1 /* Copyright (c) 2007-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include <xbt/config.hpp>
10 #include "xbt/virtu.h"
12 #include "src/mc/mc_replay.h"
13 #include "xbt/replay.h"
15 #include "src/simix/smx_private.h"
16 #include "surf/surf.h"
17 #include "simgrid/sg_config.h"
18 #include "colls/colls.h"
20 #include "src/simix/SynchroComm.hpp"
22 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
24 static int match_recv(void* a, void* b, smx_synchro_t ignored) {
25 MPI_Request ref = static_cast<MPI_Request>(a);
26 MPI_Request req = static_cast<MPI_Request>(b);
27 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
29 xbt_assert(ref, "Cannot match recv against null reference");
30 xbt_assert(req, "Cannot match recv against null request");
31 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
32 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
33 //we match, we can transfer some values
34 if(ref->src == MPI_ANY_SOURCE)
35 ref->real_src = req->src;
36 if(ref->tag == MPI_ANY_TAG)
37 ref->real_tag = req->tag;
38 if(ref->real_size < req->real_size)
41 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
42 XBT_DEBUG("match succeeded");
47 static int match_send(void* a, void* b,smx_synchro_t ignored) {
48 MPI_Request ref = static_cast<MPI_Request>(a);
49 MPI_Request req = static_cast<MPI_Request>(b);
50 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
51 xbt_assert(ref, "Cannot match send against null reference");
52 xbt_assert(req, "Cannot match send against null request");
54 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
55 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
57 if(req->src == MPI_ANY_SOURCE)
58 req->real_src = ref->src;
59 if(req->tag == MPI_ANY_TAG)
60 req->real_tag = ref->tag;
61 if(req->real_size < ref->real_size)
64 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
65 XBT_DEBUG("match succeeded");
70 // Methods used to parse and store the values for timing injections in smpi
71 // These are taken from surf/network.c and generalized to have more values for each factor
72 typedef struct s_smpi_factor_multival *smpi_os_factor_multival_t;
73 typedef struct s_smpi_factor_multival { // FIXME: this should be merged (deduplicated) with s_smpi_factor defined in network_smpi.c
76 double values[4];//arbitrary set to 4
77 } s_smpi_factor_multival_t;
79 xbt_dynar_t smpi_os_values = nullptr;
80 xbt_dynar_t smpi_or_values = nullptr;
81 xbt_dynar_t smpi_ois_values = nullptr;
83 static simgrid::config::Flag<double> smpi_wtime_sleep(
84 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
85 static simgrid::config::Flag<double> smpi_init_sleep(
86 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
87 static simgrid::config::Flag<double> smpi_iprobe_sleep(
88 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
89 static simgrid::config::Flag<double> smpi_test_sleep(
90 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
92 static int factor_cmp(const void *pa, const void *pb)
94 return ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor > (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? 1 :
95 ((static_cast<const s_smpi_factor_multival_t*>(pa))->factor < (static_cast<const s_smpi_factor_multival_t*>(pb))->factor) ? -1 : 0;
98 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
101 unsigned int iter = 0;
102 s_smpi_factor_multival_t fact;
105 xbt_dynar_t radical_elements2 = NULL;
107 xbt_dynar_t smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_multival_t), NULL);
108 xbt_dynar_t radical_elements = xbt_str_split(smpi_coef_string, ";");
109 xbt_dynar_foreach(radical_elements, iter, value) {
110 memset(&fact, 0, sizeof(s_smpi_factor_multival_t));
111 radical_elements2 = xbt_str_split(value, ":");
112 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
113 xbt_die("Malformed radical for smpi factor: '%s'", smpi_coef_string);
114 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
117 errmsg = bprintf("Invalid factor in chunk #%d: %%s", iter+1);
118 fact.factor = xbt_str_parse_int(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
120 errmsg = bprintf("Invalid factor value %d in chunk #%d: %%s", i, iter+1);
121 fact.values[fact.nb_values] = xbt_str_parse_double(xbt_dynar_get_as(radical_elements2, i, char *), errmsg);
127 xbt_dynar_push_as(smpi_factor, s_smpi_factor_multival_t, fact);
128 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
129 xbt_dynar_free(&radical_elements2);
131 xbt_dynar_free(&radical_elements);
132 xbt_dynar_sort(smpi_factor, &factor_cmp);
133 xbt_dynar_foreach(smpi_factor, iter, fact) {
134 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
139 static double smpi_os(double size)
141 if (smpi_os_values == nullptr) {
142 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
143 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
145 unsigned int iter = 0;
146 s_smpi_factor_multival_t fact;
148 // Iterate over all the sections that were specified and find the right
149 // value. (fact.factor represents the interval sizes; we want to find the
150 // section that has fact.factor <= size and no other such fact.factor <= size)
151 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
152 xbt_dynar_foreach(smpi_os_values, iter, fact) {
153 if (size <= fact.factor) { // Values already too large, use the previously
154 // computed value of current!
155 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
158 // If the next section is too large, the current section must be used.
159 // Hence, save the cost, as we might have to use it.
160 current = fact.values[0]+fact.values[1]*size;
163 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
168 static double smpi_ois(double size)
170 if (smpi_ois_values == nullptr) {
171 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
172 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
174 unsigned int iter = 0;
175 s_smpi_factor_multival_t fact;
177 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
178 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
179 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
180 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
181 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
182 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
185 // If the next section is too large, the current section must be used.
186 // Hence, save the cost, as we might have to use it.
187 current = fact.values[0]+fact.values[1]*size;
190 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
195 static double smpi_or(double size)
197 if (smpi_or_values == nullptr) {
198 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
199 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
201 unsigned int iter = 0;
202 s_smpi_factor_multival_t fact;
204 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
205 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
206 // Note: parse_factor() (used before) already sorts the dynar we iterate over!
207 xbt_dynar_foreach(smpi_or_values, iter, fact) {
208 if (size <= fact.factor) { // Values already too large, use the previously
209 // computed value of current!
210 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
213 // If the next section is too large, the current section must be used.
214 // Hence, save the cost, as we might have to use it.
215 current=fact.values[0]+fact.values[1]*size;
218 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
223 void smpi_mpi_init() {
224 if(smpi_init_sleep > 0)
225 simcall_process_sleep(smpi_init_sleep);
228 double smpi_mpi_wtime(){
230 if (smpi_process_initialized() != 0 &&
231 smpi_process_finalized() == 0 &&
232 smpi_process_get_sampling() == 0) {
234 time = SIMIX_get_clock();
235 // to avoid deadlocks if used as a break condition, such as
236 // while (MPI_Wtime(...) < time_limit) {
239 // because the time will not normally advance when only calls to MPI_Wtime
240 // are made -> deadlock (MPI_Wtime never reaches the time limit)
241 if(smpi_wtime_sleep > 0)
242 simcall_process_sleep(smpi_wtime_sleep);
245 time = SIMIX_get_clock();
250 static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
253 MPI_Request request = NULL;
255 void *old_buf = NULL;
257 request = xbt_new(s_smpi_mpi_request_t, 1);
259 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
261 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
262 // This part handles the problem of non-contiguous memory
264 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
265 if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
266 subtype->serialize(old_buf, buf, count, datatype->substruct);
271 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
272 request->old_buf = old_buf;
273 request->old_type = datatype;
275 request->size = smpi_datatype_size(datatype) * count;
276 smpi_datatype_use(datatype);
280 request->comm = comm;
281 smpi_comm_use(request->comm);
282 request->action = nullptr;
283 request->flags = flags;
284 request->detached = 0;
285 request->detached_sender = nullptr;
286 request->real_src = 0;
288 request->truncated = 0;
289 request->real_size = 0;
290 request->real_tag = 0;
291 if (flags & PERSISTENT)
292 request->refcount = 1;
294 request->refcount = 0;
295 request->op = MPI_REPLACE;
302 void smpi_empty_status(MPI_Status * status)
304 if(status != MPI_STATUS_IGNORE) {
305 status->MPI_SOURCE = MPI_ANY_SOURCE;
306 status->MPI_TAG = MPI_ANY_TAG;
307 status->MPI_ERROR = MPI_SUCCESS;
312 static void smpi_mpi_request_free_voidp(void* request)
314 MPI_Request req = static_cast<MPI_Request>(request);
315 smpi_mpi_request_free(&req);
318 /* MPI Low level calls */
319 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
320 int dst, int tag, MPI_Comm comm)
322 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
323 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
324 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
328 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
329 int dst, int tag, MPI_Comm comm)
331 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
332 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
333 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
337 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
338 int src, int tag, MPI_Comm comm)
340 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
341 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype,
342 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
343 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
347 void smpi_mpi_start(MPI_Request request)
349 smx_mailbox_t mailbox;
351 xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
352 request->flags &= ~PREPARED;
353 request->flags &= ~FINISHED;
356 if ((request->flags & RECV) != 0) {
357 print_request("New recv", request);
359 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
361 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
362 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
363 xbt_mutex_acquire(mut);
365 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
366 mailbox = smpi_process_mailbox();
367 } else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh){
368 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
369 //begin with the more appropriate one : the small one.
370 mailbox = smpi_process_mailbox_small();
371 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
372 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
374 if(action == nullptr){
375 mailbox = smpi_process_mailbox();
376 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
377 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
378 if(action == nullptr){
379 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
380 mailbox = smpi_process_mailbox_small();
383 XBT_DEBUG("yes there was something for us in the large mailbox");
386 mailbox = smpi_process_mailbox_small();
387 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
388 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
390 if(action == nullptr){
391 XBT_DEBUG("No, nothing in the permanent receive mailbox");
392 mailbox = smpi_process_mailbox();
394 XBT_DEBUG("yes there was something for us in the small mailbox");
398 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
399 double sleeptime = (request->detached != 0) ? smpi_or(request->size) : 0.0;
401 simcall_process_sleep(sleeptime);
402 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
405 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
406 request->real_size=request->size;
407 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
408 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
409 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
410 XBT_DEBUG("recv simcall posted");
412 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
413 xbt_mutex_release(mut);
415 int receiver = request->dst;
417 int rank = request->src;
418 if (TRACE_smpi_view_internals()) {
419 TRACE_smpi_send(rank, rank, receiver,request->size);
421 print_request("New send", request);
423 //if we are giving back the control to the user without waiting for completion, we have to inject timings
424 double sleeptime = 0.0;
425 if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
426 //isend and send timings may be different
427 sleeptime = ((request->flags & ISEND) != 0)? smpi_ois(request->size) : smpi_os(request->size);
431 simcall_process_sleep(sleeptime);
432 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
435 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
437 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
439 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
440 xbt_mutex_acquire(mut);
442 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
443 mailbox = smpi_process_remote_mailbox(receiver);
445 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
446 mailbox = smpi_process_remote_mailbox(receiver);
447 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
448 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
449 if (action == nullptr) {
450 if ((request->flags & SSEND) == 0){
451 mailbox = smpi_process_remote_mailbox_small(receiver);
452 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
455 mailbox = smpi_process_remote_mailbox_small(receiver);
456 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
457 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
458 if (action == nullptr) {
459 XBT_DEBUG("No, we are first, send to large mailbox");
460 mailbox = smpi_process_remote_mailbox(receiver);
465 XBT_DEBUG("Yes there was something for us in the large mailbox");
469 mailbox = smpi_process_remote_mailbox(receiver);
470 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
473 void* buf = request->buf;
474 if ( ((request->flags & SSEND) == 0) && (((request->flags & RMA) != 0) ||
475 (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
477 request->detached = 1;
478 XBT_DEBUG("Send request %p is detached", request);
480 if(request->old_type->sizeof_substruct == 0){
481 oldbuf = request->buf;
482 if (!smpi_process_get_replaying() && oldbuf != NULL && request->size!=0){
483 if((smpi_privatize_global_variables != 0)
484 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
485 && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
486 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
487 smpi_switch_data_segment(request->src);
489 buf = xbt_malloc(request->size);
490 memcpy(buf,oldbuf,request->size);
491 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
496 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
497 request->real_size=request->size;
498 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
499 buf, request->real_size, &match_send,
500 &xbt_free_f, // how to free the userdata if a detached send fails
501 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
502 : &smpi_comm_null_copy_buffer_callback, request,
503 // detach if msg size < eager/rdv switch limit
505 XBT_DEBUG("send simcall posted");
507 /* FIXME: detached sends are not traceable (request->action == nullptr) */
508 if (request->action != nullptr)
509 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
511 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
512 xbt_mutex_release(mut);
516 void smpi_mpi_startall(int count, MPI_Request * requests)
518 if(requests== nullptr)
521 for(int i = 0; i < count; i++) {
522 smpi_mpi_start(requests[i]);
526 void smpi_mpi_request_free(MPI_Request * request)
528 if((*request) != MPI_REQUEST_NULL){
529 (*request)->refcount--;
530 if((*request)->refcount<0) xbt_die("wrong refcount");
532 if((*request)->refcount==0){
533 smpi_datatype_unuse((*request)->old_type);
534 smpi_comm_unuse((*request)->comm);
535 print_request("Destroying", (*request));
537 *request = MPI_REQUEST_NULL;
539 print_request("Decrementing", (*request));
542 xbt_die("freeing an already free request");
546 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
549 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
551 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, src, dst, tag,
552 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
554 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
555 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
561 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
564 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
566 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
567 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
569 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
570 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
576 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
578 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
579 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, smpi_process_index(),
580 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
584 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
586 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
587 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
588 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
589 smpi_mpi_start(request);
593 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
595 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
596 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
597 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
598 smpi_mpi_start(request);
602 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
604 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
605 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
606 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
607 comm, PERSISTENT | RECV | PREPARED);
611 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
613 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
614 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
615 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
616 NON_PERSISTENT | RECV);
617 smpi_mpi_start(request);
621 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
623 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
624 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
625 smpi_mpi_wait(&request, status);
629 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
631 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
632 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
633 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
635 smpi_mpi_start(request);
636 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
640 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
642 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
643 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
644 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
646 smpi_mpi_start(request);
647 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
651 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
652 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
653 MPI_Comm comm, MPI_Status * status)
655 MPI_Request requests[2];
657 int myid=smpi_process_index();
658 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
659 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
662 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
663 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
664 smpi_mpi_startall(2, requests);
665 smpi_mpi_waitall(2, requests, stats);
666 smpi_mpi_request_free(&requests[0]);
667 smpi_mpi_request_free(&requests[1]);
668 if(status != MPI_STATUS_IGNORE) {
669 // Copy receive status
674 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
676 return status->count / smpi_datatype_size(datatype);
679 static void finish_wait(MPI_Request * request, MPI_Status * status)
681 MPI_Request req = *request;
682 smpi_empty_status(status);
684 if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
685 if(status != MPI_STATUS_IGNORE) {
686 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
687 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
688 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
689 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
690 // this handles the case were size in receive differs from size in send
691 status->count = req->real_size;
694 print_request("Finishing", req);
695 MPI_Datatype datatype = req->old_type;
697 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
698 if (!smpi_process_get_replaying()){
699 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
700 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
701 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
702 smpi_switch_data_segment(smpi_process_index());
706 if(datatype->sizeof_substruct != 0){
707 // This part handles the problem of non-contignous memory the unserialization at the reception
708 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
709 if(req->flags & RECV)
710 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
711 datatype->substruct, req->op);
713 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
714 int n =req->real_size/smpi_datatype_size(datatype);
715 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
721 if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
722 int rank = smpi_process_index();
723 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
724 TRACE_smpi_recv(rank, src_traced, rank);
727 if(req->detached_sender!=NULL){
728 smpi_mpi_request_free(&(req->detached_sender));
730 if(req->flags & PERSISTENT)
732 req->flags |= FINISHED;
734 smpi_mpi_request_free(request);
737 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
738 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
740 // to avoid deadlocks if used as a break condition, such as
741 // while (MPI_Test(request, flag, status) && flag) {
743 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
744 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
745 static int nsleeps = 1;
746 if(smpi_test_sleep > 0)
747 simcall_process_sleep(nsleeps*smpi_test_sleep);
749 smpi_empty_status(status);
751 if (((*request)->flags & PREPARED) == 0) {
752 if ((*request)->action != NULL)
753 flag = simcall_comm_test((*request)->action);
755 finish_wait(request, status);
756 nsleeps=1;//reset the number of sleeps we will do next time
757 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
758 *request = MPI_REQUEST_NULL;
766 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
774 *index = MPI_UNDEFINED;
775 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
776 map = xbt_new(int, count);
777 for(i = 0; i < count; i++) {
778 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
779 xbt_dynar_push(comms, &requests[i]->action);
785 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
786 static int nsleeps = 1;
787 if(smpi_test_sleep > 0)
788 simcall_process_sleep(nsleeps*smpi_test_sleep);
790 i = simcall_comm_testany(comms);
791 // not MPI_UNDEFINED, as this is a simix return code
794 finish_wait(&requests[*index], status);
795 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
796 requests[*index] = MPI_REQUEST_NULL;
803 //all requests are null or inactive, return true
805 smpi_empty_status(status);
808 xbt_dynar_free(&comms);
813 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
816 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
819 for(i=0; i<count; i++){
820 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
821 if (smpi_mpi_test(&requests[i], pstat)!=1){
824 requests[i]=MPI_REQUEST_NULL;
827 smpi_empty_status(pstat);
829 if(status != MPI_STATUSES_IGNORE) {
836 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
838 //FIXME find another way to avoid busy waiting ?
839 // the issue here is that we have to wait on a nonexistent comm
841 smpi_mpi_iprobe(source, tag, comm, &flag, status);
842 XBT_DEBUG("Busy Waiting on probing : %d", flag);
846 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
848 MPI_Request request = build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
849 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
851 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
852 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
853 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
854 static int nsleeps = 1;
855 if(smpi_iprobe_sleep > 0)
856 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
857 // behave like a receive, but don't do it
858 smx_mailbox_t mailbox;
860 print_request("New iprobe", request);
861 // We have to test both mailboxes as we don't know if we will receive one one or another
862 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
863 mailbox = smpi_process_mailbox_small();
864 XBT_DEBUG("Trying to probe the perm recv mailbox");
865 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
868 if (request->action == nullptr){
869 mailbox = smpi_process_mailbox();
870 XBT_DEBUG("trying to probe the other mailbox");
871 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
874 if (request->action != nullptr){
875 simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
876 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
878 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
879 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
880 status->MPI_TAG = req->tag;
881 status->MPI_ERROR = MPI_SUCCESS;
882 status->count = req->real_size;
884 nsleeps = 1;//reset the number of sleeps we will do next time
890 smpi_mpi_request_free(&request);
895 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
897 print_request("Waiting", *request);
898 if ((*request)->flags & PREPARED) {
899 smpi_empty_status(status);
903 if ((*request)->action != NULL)
904 // this is not a detached send
905 simcall_comm_wait((*request)->action, -1.0);
907 finish_wait(request, status);
908 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
909 *request = MPI_REQUEST_NULL;
912 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
917 int index = MPI_UNDEFINED;
921 // Wait for a request to complete
922 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
923 map = xbt_new(int, count);
924 XBT_DEBUG("Wait for one of %d", count);
925 for(i = 0; i < count; i++) {
926 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
927 if (requests[i]->action != NULL) {
928 XBT_DEBUG("Waiting any %p ", requests[i]);
929 xbt_dynar_push(comms, &requests[i]->action);
933 //This is a finished detached request, let's return this one
934 size=0;//so we free the dynar but don't do the waitany call
936 finish_wait(&requests[i], status);//cleanup if refcount = 0
937 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
938 requests[i]=MPI_REQUEST_NULL;//set to null
944 i = simcall_comm_waitany(comms);
946 // not MPI_UNDEFINED, as this is a simix return code
949 finish_wait(&requests[index], status);
950 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
951 requests[index] = MPI_REQUEST_NULL;
955 xbt_dynar_free(&comms);
958 if (index==MPI_UNDEFINED)
959 smpi_empty_status(status);
964 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
968 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
969 int retvalue = MPI_SUCCESS;
970 //tag invalid requests in the set
971 if (status != MPI_STATUSES_IGNORE) {
972 for (c = 0; c < count; c++) {
973 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
974 smpi_empty_status(&status[c]);
975 } else if (requests[c]->src == MPI_PROC_NULL) {
976 smpi_empty_status(&status[c]);
977 status[c].MPI_SOURCE = MPI_PROC_NULL;
981 for(c = 0; c < count; c++) {
983 if (MC_is_active() || MC_record_replay_is_active()) {
984 smpi_mpi_wait(&requests[c], pstat);
987 index = smpi_mpi_waitany(count, requests, pstat);
988 if (index == MPI_UNDEFINED)
990 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
991 requests[index]=MPI_REQUEST_NULL;
993 if (status != MPI_STATUSES_IGNORE) {
994 status[index] = *pstat;
995 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
996 retvalue = MPI_ERR_IN_STATUS;
1003 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1009 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1011 for(i = 0; i < incount; i++)
1013 index=smpi_mpi_waitany(incount, requests, pstat);
1014 if(index!=MPI_UNDEFINED){
1015 indices[count] = index;
1017 if(status != MPI_STATUSES_IGNORE) {
1018 status[index] = *pstat;
1020 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1021 requests[index]=MPI_REQUEST_NULL;
1023 return MPI_UNDEFINED;
1029 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1035 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1037 for(i = 0; i < incount; i++) {
1038 if((requests[i] != MPI_REQUEST_NULL)) {
1039 if(smpi_mpi_test(&requests[i], pstat)) {
1042 if(status != MPI_STATUSES_IGNORE) {
1045 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1046 requests[i]=MPI_REQUEST_NULL;
1052 if(count_dead==incount)
1053 return MPI_UNDEFINED;
1057 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
1059 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
1062 void smpi_mpi_barrier(MPI_Comm comm)
1064 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
1067 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1068 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1070 int system_tag = COLL_TAG_GATHER;
1071 int rank, size, src, index;
1072 MPI_Aint lb = 0, recvext = 0;
1073 MPI_Request *requests;
1075 rank = smpi_comm_rank(comm);
1076 size = smpi_comm_size(comm);
1078 // Send buffer to root
1079 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1081 smpi_datatype_extent(recvtype, &lb, &recvext);
1082 // Local copy from root
1083 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext, recvcount, recvtype);
1084 // Receive buffers from senders
1085 requests = xbt_new(MPI_Request, size - 1);
1087 for(src = 0; src < size; src++) {
1089 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1090 src, system_tag, comm);
1094 // Wait for completion of irecv's.
1095 smpi_mpi_startall(size - 1, requests);
1096 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1097 for(src = 0; src < size-1; src++) {
1098 smpi_mpi_request_free(&requests[src]);
1104 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1109 int rank = smpi_comm_rank(comm);
1112 /* arbitrarily choose root as rank 0 */
1113 size = smpi_comm_size(comm);
1115 displs = xbt_new(int, size);
1116 for (i = 0; i < size; i++) {
1118 count += recvcounts[i];
1120 tmpbuf=static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1122 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1123 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1125 smpi_free_tmp_buffer(tmpbuf);
1128 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1129 MPI_Datatype recvtype, int root, MPI_Comm comm)
1131 int system_tag = COLL_TAG_GATHERV;
1132 int rank, size, src, index;
1133 MPI_Aint lb = 0, recvext = 0;
1134 MPI_Request *requests;
1136 rank = smpi_comm_rank(comm);
1137 size = smpi_comm_size(comm);
1139 // Send buffer to root
1140 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1142 smpi_datatype_extent(recvtype, &lb, &recvext);
1143 // Local copy from root
1144 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1145 recvcounts[root], recvtype);
1146 // Receive buffers from senders
1147 requests = xbt_new(MPI_Request, size - 1);
1149 for(src = 0; src < size; src++) {
1151 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1152 recvcounts[src], recvtype, src, system_tag, comm);
1156 // Wait for completion of irecv's.
1157 smpi_mpi_startall(size - 1, requests);
1158 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1159 for(src = 0; src < size-1; src++) {
1160 smpi_mpi_request_free(&requests[src]);
1166 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1167 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1169 int system_tag = COLL_TAG_ALLGATHER;
1170 int rank, size, other, index;
1171 MPI_Aint lb = 0, recvext = 0;
1172 MPI_Request *requests;
1174 rank = smpi_comm_rank(comm);
1175 size = smpi_comm_size(comm);
1176 // FIXME: check for errors
1177 smpi_datatype_extent(recvtype, &lb, &recvext);
1178 // Local copy from self
1179 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
1180 // Send/Recv buffers to/from others;
1181 requests = xbt_new(MPI_Request, 2 * (size - 1));
1183 for(other = 0; other < size; other++) {
1185 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1187 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other,
1192 // Wait for completion of all comms.
1193 smpi_mpi_startall(2 * (size - 1), requests);
1194 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1195 for(other = 0; other < 2*(size-1); other++) {
1196 smpi_mpi_request_free(&requests[other]);
1201 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1202 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1204 int system_tag = COLL_TAG_ALLGATHERV;
1205 int rank, size, other, index;
1206 MPI_Aint lb = 0, recvext = 0;
1207 MPI_Request *requests;
1209 rank = smpi_comm_rank(comm);
1210 size = smpi_comm_size(comm);
1211 smpi_datatype_extent(recvtype, &lb, &recvext);
1212 // Local copy from self
1213 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1214 // Send buffers to others;
1215 requests = xbt_new(MPI_Request, 2 * (size - 1));
1217 for(other = 0; other < size; other++) {
1220 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1222 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1223 recvtype, other, system_tag, comm);
1227 // Wait for completion of all comms.
1228 smpi_mpi_startall(2 * (size - 1), requests);
1229 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1230 for(other = 0; other < 2*(size-1); other++) {
1231 smpi_mpi_request_free(&requests[other]);
1236 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1237 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1239 int system_tag = COLL_TAG_SCATTER;
1240 int rank, size, dst, index;
1241 MPI_Aint lb = 0, sendext = 0;
1242 MPI_Request *requests;
1244 rank = smpi_comm_rank(comm);
1245 size = smpi_comm_size(comm);
1247 // Recv buffer from root
1248 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1250 smpi_datatype_extent(sendtype, &lb, &sendext);
1251 // Local copy from root
1252 if(recvbuf!=MPI_IN_PLACE){
1253 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1254 sendcount, sendtype, recvbuf, recvcount, recvtype);
1256 // Send buffers to receivers
1257 requests = xbt_new(MPI_Request, size - 1);
1259 for(dst = 0; dst < size; dst++) {
1261 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst,
1266 // Wait for completion of isend's.
1267 smpi_mpi_startall(size - 1, requests);
1268 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1269 for(dst = 0; dst < size-1; dst++) {
1270 smpi_mpi_request_free(&requests[dst]);
1276 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1277 MPI_Datatype recvtype, int root, MPI_Comm comm)
1279 int system_tag = COLL_TAG_SCATTERV;
1280 int rank, size, dst, index;
1281 MPI_Aint lb = 0, sendext = 0;
1282 MPI_Request *requests;
1284 rank = smpi_comm_rank(comm);
1285 size = smpi_comm_size(comm);
1287 // Recv buffer from root
1288 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1290 smpi_datatype_extent(sendtype, &lb, &sendext);
1291 // Local copy from root
1292 if(recvbuf!=MPI_IN_PLACE){
1293 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1294 sendtype, recvbuf, recvcount, recvtype);
1296 // Send buffers to receivers
1297 requests = xbt_new(MPI_Request, size - 1);
1299 for(dst = 0; dst < size; dst++) {
1301 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1302 sendtype, dst, system_tag, comm);
1306 // Wait for completion of isend's.
1307 smpi_mpi_startall(size - 1, requests);
1308 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1309 for(dst = 0; dst < size-1; dst++) {
1310 smpi_mpi_request_free(&requests[dst]);
1316 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1319 int system_tag = COLL_TAG_REDUCE;
1320 int rank, size, src, index;
1321 MPI_Aint lb = 0, dataext = 0;
1322 MPI_Request *requests;
1325 char* sendtmpbuf = static_cast<char *>(sendbuf);
1328 rank = smpi_comm_rank(comm);
1329 size = smpi_comm_size(comm);
1330 //non commutative case, use a working algo from openmpi
1331 if(!smpi_op_is_commute(op)){
1332 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1336 if( sendbuf == MPI_IN_PLACE ) {
1337 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1338 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1342 // Send buffer to root
1343 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1345 smpi_datatype_extent(datatype, &lb, &dataext);
1346 // Local copy from root
1347 if (sendtmpbuf != NULL && recvbuf != NULL)
1348 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1349 // Receive buffers from senders
1350 requests = xbt_new(MPI_Request, size - 1);
1351 tmpbufs = xbt_new(void *, size - 1);
1353 for(src = 0; src < size; src++) {
1355 if (!smpi_process_get_replaying())
1356 tmpbufs[index] = xbt_malloc(count * dataext);
1358 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1360 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1364 // Wait for completion of irecv's.
1365 smpi_mpi_startall(size - 1, requests);
1366 for(src = 0; src < size - 1; src++) {
1367 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1368 XBT_DEBUG("finished waiting any request with index %d", index);
1369 if(index == MPI_UNDEFINED) {
1372 smpi_mpi_request_free(&requests[index]);
1374 if(op) /* op can be MPI_OP_NULL that does nothing */
1375 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1377 for(index = 0; index < size - 1; index++) {
1378 smpi_free_tmp_buffer(tmpbufs[index]);
1384 if( sendbuf == MPI_IN_PLACE ) {
1385 smpi_free_tmp_buffer(sendtmpbuf);
1389 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1391 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1392 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1395 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1397 int system_tag = -888;
1398 int rank, size, other, index;
1399 MPI_Aint lb = 0, dataext = 0;
1400 MPI_Request *requests;
1403 rank = smpi_comm_rank(comm);
1404 size = smpi_comm_size(comm);
1406 smpi_datatype_extent(datatype, &lb, &dataext);
1408 // Local copy from self
1409 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1411 // Send/Recv buffers to/from others;
1412 requests = xbt_new(MPI_Request, size - 1);
1413 tmpbufs = xbt_new(void *, rank);
1415 for(other = 0; other < rank; other++) {
1416 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1417 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1420 for(other = rank + 1; other < size; other++) {
1421 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1424 // Wait for completion of all comms.
1425 smpi_mpi_startall(size - 1, requests);
1427 if(smpi_op_is_commute(op)){
1428 for(other = 0; other < size - 1; other++) {
1429 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1430 if(index == MPI_UNDEFINED) {
1434 // #Request is below rank: it's a irecv
1435 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1439 //non commutative case, wait in order
1440 for(other = 0; other < size - 1; other++) {
1441 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1443 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1447 for(index = 0; index < rank; index++) {
1448 smpi_free_tmp_buffer(tmpbufs[index]);
1450 for(index = 0; index < size-1; index++) {
1451 smpi_mpi_request_free(&requests[index]);
1457 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1459 int system_tag = -888;
1460 int rank, size, other, index;
1461 MPI_Aint lb = 0, dataext = 0;
1462 MPI_Request *requests;
1464 int recvbuf_is_empty=1;
1465 rank = smpi_comm_rank(comm);
1466 size = smpi_comm_size(comm);
1468 smpi_datatype_extent(datatype, &lb, &dataext);
1470 // Send/Recv buffers to/from others;
1471 requests = xbt_new(MPI_Request, size - 1);
1472 tmpbufs = xbt_new(void *, rank);
1474 for(other = 0; other < rank; other++) {
1475 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1477 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1480 for(other = rank + 1; other < size; other++) {
1482 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1485 // Wait for completion of all comms.
1486 smpi_mpi_startall(size - 1, requests);
1487 if(smpi_op_is_commute(op)){
1488 for(other = 0; other < size - 1; other++) {
1489 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1490 if(index == MPI_UNDEFINED) {
1494 if(recvbuf_is_empty){
1495 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1498 // #Request is below rank: it's a irecv
1499 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1503 //non commutative case, wait in order
1504 for(other = 0; other < size - 1; other++) {
1505 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1507 if(recvbuf_is_empty){
1508 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1510 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1514 for(index = 0; index < rank; index++) {
1515 smpi_free_tmp_buffer(tmpbufs[index]);
1517 for(index = 0; index < size-1; index++) {
1518 smpi_mpi_request_free(&requests[index]);