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;
287 request->truncated = 0;
288 request->real_size = 0;
289 request->real_tag = 0;
290 if (flags & PERSISTENT)
291 request->refcount = 1;
293 request->refcount = 0;
294 request->op = MPI_REPLACE;
301 void smpi_empty_status(MPI_Status * status)
303 if(status != MPI_STATUS_IGNORE) {
304 status->MPI_SOURCE = MPI_ANY_SOURCE;
305 status->MPI_TAG = MPI_ANY_TAG;
306 status->MPI_ERROR = MPI_SUCCESS;
311 static void smpi_mpi_request_free_voidp(void* request)
313 MPI_Request req = static_cast<MPI_Request>(request);
314 smpi_mpi_request_free(&req);
317 /* MPI Low level calls */
318 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
319 int dst, int tag, MPI_Comm comm)
321 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
322 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
323 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
327 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
328 int dst, int tag, MPI_Comm comm)
330 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
331 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
332 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
336 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
337 int src, int tag, MPI_Comm comm)
339 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
340 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype,
341 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
342 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
346 void smpi_mpi_start(MPI_Request request)
348 smx_mailbox_t mailbox;
350 xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
351 request->flags &= ~PREPARED;
352 request->flags &= ~FINISHED;
355 if ((request->flags & RECV) != 0) {
356 print_request("New recv", request);
358 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
360 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
361 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
362 xbt_mutex_acquire(mut);
364 if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
365 mailbox = smpi_process_mailbox();
366 } else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh){
367 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
368 //begin with the more appropriate one : the small one.
369 mailbox = smpi_process_mailbox_small();
370 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
371 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
373 if(action == nullptr){
374 mailbox = smpi_process_mailbox();
375 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
376 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
377 if(action == nullptr){
378 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
379 mailbox = smpi_process_mailbox_small();
382 XBT_DEBUG("yes there was something for us in the large mailbox");
385 mailbox = smpi_process_mailbox_small();
386 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
387 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
389 if(action == nullptr){
390 XBT_DEBUG("No, nothing in the permanent receive mailbox");
391 mailbox = smpi_process_mailbox();
393 XBT_DEBUG("yes there was something for us in the small mailbox");
397 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
398 double sleeptime = (request->detached != 0) ? smpi_or(request->size) : 0.0;
400 simcall_process_sleep(sleeptime);
401 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
404 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
405 request->real_size=request->size;
406 request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
407 ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
408 : &smpi_comm_null_copy_buffer_callback, request, -1.0);
409 XBT_DEBUG("recv simcall posted");
411 if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
412 xbt_mutex_release(mut);
414 int receiver = request->dst;
416 int rank = request->src;
417 if (TRACE_smpi_view_internals()) {
418 TRACE_smpi_send(rank, rank, receiver,request->size);
420 print_request("New send", request);
422 //if we are giving back the control to the user without waiting for completion, we have to inject timings
423 double sleeptime = 0.0;
424 if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
425 //isend and send timings may be different
426 sleeptime = ((request->flags & ISEND) != 0)? smpi_ois(request->size) : smpi_os(request->size);
430 simcall_process_sleep(sleeptime);
431 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
434 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
436 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
438 if (async_small_thresh != 0 || (request->flags & RMA) != 0)
439 xbt_mutex_acquire(mut);
441 if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
442 mailbox = smpi_process_remote_mailbox(receiver);
444 else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
445 mailbox = smpi_process_remote_mailbox(receiver);
446 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
447 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
448 if (action == nullptr) {
449 if ((request->flags & SSEND) == 0){
450 mailbox = smpi_process_remote_mailbox_small(receiver);
451 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
454 mailbox = smpi_process_remote_mailbox_small(receiver);
455 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
456 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
457 if (action == nullptr) {
458 XBT_DEBUG("No, we are first, send to large mailbox");
459 mailbox = smpi_process_remote_mailbox(receiver);
464 XBT_DEBUG("Yes there was something for us in the large mailbox");
468 mailbox = smpi_process_remote_mailbox(receiver);
469 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
472 void* buf = request->buf;
473 if ( ((request->flags & SSEND) == 0) && (((request->flags & RMA) != 0) ||
474 (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
476 request->detached = 1;
477 XBT_DEBUG("Send request %p is detached", request);
479 if(request->old_type->sizeof_substruct == 0){
480 oldbuf = request->buf;
481 if (!smpi_process_get_replaying() && oldbuf != NULL && request->size!=0){
482 if((smpi_privatize_global_variables != 0)
483 && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
484 && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
485 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
486 smpi_switch_data_segment(request->src);
488 buf = xbt_malloc(request->size);
489 memcpy(buf,oldbuf,request->size);
490 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
495 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
496 request->real_size=request->size;
497 request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
498 buf, request->real_size, &match_send,
499 &xbt_free_f, // how to free the userdata if a detached send fails
500 !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
501 : &smpi_comm_null_copy_buffer_callback, request,
502 // detach if msg size < eager/rdv switch limit
504 XBT_DEBUG("send simcall posted");
506 /* FIXME: detached sends are not traceable (request->action == nullptr) */
507 if (request->action != nullptr)
508 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
510 if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
511 xbt_mutex_release(mut);
515 void smpi_mpi_startall(int count, MPI_Request * requests)
517 if(requests== nullptr)
520 for(int i = 0; i < count; i++) {
521 smpi_mpi_start(requests[i]);
525 void smpi_mpi_request_free(MPI_Request * request)
527 if((*request) != MPI_REQUEST_NULL){
528 (*request)->refcount--;
529 if((*request)->refcount<0) xbt_die("wrong refcount");
531 if((*request)->refcount==0){
532 smpi_datatype_unuse((*request)->old_type);
533 smpi_comm_unuse((*request)->comm);
534 print_request("Destroying", (*request));
536 *request = MPI_REQUEST_NULL;
538 print_request("Decrementing", (*request));
541 xbt_die("freeing an already free request");
545 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
548 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
550 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, src, dst, tag,
551 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
553 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
554 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
560 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
563 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
565 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
566 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
568 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
569 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
575 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
577 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
578 request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, smpi_process_index(),
579 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
583 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
585 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
586 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
587 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
588 smpi_mpi_start(request);
592 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
594 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
595 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
596 smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
597 smpi_mpi_start(request);
601 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
603 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
604 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
605 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
606 comm, PERSISTENT | RECV | PREPARED);
610 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
612 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
613 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
614 smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
615 NON_PERSISTENT | RECV);
616 smpi_mpi_start(request);
620 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
622 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
623 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
624 smpi_mpi_wait(&request, status);
628 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
630 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
631 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
632 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
634 smpi_mpi_start(request);
635 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
639 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
641 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
642 request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
643 smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
645 smpi_mpi_start(request);
646 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
650 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
651 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
652 MPI_Comm comm, MPI_Status * status)
654 MPI_Request requests[2];
656 int myid=smpi_process_index();
657 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
658 smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
661 requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
662 requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
663 smpi_mpi_startall(2, requests);
664 smpi_mpi_waitall(2, requests, stats);
665 smpi_mpi_request_free(&requests[0]);
666 smpi_mpi_request_free(&requests[1]);
667 if(status != MPI_STATUS_IGNORE) {
668 // Copy receive status
673 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
675 return status->count / smpi_datatype_size(datatype);
678 static void finish_wait(MPI_Request * request, MPI_Status * status)
680 MPI_Request req = *request;
681 smpi_empty_status(status);
683 if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
684 if(status != MPI_STATUS_IGNORE) {
685 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
686 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
687 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
688 status->MPI_ERROR = req->truncated != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
689 // this handles the case were size in receive differs from size in send
690 status->count = req->real_size;
693 print_request("Finishing", req);
694 MPI_Datatype datatype = req->old_type;
696 if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
697 if (!smpi_process_get_replaying()){
698 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
699 && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
700 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
701 smpi_switch_data_segment(smpi_process_index());
705 if(datatype->sizeof_substruct != 0){
706 // This part handles the problem of non-contignous memory the unserialization at the reception
707 s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
708 if(req->flags & RECV)
709 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
710 datatype->substruct, req->op);
712 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
713 int n =req->real_size/smpi_datatype_size(datatype);
714 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
720 if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
721 int rank = smpi_process_index();
722 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
723 TRACE_smpi_recv(rank, src_traced, rank);
726 if(req->detached_sender!=NULL){
727 smpi_mpi_request_free(&(req->detached_sender));
729 if(req->flags & PERSISTENT)
731 req->flags |= FINISHED;
733 smpi_mpi_request_free(request);
736 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
737 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
739 // to avoid deadlocks if used as a break condition, such as
740 // while (MPI_Test(request, flag, status) && flag) {
742 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
743 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
744 static int nsleeps = 1;
745 if(smpi_test_sleep > 0)
746 simcall_process_sleep(nsleeps*smpi_test_sleep);
748 smpi_empty_status(status);
750 if (((*request)->flags & PREPARED) == 0) {
751 if ((*request)->action != NULL)
752 flag = simcall_comm_test((*request)->action);
754 finish_wait(request, status);
755 nsleeps=1;//reset the number of sleeps we will do next time
756 if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
757 *request = MPI_REQUEST_NULL;
765 int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
773 *index = MPI_UNDEFINED;
774 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
775 map = xbt_new(int, count);
776 for(i = 0; i < count; i++) {
777 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
778 xbt_dynar_push(comms, &requests[i]->action);
784 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
785 static int nsleeps = 1;
786 if(smpi_test_sleep > 0)
787 simcall_process_sleep(nsleeps*smpi_test_sleep);
789 i = simcall_comm_testany(comms);
790 // not MPI_UNDEFINED, as this is a simix return code
793 finish_wait(&requests[*index], status);
794 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
795 requests[*index] = MPI_REQUEST_NULL;
802 //all requests are null or inactive, return true
804 smpi_empty_status(status);
807 xbt_dynar_free(&comms);
812 int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
815 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
818 for(i=0; i<count; i++){
819 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
820 if (smpi_mpi_test(&requests[i], pstat)!=1){
823 requests[i]=MPI_REQUEST_NULL;
826 smpi_empty_status(pstat);
828 if(status != MPI_STATUSES_IGNORE) {
835 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
837 //FIXME find another way to avoid busy waiting ?
838 // the issue here is that we have to wait on a nonexistent comm
840 smpi_mpi_iprobe(source, tag, comm, &flag, status);
841 XBT_DEBUG("Busy Waiting on probing : %d", flag);
845 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
847 MPI_Request request = build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
848 smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
850 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
851 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
852 // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
853 static int nsleeps = 1;
854 if(smpi_iprobe_sleep > 0)
855 simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
856 // behave like a receive, but don't do it
857 smx_mailbox_t mailbox;
859 print_request("New iprobe", request);
860 // We have to test both mailboxes as we don't know if we will receive one one or another
861 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
862 mailbox = smpi_process_mailbox_small();
863 XBT_DEBUG("Trying to probe the perm recv mailbox");
864 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
867 if (request->action == nullptr){
868 mailbox = smpi_process_mailbox();
869 XBT_DEBUG("trying to probe the other mailbox");
870 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
873 if (request->action != nullptr){
874 simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
875 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
877 if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
878 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
879 status->MPI_TAG = req->tag;
880 status->MPI_ERROR = MPI_SUCCESS;
881 status->count = req->real_size;
883 nsleeps = 1;//reset the number of sleeps we will do next time
889 smpi_mpi_request_free(&request);
894 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
896 print_request("Waiting", *request);
897 if ((*request)->flags & PREPARED) {
898 smpi_empty_status(status);
902 if ((*request)->action != NULL)
903 // this is not a detached send
904 simcall_comm_wait((*request)->action, -1.0);
906 finish_wait(request, status);
907 if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
908 *request = MPI_REQUEST_NULL;
911 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
916 int index = MPI_UNDEFINED;
920 // Wait for a request to complete
921 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
922 map = xbt_new(int, count);
923 XBT_DEBUG("Wait for one of %d", count);
924 for(i = 0; i < count; i++) {
925 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
926 if (requests[i]->action != NULL) {
927 XBT_DEBUG("Waiting any %p ", requests[i]);
928 xbt_dynar_push(comms, &requests[i]->action);
932 //This is a finished detached request, let's return this one
933 size=0;//so we free the dynar but don't do the waitany call
935 finish_wait(&requests[i], status);//cleanup if refcount = 0
936 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
937 requests[i]=MPI_REQUEST_NULL;//set to null
943 i = simcall_comm_waitany(comms);
945 // not MPI_UNDEFINED, as this is a simix return code
948 finish_wait(&requests[index], status);
949 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
950 requests[index] = MPI_REQUEST_NULL;
954 xbt_dynar_free(&comms);
957 if (index==MPI_UNDEFINED)
958 smpi_empty_status(status);
963 int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
967 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
968 int retvalue = MPI_SUCCESS;
969 //tag invalid requests in the set
970 if (status != MPI_STATUSES_IGNORE) {
971 for (c = 0; c < count; c++) {
972 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
973 smpi_empty_status(&status[c]);
974 } else if (requests[c]->src == MPI_PROC_NULL) {
975 smpi_empty_status(&status[c]);
976 status[c].MPI_SOURCE = MPI_PROC_NULL;
980 for(c = 0; c < count; c++) {
982 if (MC_is_active() || MC_record_replay_is_active()) {
983 smpi_mpi_wait(&requests[c], pstat);
986 index = smpi_mpi_waitany(count, requests, pstat);
987 if (index == MPI_UNDEFINED)
989 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
990 requests[index]=MPI_REQUEST_NULL;
992 if (status != MPI_STATUSES_IGNORE) {
993 status[index] = *pstat;
994 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
995 retvalue = MPI_ERR_IN_STATUS;
1002 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1008 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1010 for(i = 0; i < incount; i++)
1012 index=smpi_mpi_waitany(incount, requests, pstat);
1013 if(index!=MPI_UNDEFINED){
1014 indices[count] = index;
1016 if(status != MPI_STATUSES_IGNORE) {
1017 status[index] = *pstat;
1019 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1020 requests[index]=MPI_REQUEST_NULL;
1022 return MPI_UNDEFINED;
1028 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1034 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1036 for(i = 0; i < incount; i++) {
1037 if((requests[i] != MPI_REQUEST_NULL)) {
1038 if(smpi_mpi_test(&requests[i], pstat)) {
1041 if(status != MPI_STATUSES_IGNORE) {
1044 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1045 requests[i]=MPI_REQUEST_NULL;
1051 if(count_dead==incount)
1052 return MPI_UNDEFINED;
1056 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
1058 smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm);
1061 void smpi_mpi_barrier(MPI_Comm comm)
1063 smpi_coll_tuned_barrier_ompi_basic_linear(comm);
1066 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1067 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1069 int system_tag = COLL_TAG_GATHER;
1070 int rank, size, src, index;
1071 MPI_Aint lb = 0, recvext = 0;
1072 MPI_Request *requests;
1074 rank = smpi_comm_rank(comm);
1075 size = smpi_comm_size(comm);
1077 // Send buffer to root
1078 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1080 smpi_datatype_extent(recvtype, &lb, &recvext);
1081 // Local copy from root
1082 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext, recvcount, recvtype);
1083 // Receive buffers from senders
1084 requests = xbt_new(MPI_Request, size - 1);
1086 for(src = 0; src < size; src++) {
1088 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + src * recvcount * recvext, recvcount, recvtype,
1089 src, system_tag, comm);
1093 // Wait for completion of irecv's.
1094 smpi_mpi_startall(size - 1, requests);
1095 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1096 for(src = 0; src < size-1; src++) {
1097 smpi_mpi_request_free(&requests[src]);
1103 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
1108 int rank = smpi_comm_rank(comm);
1111 /* arbitrarily choose root as rank 0 */
1112 size = smpi_comm_size(comm);
1114 displs = xbt_new(int, size);
1115 for (i = 0; i < size; i++) {
1117 count += recvcounts[i];
1119 tmpbuf=static_cast<void*>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1121 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1122 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm);
1124 smpi_free_tmp_buffer(tmpbuf);
1127 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs,
1128 MPI_Datatype recvtype, int root, MPI_Comm comm)
1130 int system_tag = COLL_TAG_GATHERV;
1131 int rank, size, src, index;
1132 MPI_Aint lb = 0, recvext = 0;
1133 MPI_Request *requests;
1135 rank = smpi_comm_rank(comm);
1136 size = smpi_comm_size(comm);
1138 // Send buffer to root
1139 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1141 smpi_datatype_extent(recvtype, &lb, &recvext);
1142 // Local copy from root
1143 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
1144 recvcounts[root], recvtype);
1145 // Receive buffers from senders
1146 requests = xbt_new(MPI_Request, size - 1);
1148 for(src = 0; src < size; src++) {
1150 requests[index] = smpi_irecv_init(static_cast<char*>(recvbuf) + displs[src] * recvext,
1151 recvcounts[src], recvtype, src, system_tag, comm);
1155 // Wait for completion of irecv's.
1156 smpi_mpi_startall(size - 1, requests);
1157 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1158 for(src = 0; src < size-1; src++) {
1159 smpi_mpi_request_free(&requests[src]);
1165 void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1166 void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
1168 int system_tag = COLL_TAG_ALLGATHER;
1169 int rank, size, other, index;
1170 MPI_Aint lb = 0, recvext = 0;
1171 MPI_Request *requests;
1173 rank = smpi_comm_rank(comm);
1174 size = smpi_comm_size(comm);
1175 // FIXME: check for errors
1176 smpi_datatype_extent(recvtype, &lb, &recvext);
1177 // Local copy from self
1178 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
1179 // Send/Recv buffers to/from others;
1180 requests = xbt_new(MPI_Request, 2 * (size - 1));
1182 for(other = 0; other < size; other++) {
1184 requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm);
1186 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other,
1191 // Wait for completion of all comms.
1192 smpi_mpi_startall(2 * (size - 1), requests);
1193 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1194 for(other = 0; other < 2*(size-1); other++) {
1195 smpi_mpi_request_free(&requests[other]);
1200 void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
1201 int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm)
1203 int system_tag = COLL_TAG_ALLGATHERV;
1204 int rank, size, other, index;
1205 MPI_Aint lb = 0, recvext = 0;
1206 MPI_Request *requests;
1208 rank = smpi_comm_rank(comm);
1209 size = smpi_comm_size(comm);
1210 smpi_datatype_extent(recvtype, &lb, &recvext);
1211 // Local copy from self
1212 smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype);
1213 // Send buffers to others;
1214 requests = xbt_new(MPI_Request, 2 * (size - 1));
1216 for(other = 0; other < size; other++) {
1219 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
1221 requests[index] = smpi_irecv_init(static_cast<char *>(recvbuf) + displs[other] * recvext, recvcounts[other],
1222 recvtype, other, system_tag, comm);
1226 // Wait for completion of all comms.
1227 smpi_mpi_startall(2 * (size - 1), requests);
1228 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1229 for(other = 0; other < 2*(size-1); other++) {
1230 smpi_mpi_request_free(&requests[other]);
1235 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1236 void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
1238 int system_tag = COLL_TAG_SCATTER;
1239 int rank, size, dst, index;
1240 MPI_Aint lb = 0, sendext = 0;
1241 MPI_Request *requests;
1243 rank = smpi_comm_rank(comm);
1244 size = smpi_comm_size(comm);
1246 // Recv buffer from root
1247 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1249 smpi_datatype_extent(sendtype, &lb, &sendext);
1250 // Local copy from root
1251 if(recvbuf!=MPI_IN_PLACE){
1252 smpi_datatype_copy(static_cast<char *>(sendbuf) + root * sendcount * sendext,
1253 sendcount, sendtype, recvbuf, recvcount, recvtype);
1255 // Send buffers to receivers
1256 requests = xbt_new(MPI_Request, size - 1);
1258 for(dst = 0; dst < size; dst++) {
1260 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst,
1265 // Wait for completion of isend's.
1266 smpi_mpi_startall(size - 1, requests);
1267 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1268 for(dst = 0; dst < size-1; dst++) {
1269 smpi_mpi_request_free(&requests[dst]);
1275 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount,
1276 MPI_Datatype recvtype, int root, MPI_Comm comm)
1278 int system_tag = COLL_TAG_SCATTERV;
1279 int rank, size, dst, index;
1280 MPI_Aint lb = 0, sendext = 0;
1281 MPI_Request *requests;
1283 rank = smpi_comm_rank(comm);
1284 size = smpi_comm_size(comm);
1286 // Recv buffer from root
1287 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
1289 smpi_datatype_extent(sendtype, &lb, &sendext);
1290 // Local copy from root
1291 if(recvbuf!=MPI_IN_PLACE){
1292 smpi_datatype_copy(static_cast<char *>(sendbuf) + displs[root] * sendext, sendcounts[root],
1293 sendtype, recvbuf, recvcount, recvtype);
1295 // Send buffers to receivers
1296 requests = xbt_new(MPI_Request, size - 1);
1298 for(dst = 0; dst < size; dst++) {
1300 requests[index] = smpi_isend_init(static_cast<char *>(sendbuf) + displs[dst] * sendext, sendcounts[dst],
1301 sendtype, dst, system_tag, comm);
1305 // Wait for completion of isend's.
1306 smpi_mpi_startall(size - 1, requests);
1307 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1308 for(dst = 0; dst < size-1; dst++) {
1309 smpi_mpi_request_free(&requests[dst]);
1315 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
1318 int system_tag = COLL_TAG_REDUCE;
1319 int rank, size, src, index;
1320 MPI_Aint lb = 0, dataext = 0;
1321 MPI_Request *requests;
1324 char* sendtmpbuf = static_cast<char *>(sendbuf);
1327 rank = smpi_comm_rank(comm);
1328 size = smpi_comm_size(comm);
1329 //non commutative case, use a working algo from openmpi
1330 if(!smpi_op_is_commute(op)){
1331 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
1335 if( sendbuf == MPI_IN_PLACE ) {
1336 sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)));
1337 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1341 // Send buffer to root
1342 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1344 smpi_datatype_extent(datatype, &lb, &dataext);
1345 // Local copy from root
1346 if (sendtmpbuf != NULL && recvbuf != NULL)
1347 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1348 // Receive buffers from senders
1349 requests = xbt_new(MPI_Request, size - 1);
1350 tmpbufs = xbt_new(void *, size - 1);
1352 for(src = 0; src < size; src++) {
1354 if (!smpi_process_get_replaying())
1355 tmpbufs[index] = xbt_malloc(count * dataext);
1357 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1359 smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
1363 // Wait for completion of irecv's.
1364 smpi_mpi_startall(size - 1, requests);
1365 for(src = 0; src < size - 1; src++) {
1366 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1367 XBT_DEBUG("finished waiting any request with index %d", index);
1368 if(index == MPI_UNDEFINED) {
1371 smpi_mpi_request_free(&requests[index]);
1373 if(op) /* op can be MPI_OP_NULL that does nothing */
1374 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1376 for(index = 0; index < size - 1; index++) {
1377 smpi_free_tmp_buffer(tmpbufs[index]);
1383 if( sendbuf == MPI_IN_PLACE ) {
1384 smpi_free_tmp_buffer(sendtmpbuf);
1388 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1390 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1391 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1394 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1396 int system_tag = -888;
1397 int rank, size, other, index;
1398 MPI_Aint lb = 0, dataext = 0;
1399 MPI_Request *requests;
1402 rank = smpi_comm_rank(comm);
1403 size = smpi_comm_size(comm);
1405 smpi_datatype_extent(datatype, &lb, &dataext);
1407 // Local copy from self
1408 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1410 // Send/Recv buffers to/from others;
1411 requests = xbt_new(MPI_Request, size - 1);
1412 tmpbufs = xbt_new(void *, rank);
1414 for(other = 0; other < rank; other++) {
1415 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1416 requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1419 for(other = rank + 1; other < size; other++) {
1420 requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1423 // Wait for completion of all comms.
1424 smpi_mpi_startall(size - 1, requests);
1426 if(smpi_op_is_commute(op)){
1427 for(other = 0; other < size - 1; other++) {
1428 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1429 if(index == MPI_UNDEFINED) {
1433 // #Request is below rank: it's a irecv
1434 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1438 //non commutative case, wait in order
1439 for(other = 0; other < size - 1; other++) {
1440 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1442 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1446 for(index = 0; index < rank; index++) {
1447 smpi_free_tmp_buffer(tmpbufs[index]);
1449 for(index = 0; index < size-1; index++) {
1450 smpi_mpi_request_free(&requests[index]);
1456 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1458 int system_tag = -888;
1459 int rank, size, other, index;
1460 MPI_Aint lb = 0, dataext = 0;
1461 MPI_Request *requests;
1463 int recvbuf_is_empty=1;
1464 rank = smpi_comm_rank(comm);
1465 size = smpi_comm_size(comm);
1467 smpi_datatype_extent(datatype, &lb, &dataext);
1469 // Send/Recv buffers to/from others;
1470 requests = xbt_new(MPI_Request, size - 1);
1471 tmpbufs = xbt_new(void *, rank);
1473 for(other = 0; other < rank; other++) {
1474 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1476 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
1479 for(other = rank + 1; other < size; other++) {
1481 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1484 // Wait for completion of all comms.
1485 smpi_mpi_startall(size - 1, requests);
1486 if(smpi_op_is_commute(op)){
1487 for(other = 0; other < size - 1; other++) {
1488 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1489 if(index == MPI_UNDEFINED) {
1493 if(recvbuf_is_empty){
1494 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1497 // #Request is below rank: it's a irecv
1498 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1502 //non commutative case, wait in order
1503 for(other = 0; other < size - 1; other++) {
1504 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1506 if(recvbuf_is_empty){
1507 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1509 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1513 for(index = 0; index < rank; index++) {
1514 smpi_free_tmp_buffer(tmpbufs[index]);
1516 for(index = 0; index < size-1; index++) {
1517 smpi_mpi_request_free(&requests[index]);