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>
11 #include "xbt/virtu.h"
13 #include "src/mc/mc_replay.h"
14 #include "xbt/replay.h"
16 #include "src/simix/smx_private.h"
17 #include "surf/surf.h"
18 #include "simgrid/sg_config.h"
19 #include "smpi/smpi_utils.hpp"
20 #include "colls/colls.h"
22 #include "src/kernel/activity/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_activity_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_activity_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)){
58 if(req->src == MPI_ANY_SOURCE)
59 req->real_src = ref->src;
60 if(req->tag == MPI_ANY_TAG)
61 req->real_tag = ref->tag;
62 if(req->real_size < ref->real_size)
65 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
66 XBT_DEBUG("match succeeded");
72 std::vector<s_smpi_factor_t> smpi_os_values;
73 std::vector<s_smpi_factor_t> smpi_or_values;
74 std::vector<s_smpi_factor_t> smpi_ois_values;
76 static simgrid::config::Flag<double> smpi_wtime_sleep(
77 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
78 static simgrid::config::Flag<double> smpi_init_sleep(
79 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
80 static simgrid::config::Flag<double> smpi_iprobe_sleep(
81 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
82 static simgrid::config::Flag<double> smpi_test_sleep(
83 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
86 static double smpi_os(size_t size)
88 if (smpi_os_values.empty()) {
89 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
91 double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size;
92 // Iterate over all the sections that were specified and find the right
93 // value. (fact.factor represents the interval sizes; we want to find the
94 // section that has fact.factor <= size and no other such fact.factor <= size)
95 // Note: parse_factor() (used before) already sorts the vector we iterate over!
96 for (auto& fact : smpi_os_values) {
97 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
98 XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
101 // If the next section is too large, the current section must be used.
102 // Hence, save the cost, as we might have to use it.
103 current = fact.values[0]+fact.values[1]*size;
106 XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
111 static double smpi_ois(size_t size)
113 if (smpi_ois_values.empty()) {
114 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
116 double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size;
117 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
118 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
119 // Note: parse_factor() (used before) already sorts the vector we iterate over!
120 for (auto& fact : smpi_ois_values) {
121 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
122 XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
125 // If the next section is too large, the current section must be used.
126 // Hence, save the cost, as we might have to use it.
127 current = fact.values[0]+fact.values[1]*size;
130 XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
135 static double smpi_or(size_t size)
137 if (smpi_or_values.empty()) {
138 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
141 double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size;
143 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
144 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
145 // Note: parse_factor() (used before) already sorts the vector we iterate over!
146 for (auto fact : smpi_or_values) {
147 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
148 XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
151 // If the next section is too large, the current section must be used.
152 // Hence, save the cost, as we might have to use it.
153 current=fact.values[0]+fact.values[1]*size;
156 XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
161 void smpi_mpi_init() {
162 if(smpi_init_sleep > 0)
163 simcall_process_sleep(smpi_init_sleep);
166 double smpi_mpi_wtime(){
168 if (smpi_process_initialized() != 0 && smpi_process_finalized() == 0 && smpi_process_get_sampling() == 0) {
170 time = SIMIX_get_clock();
171 // to avoid deadlocks if used as a break condition, such as
172 // while (MPI_Wtime(...) < time_limit) {
175 // because the time will not normally advance when only calls to MPI_Wtime
176 // are made -> deadlock (MPI_Wtime never reaches the time limit)
177 if(smpi_wtime_sleep > 0)
178 simcall_process_sleep(smpi_wtime_sleep);
181 time = SIMIX_get_clock();
186 static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
189 MPI_Request request = nullptr;
191 void *old_buf = nullptr;
193 request = xbt_new(s_smpi_mpi_request_t, 1);
195 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
197 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
198 // This part handles the problem of non-contiguous memory
200 buf = count==0 ? nullptr : xbt_malloc(count*smpi_datatype_size(datatype));
201 if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
202 subtype->serialize(old_buf, buf, count, datatype->substruct);
207 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
208 request->old_buf = old_buf;
209 request->old_type = datatype;
211 request->size = smpi_datatype_size(datatype) * count;
212 smpi_datatype_use(datatype);
216 request->comm = comm;
217 smpi_comm_use(request->comm);
218 request->action = nullptr;
219 request->flags = flags;
220 request->detached = 0;
221 request->detached_sender = nullptr;
222 request->real_src = 0;
223 request->truncated = 0;
224 request->real_size = 0;
225 request->real_tag = 0;
226 if (flags & PERSISTENT)
227 request->refcount = 1;
229 request->refcount = 0;
230 request->op = MPI_REPLACE;
237 void smpi_empty_status(MPI_Status * status)
239 if(status != MPI_STATUS_IGNORE) {
240 status->MPI_SOURCE = MPI_ANY_SOURCE;
241 status->MPI_TAG = MPI_ANY_TAG;
242 status->MPI_ERROR = MPI_SUCCESS;
247 static void smpi_mpi_request_free_voidp(void* request)
249 MPI_Request req = static_cast<MPI_Request>(request);
250 smpi_mpi_request_free(&req);
253 /* MPI Low level calls */
254 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
256 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
257 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
258 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
262 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
264 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
265 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
266 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
270 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
272 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
273 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
274 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
275 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
279 void smpi_mpi_start(MPI_Request request)
281 smx_mailbox_t mailbox;
283 xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
284 request->flags &= ~PREPARED;
285 request->flags &= ~FINISHED;
288 if ((request->flags & RECV) != 0) {
289 print_request("New recv", request);
291 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
293 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
294 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
295 xbt_mutex_acquire(mut);
297 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
298 mailbox = smpi_process_mailbox();
300 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) {
301 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
302 //begin with the more appropriate one : the small one.
303 mailbox = smpi_process_mailbox_small();
304 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
305 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv,
306 static_cast<void*>(request));
308 if (action == nullptr) {
309 mailbox = smpi_process_mailbox();
310 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
311 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
312 if (action == nullptr) {
313 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
314 mailbox = smpi_process_mailbox_small();
317 XBT_DEBUG("yes there was something for us in the large mailbox");
320 mailbox = smpi_process_mailbox_small();
321 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
322 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
324 if (action == nullptr) {
325 XBT_DEBUG("No, nothing in the permanent receive mailbox");
326 mailbox = smpi_process_mailbox();
328 XBT_DEBUG("yes there was something for us in the small mailbox");
332 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
333 request->real_size=request->size;
334 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
335 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
336 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
337 XBT_DEBUG("recv simcall posted");
339 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
340 xbt_mutex_release(mut);
341 } else { /* the RECV flag was not set, so this is a send */
342 int receiver = request->dst;
344 int rank = request->src;
345 if (TRACE_smpi_view_internals()) {
346 TRACE_smpi_send(rank, rank, receiver, request->tag, request->size);
348 print_request("New send", request);
350 void* buf = request->buf;
351 if ((request->flags & SSEND) == 0 && ( (request->flags & RMA) != 0
352 || static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
353 void *oldbuf = nullptr;
354 request->detached = 1;
355 XBT_DEBUG("Send request %p is detached", request);
357 if(request->old_type->sizeof_substruct == 0){
358 oldbuf = request->buf;
359 if (!smpi_process_get_replaying() && oldbuf != nullptr && request->size!=0){
360 if((smpi_privatize_global_variables != 0)
361 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
362 && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
363 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
364 smpi_switch_data_segment(request->src);
366 buf = xbt_malloc(request->size);
367 memcpy(buf,oldbuf,request->size);
368 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
373 //if we are giving back the control to the user without waiting for completion, we have to inject timings
374 double sleeptime = 0.0;
375 if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
376 //isend and send timings may be different
377 sleeptime = ((request->flags & ISEND) != 0) ? smpi_ois(request->size) : smpi_os(request->size);
381 simcall_process_sleep(sleeptime);
382 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, sleeptime);
385 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
387 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
389 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
390 xbt_mutex_acquire(mut);
392 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
393 mailbox = smpi_process_remote_mailbox(receiver);
394 } else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
395 mailbox = smpi_process_remote_mailbox(receiver);
396 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
397 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send,
398 static_cast<void*>(request));
399 if (action == nullptr) {
400 if ((request->flags & SSEND) == 0){
401 mailbox = smpi_process_remote_mailbox_small(receiver);
402 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
404 mailbox = smpi_process_remote_mailbox_small(receiver);
405 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
406 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
407 if (action == nullptr) {
408 XBT_DEBUG("No, we are first, send to large mailbox");
409 mailbox = smpi_process_remote_mailbox(receiver);
413 XBT_DEBUG("Yes there was something for us in the large mailbox");
416 mailbox = smpi_process_remote_mailbox(receiver);
417 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
420 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
421 request->real_size=request->size;
422 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
423 buf, request->real_size, &match_send,
424 &xbt_free_f, // how to free the userdata if a detached send fails
425 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
426 : &smpi_comm_null_copy_buffer_callback, request,
427 // detach if msg size < eager/rdv switch limit
429 XBT_DEBUG("send simcall posted");
431 /* FIXME: detached sends are not traceable (request->action == nullptr) */
432 if (request->action != nullptr)
433 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
435 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
436 xbt_mutex_release(mut);
440 void smpi_mpi_startall(int count, MPI_Request * requests)
442 if(requests== nullptr)
445 for(int i = 0; i < count; i++) {
446 smpi_mpi_start(requests[i]);
450 void smpi_mpi_request_free(MPI_Request * request)
452 if((*request) != MPI_REQUEST_NULL){
453 (*request)->refcount--;
454 if((*request)->refcount<0) xbt_die("wrong refcount");
456 if((*request)->refcount==0){
457 smpi_datatype_unuse((*request)->old_type);
458 smpi_comm_unuse((*request)->comm);
459 print_request("Destroying", (*request));
461 *request = MPI_REQUEST_NULL;
463 print_request("Decrementing", (*request));
466 xbt_die("freeing an already free request");
470 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
473 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
475 request = build_request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
476 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
478 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
479 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
485 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
488 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
490 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
491 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
493 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
494 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
500 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
502 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
503 request = build_request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process_index(),
504 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
508 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
510 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
511 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
512 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
513 smpi_mpi_start(request);
517 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
519 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
520 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
521 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
522 smpi_mpi_start(request);
526 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
528 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
529 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
530 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
531 comm, PERSISTENT | RECV | PREPARED);
535 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
537 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
538 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
539 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
540 NON_PERSISTENT | RECV);
541 smpi_mpi_start(request);
545 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
547 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
548 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
549 smpi_mpi_wait(&request, status);
553 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
555 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
556 request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
557 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
559 smpi_mpi_start(request);
560 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
564 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
566 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, smpi_process_index(),
568 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
570 smpi_mpi_start(request);
571 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
575 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
576 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
577 MPI_Comm comm, MPI_Status * status)
579 MPI_Request requests[2];
581 int myid=smpi_process_index();
582 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
583 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
586 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
587 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
588 smpi_mpi_startall(2, requests);
589 smpi_mpi_waitall(2, requests, stats);
590 smpi_mpi_request_free(&requests[0]);
591 smpi_mpi_request_free(&requests[1]);
592 if(status != MPI_STATUS_IGNORE) {
593 // Copy receive status
598 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
600 return status->count / smpi_datatype_size(datatype);
603 static void finish_wait(MPI_Request * request, MPI_Status * status)
605 MPI_Request req = *request;
606 smpi_empty_status(status);
608 if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
609 if(status != MPI_STATUS_IGNORE) {
610 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
611 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
612 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
613 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
614 // this handles the case were size in receive differs from size in send
615 status->count = req->real_size;
618 print_request("Finishing", req);
619 MPI_Datatype datatype = req->old_type;
621 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
622 if (!smpi_process_get_replaying()){
623 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
624 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
625 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
626 smpi_switch_data_segment(smpi_process_index());
630 if(datatype->sizeof_substruct != 0){
631 // This part handles the problem of non-contignous memory the unserialization at the reception
632 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
633 if(req->flags & RECV)
634 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
635 datatype->substruct, req->op);
637 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
638 int n =req->real_size/smpi_datatype_size(datatype);
639 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
645 if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
646 int rank = smpi_process_index();
647 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
648 TRACE_smpi_recv(rank, src_traced, rank,req->tag);
651 if(req->detached_sender != nullptr){
652 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
653 double sleeptime = smpi_or(req->real_size);
655 simcall_process_sleep(sleeptime);
656 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size, sleeptime);
658 smpi_mpi_request_free(&(req->detached_sender));
660 if(req->flags & PERSISTENT)
661 req->action = nullptr;
662 req->flags |= FINISHED;
664 smpi_mpi_request_free(request);
667 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
668 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
670 // to avoid deadlocks if used as a break condition, such as
671 // while (MPI_Test(request, flag, status) && flag) {
673 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
674 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
675 static int nsleeps = 1;
676 if(smpi_test_sleep > 0)
677 simcall_process_sleep(nsleeps*smpi_test_sleep);
679 smpi_empty_status(status);
681 if (((*request)->flags & PREPARED) == 0) {
682 if ((*request)->action != nullptr)
683 flag = simcall_comm_test((*request)->action);
685 finish_wait(request, status);
686 nsleeps=1;//reset the number of sleeps we will do next time
687 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
688 *request = MPI_REQUEST_NULL;
689 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
696 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
698 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
699 comms.reserve(count);
704 *index = MPI_UNDEFINED;
706 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
707 for(i = 0; i < count; i++) {
708 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
709 comms.push_back(requests[i]->action);
714 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
715 static int nsleeps = 1;
716 if(smpi_test_sleep > 0)
717 simcall_process_sleep(nsleeps*smpi_test_sleep);
719 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
720 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
722 finish_wait(&requests[*index], status);
725 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) {
726 requests[*index] = MPI_REQUEST_NULL;
732 //all requests are null or inactive, return true
734 smpi_empty_status(status);
740 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
743 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
745 for(int i=0; i<count; i++){
746 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
747 if (smpi_mpi_test(&requests[i], pstat)!=1){
750 requests[i]=MPI_REQUEST_NULL;
753 smpi_empty_status(pstat);
755 if(status != MPI_STATUSES_IGNORE) {
762 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
764 //FIXME find another way to avoid busy waiting ?
765 // the issue here is that we have to wait on a nonexistent comm
767 smpi_mpi_iprobe(source, tag, comm, &flag, status);
768 XBT_DEBUG("Busy Waiting on probing : %d", flag);
772 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
773 MPI_Request request = build_request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
774 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
776 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
777 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
778 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
779 static int nsleeps = 1;
780 if(smpi_iprobe_sleep > 0)
781 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
782 // behave like a receive, but don't do it
783 smx_mailbox_t mailbox;
785 print_request("New iprobe", request);
786 // We have to test both mailboxes as we don't know if we will receive one one or another
787 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
788 mailbox = smpi_process_mailbox_small();
789 XBT_DEBUG("Trying to probe the perm recv mailbox");
790 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv,
791 static_cast<void*>(request));
794 if (request->action == nullptr){
795 mailbox = smpi_process_mailbox();
796 XBT_DEBUG("trying to probe the other mailbox");
797 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv,
798 static_cast<void*>(request));
801 if (request->action != nullptr){
802 simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action);
803 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
805 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
806 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
807 status->MPI_TAG = req->tag;
808 status->MPI_ERROR = MPI_SUCCESS;
809 status->count = req->real_size;
811 nsleeps = 1;//reset the number of sleeps we will do next time
815 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
818 smpi_mpi_request_free(&request);
823 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
825 print_request("Waiting", *request);
826 if ((*request)->flags & PREPARED) {
827 smpi_empty_status(status);
831 if ((*request)->action != nullptr)
832 // this is not a detached send
833 simcall_comm_wait((*request)->action, -1.0);
835 finish_wait(request, status);
836 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
837 *request = MPI_REQUEST_NULL;
840 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
845 int index = MPI_UNDEFINED;
849 // Wait for a request to complete
850 comms = xbt_dynar_new(sizeof(smx_activity_t), nullptr);
851 map = xbt_new(int, count);
852 XBT_DEBUG("Wait for one of %d", count);
853 for(i = 0; i < count; i++) {
854 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
855 if (requests[i]->action != nullptr) {
856 XBT_DEBUG("Waiting any %p ", requests[i]);
857 xbt_dynar_push(comms, &requests[i]->action);
861 //This is a finished detached request, let's return this one
862 size=0;//so we free the dynar but don't do the waitany call
864 finish_wait(&requests[i], status);//cleanup if refcount = 0
865 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
866 requests[i]=MPI_REQUEST_NULL;//set to null
872 i = simcall_comm_waitany(comms, -1);
874 // not MPI_UNDEFINED, as this is a simix return code
877 finish_wait(&requests[index], status);
878 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
879 requests[index] = MPI_REQUEST_NULL;
883 xbt_dynar_free(&comms);
886 if (index==MPI_UNDEFINED)
887 smpi_empty_status(status);
892 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
896 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
897 int retvalue = MPI_SUCCESS;
898 //tag invalid requests in the set
899 if (status != MPI_STATUSES_IGNORE) {
900 for (c = 0; c < count; c++) {
901 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
902 smpi_empty_status(&status[c]);
903 } else if (requests[c]->src == MPI_PROC_NULL) {
904 smpi_empty_status(&status[c]);
905 status[c].MPI_SOURCE = MPI_PROC_NULL;
909 for(c = 0; c < count; c++) {
911 if (MC_is_active() || MC_record_replay_is_active()) {
912 smpi_mpi_wait(&requests[c], pstat);
915 index = smpi_mpi_waitany(count, requests, pstat);
916 if (index == MPI_UNDEFINED)
918 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
919 requests[index]=MPI_REQUEST_NULL;
921 if (status != MPI_STATUSES_IGNORE) {
922 status[index] = *pstat;
923 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
924 retvalue = MPI_ERR_IN_STATUS;
931 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
937 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
939 for(i = 0; i < incount; i++)
941 index=smpi_mpi_waitany(incount, requests, pstat);
942 if(index!=MPI_UNDEFINED){
943 indices[count] = index;
945 if(status != MPI_STATUSES_IGNORE) {
946 status[index] = *pstat;
948 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
949 requests[index]=MPI_REQUEST_NULL;
951 return MPI_UNDEFINED;
957 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
963 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
965 for(i = 0; i < incount; i++) {
966 if((requests[i] != MPI_REQUEST_NULL)) {
967 if(smpi_mpi_test(&requests[i], pstat)) {
970 if(status != MPI_STATUSES_IGNORE) {
973 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
974 requests[i]=MPI_REQUEST_NULL;
980 if(count_dead==incount)
981 return MPI_UNDEFINED;
985 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
987 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
990 void smpi_mpi_barrier(MPI_Comm comm)
992 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
995 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
996 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
998 int system_tag = COLL_TAG_GATHER;
1000 MPI_Aint recvext = 0;
1002 int rank = smpi_comm_rank(comm);
1003 int size = smpi_comm_size(comm);
1005 // Send buffer to root
1006 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1008 smpi_datatype_extent(recvtype, &lb, &recvext);
1009 // Local copy from root
1010 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext,
1011 recvcount, recvtype);
1012 // Receive buffers from senders
1013 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1015 for (int src = 0; src < size; src++) {
1017 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1018 src, system_tag, comm);
1022 // Wait for completion of irecv's.
1023 smpi_mpi_startall(size - 1, requests);
1024 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1025 for (int src = 0; src < size-1; src++) {
1026 smpi_mpi_request_free(&requests[src]);
1032 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1035 int rank = smpi_comm_rank(comm);
1037 /* arbitrarily choose root as rank 0 */
1038 int size = smpi_comm_size(comm);
1040 int *displs = xbt_new(int, size);
1041 for (int i = 0; i < size; i++) {
1043 count += recvcounts[i];
1045 void *tmpbuf = static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1047 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1048 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1050 smpi_free_tmp_buffer(tmpbuf);
1053 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1054 MPI_Datatype recvtype, int root, MPI_Comm comm)
1056 int system_tag = COLL_TAG_GATHERV;
1058 MPI_Aint recvext = 0;
1060 int rank = smpi_comm_rank(comm);
1061 int size = smpi_comm_size(comm);
1063 // Send buffer to root
1064 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1066 smpi_datatype_extent(recvtype, &lb, &recvext);
1067 // Local copy from root
1068 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1069 recvcounts[root], recvtype);
1070 // Receive buffers from senders
1071 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1073 for (int src = 0; src < size; src++) {
1075 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1076 recvcounts[src], recvtype, src, system_tag, comm);
1080 // Wait for completion of irecv's.
1081 smpi_mpi_startall(size - 1, requests);
1082 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1083 for (int src = 0; src < size-1; src++) {
1084 smpi_mpi_request_free(&requests[src]);
1090 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1091 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1093 int system_tag = COLL_TAG_ALLGATHER;
1095 MPI_Aint recvext = 0;
1096 MPI_Request *requests;
1098 int rank = smpi_comm_rank(comm);
1099 int size = smpi_comm_size(comm);
1100 // FIXME: check for errors
1101 smpi_datatype_extent(recvtype, &lb, &recvext);
1102 // Local copy from self
1103 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount,
1105 // Send/Recv buffers to/from others;
1106 requests = xbt_new(MPI_Request, 2 * (size - 1));
1108 for (int other = 0; other < size; other++) {
1110 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1112 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype,
1113 other, system_tag, comm);
1117 // Wait for completion of all comms.
1118 smpi_mpi_startall(2 * (size - 1), requests);
1119 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1120 for (int other = 0; other < 2*(size-1); other++) {
1121 smpi_mpi_request_free(&requests[other]);
1126 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1127 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1129 int system_tag = COLL_TAG_ALLGATHERV;
1131 MPI_Aint recvext = 0;
1133 int rank = smpi_comm_rank(comm);
1134 int size = smpi_comm_size(comm);
1135 smpi_datatype_extent(recvtype, &lb, &recvext);
1136 // Local copy from self
1137 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1138 static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1139 // Send buffers to others;
1140 MPI_Request *requests = xbt_new(MPI_Request, 2 * (size - 1));
1142 for (int other = 0; other < size; other++) {
1145 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1147 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1148 recvtype, other, system_tag, comm);
1152 // Wait for completion of all comms.
1153 smpi_mpi_startall(2 * (size - 1), requests);
1154 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1155 for (int other = 0; other < 2*(size-1); other++) {
1156 smpi_mpi_request_free(&requests[other]);
1161 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1162 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1164 int system_tag = COLL_TAG_SCATTER;
1166 MPI_Aint sendext = 0;
1167 MPI_Request *requests;
1169 int rank = smpi_comm_rank(comm);
1170 int size = smpi_comm_size(comm);
1172 // Recv buffer from root
1173 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1175 smpi_datatype_extent(sendtype, &lb, &sendext);
1176 // Local copy from root
1177 if(recvbuf!=MPI_IN_PLACE){
1178 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1179 sendcount, sendtype, recvbuf, recvcount, recvtype);
1181 // Send buffers to receivers
1182 requests = xbt_new(MPI_Request, size - 1);
1184 for(int dst = 0; dst < size; dst++) {
1186 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype,
1187 dst, system_tag, comm);
1191 // Wait for completion of isend's.
1192 smpi_mpi_startall(size - 1, requests);
1193 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1194 for (int dst = 0; dst < size-1; dst++) {
1195 smpi_mpi_request_free(&requests[dst]);
1201 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1202 MPI_Datatype recvtype, int root, MPI_Comm comm)
1204 int system_tag = COLL_TAG_SCATTERV;
1206 MPI_Aint sendext = 0;
1208 int rank = smpi_comm_rank(comm);
1209 int size = smpi_comm_size(comm);
1211 // Recv buffer from root
1212 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1214 smpi_datatype_extent(sendtype, &lb, &sendext);
1215 // Local copy from root
1216 if(recvbuf!=MPI_IN_PLACE){
1217 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1218 sendtype, recvbuf, recvcount, recvtype);
1220 // Send buffers to receivers
1221 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1223 for (int dst = 0; dst < size; dst++) {
1225 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1226 sendtype, dst, system_tag, comm);
1230 // Wait for completion of isend's.
1231 smpi_mpi_startall(size - 1, requests);
1232 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1233 for (int dst = 0; dst < size-1; dst++) {
1234 smpi_mpi_request_free(&requests[dst]);
1240 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1243 int system_tag = COLL_TAG_REDUCE;
1245 MPI_Aint dataext = 0;
1247 char* sendtmpbuf = static_cast<char *>(sendbuf);
1249 int rank = smpi_comm_rank(comm);
1250 int size = smpi_comm_size(comm);
1251 //non commutative case, use a working algo from openmpi
1252 if(!smpi_op_is_commute(op)){
1253 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1257 if( sendbuf == MPI_IN_PLACE ) {
1258 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1259 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1263 // Send buffer to root
1264 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1266 smpi_datatype_extent(datatype, &lb, &dataext);
1267 // Local copy from root
1268 if (sendtmpbuf != nullptr && recvbuf != nullptr)
1269 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1270 // Receive buffers from senders
1271 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1272 void **tmpbufs = xbt_new(void *, size - 1);
1274 for (int src = 0; src < size; src++) {
1276 if (!smpi_process_get_replaying())
1277 tmpbufs[index] = xbt_malloc(count * dataext);
1279 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1281 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1285 // Wait for completion of irecv's.
1286 smpi_mpi_startall(size - 1, requests);
1287 for (int src = 0; src < size - 1; src++) {
1288 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1289 XBT_DEBUG("finished waiting any request with index %d", index);
1290 if(index == MPI_UNDEFINED) {
1293 smpi_mpi_request_free(&requests[index]);
1295 if(op) /* op can be MPI_OP_NULL that does nothing */
1296 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1298 for(index = 0; index < size - 1; index++) {
1299 smpi_free_tmp_buffer(tmpbufs[index]);
1305 if( sendbuf == MPI_IN_PLACE ) {
1306 smpi_free_tmp_buffer(sendtmpbuf);
1310 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1312 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1313 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1316 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1318 int system_tag = -888;
1319 MPI_Aint lb = 0, dataext = 0;
1321 int rank = smpi_comm_rank(comm);
1322 int size = smpi_comm_size(comm);
1324 smpi_datatype_extent(datatype, &lb, &dataext);
1326 // Local copy from self
1327 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1329 // Send/Recv buffers to/from others;
1330 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1331 void **tmpbufs = xbt_new(void *, rank);
1333 for (int other = 0; other < rank; other++) {
1334 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1335 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1338 for (int other = rank + 1; other < size; other++) {
1339 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1342 // Wait for completion of all comms.
1343 smpi_mpi_startall(size - 1, requests);
1345 if(smpi_op_is_commute(op)){
1346 for (int other = 0; other < size - 1; other++) {
1347 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1348 if(index == MPI_UNDEFINED) {
1352 // #Request is below rank: it's a irecv
1353 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1357 //non commutative case, wait in order
1358 for (int other = 0; other < size - 1; other++) {
1359 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1361 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1365 for(index = 0; index < rank; index++) {
1366 smpi_free_tmp_buffer(tmpbufs[index]);
1368 for(index = 0; index < size-1; index++) {
1369 smpi_mpi_request_free(&requests[index]);
1375 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1377 int system_tag = -888;
1378 MPI_Aint lb = 0, dataext = 0;
1379 int recvbuf_is_empty=1;
1380 int rank = smpi_comm_rank(comm);
1381 int size = smpi_comm_size(comm);
1383 smpi_datatype_extent(datatype, &lb, &dataext);
1385 // Send/Recv buffers to/from others;
1386 MPI_Request *requests = xbt_new(MPI_Request, size - 1);
1387 void **tmpbufs = xbt_new(void *, rank);
1389 for (int other = 0; other < rank; other++) {
1390 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1391 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1394 for (int other = rank + 1; other < size; other++) {
1395 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1398 // Wait for completion of all comms.
1399 smpi_mpi_startall(size - 1, requests);
1400 if(smpi_op_is_commute(op)){
1401 for (int other = 0; other < size - 1; other++) {
1402 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1403 if(index == MPI_UNDEFINED) {
1407 if(recvbuf_is_empty){
1408 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1411 // #Request is below rank: it's a irecv
1412 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1416 //non commutative case, wait in order
1417 for (int other = 0; other < size - 1; other++) {
1418 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1420 if(recvbuf_is_empty){
1421 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1424 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1428 for(index = 0; index < rank; index++) {
1429 smpi_free_tmp_buffer(tmpbufs[index]);
1431 for(index = 0; index < size-1; index++) {
1432 smpi_mpi_request_free(&requests[index]);