call afterwards
1073 if (requests[index] != MPI_REQUEST_NULL && (requests[index])->flags_ & MPI_REQ_NBC)
1074 finish_nbc_requests(&requests[index], 0);
1075 finish_wait(&requests[i], status); // cleanup if refcount = 0
1076 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1077 requests[i] = MPI_REQUEST_NULL; // set to null
1082 if (not comms.empty()) {
1083 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1086 i = simcall_comm_waitany(comms.data(), comms.size(), -1);
1087 } catch (const CancelException&) {
1088 XBT_INFO("request cancelled");
1092 // not MPI_UNDEFINED, as this is a simix return code
1095 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1096 if ((requests[index] == MPI_REQUEST_NULL) ||
1097 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1098 finish_wait(&requests[index],status);
1099 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1100 requests[index] = MPI_REQUEST_NULL;
1107 if (index==MPI_UNDEFINED)
1108 Status::empty(status);
1113 static int sort_accumulates(const Request* a, const Request* b)
1115 return (a->tag() > b->tag());
1118 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1120 std::vector<MPI_Request> accumulates;
1123 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1124 int retvalue = MPI_SUCCESS;
1125 //tag invalid requests in the set
1126 if (status != MPI_STATUSES_IGNORE) {
1127 for (int c = 0; c < count; c++) {
1128 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1129 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1130 Status::empty(&status[c]);
1131 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1132 Status::empty(&status[c]);
1133 status[c].MPI_SOURCE = MPI_PROC_NULL;
1137 for (int c = 0; c < count; c++) {
1138 if (MC_is_active() || MC_record_replay_is_active()) {
1139 wait(&requests[c],pstat);
1142 index = waitany(count, requests, pstat);
1144 if (index == MPI_UNDEFINED)
1147 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1148 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1149 accumulates.push_back(requests[index]);
1150 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1151 requests[index] = MPI_REQUEST_NULL;
1153 if (status != MPI_STATUSES_IGNORE) {
1154 status[index] = *pstat;
1155 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1156 retvalue = MPI_ERR_IN_STATUS;
1160 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1161 for (auto& req : accumulates)
1162 finish_wait(&req, status);
1167 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1173 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1174 index = waitany(incount, requests, pstat);
1175 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1176 if(status != MPI_STATUSES_IGNORE) {
1177 status[count] = *pstat;
1179 indices[count] = index;
1181 for (int i = 0; i < incount; i++) {
1182 if (i!=index && requests[i] != MPI_REQUEST_NULL
1183 && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1184 test(&requests[i], pstat,&flag);
1187 if(status != MPI_STATUSES_IGNORE) {
1188 status[count] = *pstat;
1190 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1191 requests[i]=MPI_REQUEST_NULL;
1199 MPI_Request Request::f2c(int id)
1201 if(id==MPI_FORTRAN_REQUEST_NULL)
1202 return MPI_REQUEST_NULL;
1203 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1206 void Request::free_f(int id)
1208 if (id != MPI_FORTRAN_REQUEST_NULL) {
1209 F2C::lookup()->erase(id);
1213 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1217 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1218 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1222 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1228 if(req != MPI_REQUEST_NULL &&
1229 status != MPI_STATUS_IGNORE) {
1230 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1231 status->MPI_SOURCE = req->comm_->group()->rank(src);
1232 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1233 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1234 status->count = req->real_size_;
1239 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1240 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1242 *request = new Request();
1243 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1244 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1245 (*request)->refcount_ = 1;
1246 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1247 ((*request)->generalized_funcs)->query_fn=query_fn;
1248 ((*request)->generalized_funcs)->free_fn=free_fn;
1249 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1250 ((*request)->generalized_funcs)->extra_state=extra_state;
1251 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1252 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1256 int Request::grequest_complete(MPI_Request request)
1258 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1259 return MPI_ERR_REQUEST;
1260 request->generalized_funcs->mutex->lock();
1261 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1262 request->generalized_funcs->cond->notify_one();
1263 request->generalized_funcs->mutex->unlock();
1267 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1268 if (not reqs.empty()) {
1269 nbc_requests_ = reqs;
1270 Request::startall(reqs.size(), reqs.data());
1274 std::vector<MPI_Request> Request::get_nbc_requests() const
1276 return nbc_requests_;