1 /* Copyright (c) 2007-2014. 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. */
10 #include "xbt/replay.h"
12 #include "simix/smx_private.h"
13 #include "surf/surf.h"
14 #include "simgrid/sg_config.h"
15 #include "colls/colls.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
20 static int match_recv(void* a, void* b, smx_synchro_t ignored) {
21 MPI_Request ref = (MPI_Request)a;
22 MPI_Request req = (MPI_Request)b;
23 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
25 xbt_assert(ref, "Cannot match recv against null reference");
26 xbt_assert(req, "Cannot match recv against null request");
27 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
28 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
29 //we match, we can transfer some values
30 // FIXME : move this to the copy function ?
31 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
32 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
33 if(ref->real_size < req->real_size) ref->truncated = 1;
35 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
37 XBT_DEBUG("match succeeded");
42 static int match_send(void* a, void* b,smx_synchro_t ignored) {
43 MPI_Request ref = (MPI_Request)a;
44 MPI_Request req = (MPI_Request)b;
45 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
46 xbt_assert(ref, "Cannot match send against null reference");
47 xbt_assert(req, "Cannot match send against null request");
49 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
50 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
52 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
53 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
54 if(req->real_size < ref->real_size) req->truncated = 1;
56 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
58 XBT_DEBUG("match succeeded");
64 typedef struct s_smpi_factor *smpi_factor_t;
65 typedef struct s_smpi_factor {
68 double values[4];//arbitrary set to 4
70 xbt_dynar_t smpi_os_values = NULL;
71 xbt_dynar_t smpi_or_values = NULL;
72 xbt_dynar_t smpi_ois_values = NULL;
74 double smpi_wtime_sleep = 0.0;
75 double smpi_iprobe_sleep = 1e-4;
76 double smpi_test_sleep = 1e-4;
79 // Methods used to parse and store the values for timing injections in smpi
80 // These are taken from surf/network.c and generalized to have more factors
81 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
83 static int factor_cmp(const void *pa, const void *pb)
85 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
86 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
90 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
93 unsigned int iter = 0;
97 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
99 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
100 radical_elements = xbt_str_split(smpi_coef_string, ";");
101 xbt_dynar_foreach(radical_elements, iter, value) {
102 memset(&fact, 0, sizeof(s_smpi_factor_t));
103 radical_elements2 = xbt_str_split(value, ":");
104 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
105 xbt_die("Malformed radical for smpi factor!");
106 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
108 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
110 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
115 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
116 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
117 xbt_dynar_free(&radical_elements2);
119 xbt_dynar_free(&radical_elements);
121 xbt_dynar_sort(smpi_factor, &factor_cmp);
122 xbt_dynar_foreach(smpi_factor, iter, fact) {
123 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
128 static double smpi_os(double size)
130 if (!smpi_os_values) {
131 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
132 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
134 unsigned int iter = 0;
135 s_smpi_factor_t fact;
137 xbt_dynar_foreach(smpi_os_values, iter, fact) {
138 if (size <= fact.factor) {
139 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
142 current=fact.values[0]+fact.values[1]*size;
145 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
150 static double smpi_ois(double size)
152 if (!smpi_ois_values) {
153 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
154 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
156 unsigned int iter = 0;
157 s_smpi_factor_t fact;
159 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
160 if (size <= fact.factor) {
161 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
164 current=fact.values[0]+fact.values[1]*size;
167 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
172 static double smpi_or(double size)
174 if (!smpi_or_values) {
175 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
176 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
178 unsigned int iter = 0;
179 s_smpi_factor_t fact;
181 xbt_dynar_foreach(smpi_or_values, iter, fact) {
182 if (size <= fact.factor) {
183 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
186 current=fact.values[0]+fact.values[1]*size;
188 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
193 double smpi_mpi_wtime(){
195 if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
197 time = SIMIX_get_clock();
198 //to avoid deadlocks if called too many times
199 if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep);
202 time = SIMIX_get_clock();
207 static MPI_Request build_request(void *buf, int count,
208 MPI_Datatype datatype, int src, int dst,
209 int tag, MPI_Comm comm, unsigned flags)
211 MPI_Request request = NULL;
213 void *old_buf = NULL;
215 request = xbt_new(s_smpi_mpi_request_t, 1);
217 s_smpi_subtype_t *subtype = datatype->substruct;
219 if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
220 // This part handles the problem of non-contiguous memory
222 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
223 if ((datatype->has_subtype == 1) && (flags & SEND)) {
224 subtype->serialize(old_buf, buf, count, datatype->substruct);
229 // This part handles the problem of non-contiguous memory (for the
230 // unserialisation at the reception)
231 request->old_buf = old_buf;
232 request->old_type = datatype;
234 request->size = smpi_datatype_size(datatype) * count;
238 request->comm = comm;
239 request->action = NULL;
240 request->flags = flags;
241 request->detached = 0;
242 request->detached_sender = NULL;
243 request->real_src = 0;
245 request->truncated = 0;
246 request->real_size = 0;
247 request->real_tag = 0;
248 if(flags & PERSISTENT)
249 request->refcount = 1;
251 request->refcount = 0;
252 request->op = MPI_REPLACE;
257 if (flags & SEND) smpi_datatype_unuse(datatype);
263 void smpi_empty_status(MPI_Status * status)
265 if(status != MPI_STATUS_IGNORE) {
266 status->MPI_SOURCE = MPI_ANY_SOURCE;
267 status->MPI_TAG = MPI_ANY_TAG;
268 status->MPI_ERROR = MPI_SUCCESS;
273 void smpi_action_trace_run(char *path)
277 xbt_dict_cursor_t cursor;
281 action_fp = fopen(path, "r");
282 if (action_fp == NULL)
283 xbt_die("Cannot open %s: %s", path, strerror(errno));
286 if (!xbt_dict_is_empty(action_queues)) {
288 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
291 xbt_dict_foreach(action_queues, cursor, name, todo) {
292 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
298 xbt_dict_free(&action_queues);
299 action_queues = xbt_dict_new_homogeneous(NULL);
302 static void smpi_mpi_request_free_voidp(void* request)
304 MPI_Request req = request;
305 smpi_mpi_request_free(&req);
308 /* MPI Low level calls */
309 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
310 int dst, int tag, MPI_Comm comm)
312 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
313 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
314 comm, PERSISTENT | SEND | PREPARED);
318 MPI_Request smpi_mpi_ssend_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 ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
323 comm, PERSISTENT | SSEND | SEND | PREPARED);
327 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
328 int src, 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 ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
332 comm, PERSISTENT | RECV | PREPARED);
336 void smpi_mpi_start(MPI_Request request)
340 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
341 request->flags &= ~PREPARED;
342 request->flags &= ~FINISHED;
345 if (request->flags & RECV) {
346 print_request("New recv", request);
348 xbt_mutex_t mut=smpi_process_mailboxes_mutex();
349 xbt_mutex_acquire(mut);
351 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
352 //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
353 mailbox = smpi_process_mailbox_small();
354 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
355 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
358 mailbox = smpi_process_mailbox();
359 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
360 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
362 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
363 mailbox = smpi_process_mailbox_small();
366 XBT_DEBUG("yes there was something for us in the large mailbox");
369 mailbox = smpi_process_mailbox_small();
370 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
371 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
374 XBT_DEBUG("No, nothing in the permanent receive mailbox");
375 mailbox = smpi_process_mailbox();
377 XBT_DEBUG("yes there was something for us in the small mailbox");
381 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
382 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
384 simcall_process_sleep(sleeptime);
385 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
388 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
389 request->real_size=request->size;
390 smpi_datatype_use(request->old_type);
391 smpi_comm_use(request->comm);
392 request->action = simcall_comm_irecv(mailbox, request->buf,
393 &request->real_size, &match_recv,
394 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
395 : &smpi_comm_null_copy_buffer_callback,
397 XBT_DEBUG("recv simcall posted");
399 xbt_mutex_release(mut);
403 int receiver = request->dst;
406 int rank = request->src;
407 if (TRACE_smpi_view_internals()) {
408 TRACE_smpi_send(rank, rank, receiver,request->size);
411 print_request("New send", request);
413 //if we are giving back the control to the user without waiting for completion, we have to inject timings
414 double sleeptime = 0.0;
415 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
416 //isend and send timings may be different
417 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
420 if(sleeptime != 0.0){
421 simcall_process_sleep(sleeptime);
422 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
425 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
426 xbt_mutex_acquire(mut);
428 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
429 mailbox = smpi_process_remote_mailbox(receiver);
430 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
431 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
433 if (! (request->flags & SSEND)){
434 mailbox = smpi_process_remote_mailbox_small(receiver);
435 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
437 mailbox = smpi_process_remote_mailbox_small(receiver);
438 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
439 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
441 XBT_DEBUG("No, we are first, send to large mailbox");
442 mailbox = smpi_process_remote_mailbox(receiver);
446 XBT_DEBUG("Yes there was something for us in the large mailbox");
449 mailbox = smpi_process_remote_mailbox(receiver);
450 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
453 void* buf = request->buf;
454 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
456 request->detached = 1;
457 XBT_DEBUG("Send request %p is detached", request);
459 if(request->old_type->has_subtype == 0){
460 oldbuf = request->buf;
461 if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
462 if((smpi_privatize_global_variables)
463 && ((char*)request->buf >= start_data_exe)
464 && ((char*)request->buf < start_data_exe + size_data_exe )){
465 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
466 smpi_switch_data_segment(request->src);
468 buf = xbt_malloc(request->size);
469 memcpy(buf,oldbuf,request->size);
470 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
475 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
476 request->real_size=request->size;
477 smpi_datatype_use(request->old_type);
478 smpi_comm_use(request->comm);
480 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
481 buf, request->real_size,
483 &xbt_free_f, // how to free the userdata if a detached send fails
484 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
485 : &smpi_comm_null_copy_buffer_callback,
487 // detach if msg size < eager/rdv switch limit
489 XBT_DEBUG("send simcall posted");
494 /* FIXME: detached sends are not traceable (request->action == NULL) */
496 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
499 xbt_mutex_release(mut);
504 void smpi_mpi_startall(int count, MPI_Request * requests)
507 if(requests==NULL) return;
509 for(i = 0; i < count; i++) {
510 smpi_mpi_start(requests[i]);
514 void smpi_mpi_request_free(MPI_Request * request)
516 if((*request) != MPI_REQUEST_NULL){
517 (*request)->refcount--;
518 if((*request)->refcount<0) xbt_die("wrong refcount");
520 if((*request)->refcount==0){
521 print_request("Destroying", (*request));
523 *request = MPI_REQUEST_NULL;
525 print_request("Decrementing", (*request));
528 xbt_die("freeing an already free request");
533 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
534 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
536 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
538 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
539 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
541 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
542 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
548 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
549 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
551 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
553 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
554 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
556 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
557 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
564 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
565 int dst, int tag, MPI_Comm comm)
567 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
568 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
569 comm, PERSISTENT | ISEND | SEND | PREPARED);
573 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
574 int dst, int tag, MPI_Comm comm)
576 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
577 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
578 comm, NON_PERSISTENT | ISEND | SEND);
579 smpi_mpi_start(request);
583 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
584 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 ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
588 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
589 smpi_mpi_start(request);
593 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
594 int src, int tag, MPI_Comm comm)
596 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
597 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
598 comm, PERSISTENT | RECV | PREPARED);
602 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
603 int src, int tag, MPI_Comm comm)
605 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
606 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
607 comm, NON_PERSISTENT | RECV);
608 smpi_mpi_start(request);
612 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
613 int tag, MPI_Comm comm, MPI_Status * status)
615 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
616 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
617 smpi_mpi_wait(&request, status);
623 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
624 int tag, MPI_Comm comm)
626 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
627 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
628 comm, NON_PERSISTENT | SEND);
630 smpi_mpi_start(request);
631 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
635 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
636 int dst, int tag, MPI_Comm comm)
638 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
639 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
640 comm, NON_PERSISTENT | SSEND | SEND);
642 smpi_mpi_start(request);
643 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
647 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
648 int dst, int sendtag, void *recvbuf, int recvcount,
649 MPI_Datatype recvtype, int src, int recvtag,
650 MPI_Comm comm, MPI_Status * status)
652 MPI_Request requests[2];
654 int myid=smpi_process_index();
655 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
656 smpi_datatype_copy(sendbuf, sendcount, sendtype,
657 recvbuf, recvcount, recvtype);
661 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
663 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 && req->flags & SEND)
685 && !(req->flags & PREPARED)){
686 if(status != MPI_STATUS_IGNORE) {
687 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
688 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
689 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
690 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
691 // this handles the case were size in receive differs from size in send
692 // FIXME: really this should just contain the count of receive-type blocks,
694 status->count = req->real_size;
697 print_request("Finishing", req);
698 MPI_Datatype datatype = req->old_type;
700 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
701 if (!smpi_process_get_replaying()){
702 if( smpi_privatize_global_variables
703 && ((char*)req->old_buf >= start_data_exe)
704 && ((char*)req->old_buf < start_data_exe + size_data_exe )
706 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
707 smpi_switch_data_segment(smpi_process_index());
711 if(datatype->has_subtype == 1){
712 // This part handles the problem of non-contignous memory
713 // the unserialization at the reception
714 s_smpi_subtype_t *subtype = datatype->substruct;
715 if(req->flags & RECV)
716 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
717 if(req->detached == 0) free(req->buf);
718 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
719 int n =req->real_size/smpi_datatype_size(datatype);
720 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
723 smpi_comm_unuse(req->comm);
724 smpi_datatype_unuse(datatype);
729 if (TRACE_smpi_view_internals()) {
730 if(req->flags & RECV){
731 int rank = smpi_process_index();
732 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
733 TRACE_smpi_recv(rank, src_traced, rank);
738 if(req->detached_sender!=NULL){
739 smpi_mpi_request_free(&(req->detached_sender));
741 if(req->flags & PERSISTENT)
743 req->flags |= FINISHED;
745 smpi_mpi_request_free(request);
749 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
752 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
755 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
756 static int nsleeps = 1;
757 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
759 smpi_empty_status(status);
761 if (!((*request)->flags & PREPARED)) {
762 if ((*request)->action != NULL)
763 flag = simcall_comm_test((*request)->action);
765 finish_wait(request, status);
766 nsleeps=1;//reset the number of sleeps we will do next time
767 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
768 *request = MPI_REQUEST_NULL;
776 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
783 *index = MPI_UNDEFINED;
785 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
786 map = xbt_new(int, count);
788 for(i = 0; i < count; i++) {
789 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
790 !(requests[i]->flags & PREPARED)) {
791 xbt_dynar_push(comms, &requests[i]->action);
797 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
798 static int nsleeps = 1;
799 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
801 i = simcall_comm_testany(comms);
802 // not MPI_UNDEFINED, as this is a simix return code
805 finish_wait(&requests[*index], status);
806 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
807 requests[*index] = MPI_REQUEST_NULL;
814 //all requests are null or inactive, return true
816 smpi_empty_status(status);
819 xbt_dynar_free(&comms);
825 int smpi_mpi_testall(int count, MPI_Request requests[],
829 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
832 for(i=0; i<count; i++){
833 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
834 if (smpi_mpi_test(&requests[i], pstat)!=1){
837 requests[i]=MPI_REQUEST_NULL;
840 smpi_empty_status(pstat);
842 if(status != MPI_STATUSES_IGNORE) {
849 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
851 //FIXME find another wait to avoid busy waiting ?
852 // the issue here is that we have to wait on a nonexistent comm
854 smpi_mpi_iprobe(source, tag, comm, &flag, status);
855 XBT_DEBUG("Busy Waiting on probing : %d", flag);
859 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
861 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
862 comm, PERSISTENT | RECV);
864 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
865 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
866 static int nsleeps = 1;
867 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
868 // behave like a receive, but don't do it
871 print_request("New iprobe", request);
872 // We have to test both mailboxes as we don't know if we will receive one one or another
873 if (sg_cfg_get_int("smpi/async_small_thres")>0){
874 mailbox = smpi_process_mailbox_small();
875 XBT_DEBUG("trying to probe the perm recv mailbox");
876 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
878 if (request->action==NULL){
879 mailbox = smpi_process_mailbox();
880 XBT_DEBUG("trying to probe the other mailbox");
881 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
885 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
887 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
888 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
889 status->MPI_TAG = req->tag;
890 status->MPI_ERROR = MPI_SUCCESS;
891 status->count = req->real_size;
893 nsleeps=1;//reset the number of sleeps we will do next time
899 smpi_mpi_request_free(&request);
904 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
906 print_request("Waiting", *request);
907 if ((*request)->flags & PREPARED) {
908 smpi_empty_status(status);
912 if ((*request)->action != NULL) { // this is not a detached send
913 simcall_comm_wait((*request)->action, -1.0);
915 if((MC_is_active() || MC_record_replay_is_active()) && (*request)->action)
916 (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison
919 finish_wait(request, status);
920 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
921 *request = MPI_REQUEST_NULL;
922 // FIXME for a detached send, finish_wait is not called:
925 int smpi_mpi_waitany(int count, MPI_Request requests[],
932 index = MPI_UNDEFINED;
934 // Wait for a request to complete
935 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
936 map = xbt_new(int, count);
938 XBT_DEBUG("Wait for one of %d", count);
939 for(i = 0; i < count; i++) {
940 if (requests[i] != MPI_REQUEST_NULL
941 && !(requests[i]->flags & PREPARED)
942 && !(requests[i]->flags & FINISHED)) {
943 if (requests[i]->action != NULL) {
944 XBT_DEBUG("Waiting any %p ", requests[i]);
945 xbt_dynar_push(comms, &requests[i]->action);
949 //This is a finished detached request, let's return this one
950 size=0;//so we free the dynar but don't do the waitany call
952 finish_wait(&requests[i], status);//cleanup if refcount = 0
953 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
954 requests[i]=MPI_REQUEST_NULL;//set to null
960 i = simcall_comm_waitany(comms);
962 // not MPI_UNDEFINED, as this is a simix return code
965 finish_wait(&requests[index], status);
966 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
967 requests[index] = MPI_REQUEST_NULL;
971 xbt_dynar_free(&comms);
974 if (index==MPI_UNDEFINED)
975 smpi_empty_status(status);
980 int smpi_mpi_waitall(int count, MPI_Request requests[],
985 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
986 int retvalue = MPI_SUCCESS;
987 //tag invalid requests in the set
988 if (status != MPI_STATUSES_IGNORE) {
989 for (c = 0; c < count; c++) {
990 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
991 (requests[c]->flags & PREPARED)) {
992 smpi_empty_status(&status[c]);
993 } else if (requests[c]->src == MPI_PROC_NULL) {
994 smpi_empty_status(&status[c]);
995 status[c].MPI_SOURCE = MPI_PROC_NULL;
999 for(c = 0; c < count; c++) {
1001 if (MC_is_active() || MC_record_replay_is_active()) {
1002 smpi_mpi_wait(&requests[c], pstat);
1005 index = smpi_mpi_waitany(count, requests, pstat);
1006 if (index == MPI_UNDEFINED)
1008 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1009 requests[index]=MPI_REQUEST_NULL;
1011 if (status != MPI_STATUSES_IGNORE) {
1012 status[index] = *pstat;
1013 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1014 retvalue = MPI_ERR_IN_STATUS;
1021 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
1022 MPI_Status status[])
1024 int i, count, index;
1026 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1029 for(i = 0; i < incount; i++)
1031 index=smpi_mpi_waitany(incount, requests, pstat);
1032 if(index!=MPI_UNDEFINED){
1033 indices[count] = index;
1035 if(status != MPI_STATUSES_IGNORE) {
1036 status[index] = *pstat;
1038 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1039 requests[index]=MPI_REQUEST_NULL;
1041 return MPI_UNDEFINED;
1047 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
1048 MPI_Status status[])
1050 int i, count, count_dead;
1052 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1056 for(i = 0; i < incount; i++) {
1057 if((requests[i] != MPI_REQUEST_NULL)) {
1058 if(smpi_mpi_test(&requests[i], pstat)) {
1061 if(status != MPI_STATUSES_IGNORE) {
1064 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1065 requests[i]=MPI_REQUEST_NULL;
1071 if(count_dead==incount)return MPI_UNDEFINED;
1075 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1078 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1079 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1082 void smpi_mpi_barrier(MPI_Comm comm)
1084 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1085 nary_tree_barrier(comm, 4);
1088 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1089 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1090 int root, MPI_Comm comm)
1092 int system_tag = COLL_TAG_GATHER;
1093 int rank, size, src, index;
1094 MPI_Aint lb = 0, recvext = 0;
1095 MPI_Request *requests;
1097 rank = smpi_comm_rank(comm);
1098 size = smpi_comm_size(comm);
1100 // Send buffer to root
1101 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1103 // FIXME: check for errors
1104 smpi_datatype_extent(recvtype, &lb, &recvext);
1105 // Local copy from root
1106 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1107 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1108 // Receive buffers from senders
1109 requests = xbt_new(MPI_Request, size - 1);
1111 for(src = 0; src < size; src++) {
1113 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1114 recvcount, recvtype,
1115 src, system_tag, comm);
1119 // Wait for completion of irecv's.
1120 smpi_mpi_startall(size - 1, requests);
1121 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1122 for(src = 0; src < size-1; src++) {
1123 smpi_mpi_request_free(&requests[src]);
1130 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1131 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1135 int rank = smpi_process_index();
1138 /* arbitrarily choose root as rank 0 */
1139 size = smpi_comm_size(comm);
1141 displs = xbt_new(int, size);
1142 for (i = 0; i < size; i++) {
1144 count += recvcounts[i];
1146 tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1148 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1149 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1150 recvcounts[rank], datatype, 0, comm);
1152 smpi_free_tmp_buffer(tmpbuf);
1155 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1156 void *recvbuf, int *recvcounts, int *displs,
1157 MPI_Datatype recvtype, int root, MPI_Comm comm)
1159 int system_tag = COLL_TAG_GATHERV;
1160 int rank, size, src, index;
1161 MPI_Aint lb = 0, recvext = 0;
1162 MPI_Request *requests;
1164 rank = smpi_comm_rank(comm);
1165 size = smpi_comm_size(comm);
1167 // Send buffer to root
1168 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1170 // FIXME: check for errors
1171 smpi_datatype_extent(recvtype, &lb, &recvext);
1172 // Local copy from root
1173 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1174 (char *)recvbuf + displs[root] * recvext,
1175 recvcounts[root], recvtype);
1176 // Receive buffers from senders
1177 requests = xbt_new(MPI_Request, size - 1);
1179 for(src = 0; src < size; src++) {
1182 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1183 recvcounts[src], recvtype, src, system_tag, comm);
1187 // Wait for completion of irecv's.
1188 smpi_mpi_startall(size - 1, requests);
1189 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1190 for(src = 0; src < size-1; src++) {
1191 smpi_mpi_request_free(&requests[src]);
1197 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1198 MPI_Datatype sendtype, void *recvbuf,
1199 int recvcount, MPI_Datatype recvtype,
1202 int system_tag = COLL_TAG_ALLGATHER;
1203 int rank, size, other, index;
1204 MPI_Aint lb = 0, recvext = 0;
1205 MPI_Request *requests;
1207 rank = smpi_comm_rank(comm);
1208 size = smpi_comm_size(comm);
1209 // FIXME: check for errors
1210 smpi_datatype_extent(recvtype, &lb, &recvext);
1211 // Local copy from self
1212 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1213 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1215 // Send/Recv buffers to/from others;
1216 requests = xbt_new(MPI_Request, 2 * (size - 1));
1218 for(other = 0; other < size; other++) {
1221 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1224 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1225 recvcount, recvtype, other,
1230 // Wait for completion of all comms.
1231 smpi_mpi_startall(2 * (size - 1), requests);
1232 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1233 for(other = 0; other < 2*(size-1); other++) {
1234 smpi_mpi_request_free(&requests[other]);
1239 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1240 MPI_Datatype sendtype, void *recvbuf,
1241 int *recvcounts, int *displs,
1242 MPI_Datatype recvtype, MPI_Comm comm)
1244 int system_tag = COLL_TAG_ALLGATHERV;
1245 int rank, size, other, index;
1246 MPI_Aint lb = 0, recvext = 0;
1247 MPI_Request *requests;
1249 rank = smpi_comm_rank(comm);
1250 size = smpi_comm_size(comm);
1251 // FIXME: check for errors
1252 smpi_datatype_extent(recvtype, &lb, &recvext);
1253 // Local copy from self
1254 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1255 (char *)recvbuf + displs[rank] * recvext,
1256 recvcounts[rank], recvtype);
1257 // Send buffers to others;
1258 requests = xbt_new(MPI_Request, 2 * (size - 1));
1260 for(other = 0; other < size; other++) {
1263 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1267 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1268 recvtype, other, system_tag, comm);
1272 // Wait for completion of all comms.
1273 smpi_mpi_startall(2 * (size - 1), requests);
1274 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1275 for(other = 0; other < 2*(size-1); other++) {
1276 smpi_mpi_request_free(&requests[other]);
1281 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1282 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1283 int root, MPI_Comm comm)
1285 int system_tag = COLL_TAG_SCATTER;
1286 int rank, size, dst, index;
1287 MPI_Aint lb = 0, sendext = 0;
1288 MPI_Request *requests;
1290 rank = smpi_comm_rank(comm);
1291 size = smpi_comm_size(comm);
1293 // Recv buffer from root
1294 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1297 // FIXME: check for errors
1298 smpi_datatype_extent(sendtype, &lb, &sendext);
1299 // Local copy from root
1300 if(recvbuf!=MPI_IN_PLACE){
1301 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1302 sendcount, sendtype, recvbuf, recvcount, recvtype);
1304 // Send buffers to receivers
1305 requests = xbt_new(MPI_Request, size - 1);
1307 for(dst = 0; dst < size; dst++) {
1309 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1310 sendcount, sendtype, dst,
1315 // Wait for completion of isend's.
1316 smpi_mpi_startall(size - 1, requests);
1317 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1318 for(dst = 0; dst < size-1; dst++) {
1319 smpi_mpi_request_free(&requests[dst]);
1325 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1326 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1327 MPI_Datatype recvtype, int root, MPI_Comm comm)
1329 int system_tag = COLL_TAG_SCATTERV;
1330 int rank, size, dst, index;
1331 MPI_Aint lb = 0, sendext = 0;
1332 MPI_Request *requests;
1334 rank = smpi_comm_rank(comm);
1335 size = smpi_comm_size(comm);
1337 // Recv buffer from root
1338 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1341 // FIXME: check for errors
1342 smpi_datatype_extent(sendtype, &lb, &sendext);
1343 // Local copy from root
1344 if(recvbuf!=MPI_IN_PLACE){
1345 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1346 sendtype, recvbuf, recvcount, recvtype);
1348 // Send buffers to receivers
1349 requests = xbt_new(MPI_Request, size - 1);
1351 for(dst = 0; dst < size; dst++) {
1354 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1355 sendtype, dst, system_tag, comm);
1359 // Wait for completion of isend's.
1360 smpi_mpi_startall(size - 1, requests);
1361 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1362 for(dst = 0; dst < size-1; dst++) {
1363 smpi_mpi_request_free(&requests[dst]);
1369 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1370 MPI_Datatype datatype, MPI_Op op, int root,
1373 int system_tag = COLL_TAG_REDUCE;
1374 int rank, size, src, index;
1375 MPI_Aint lb = 0, dataext = 0;
1376 MPI_Request *requests;
1380 char* sendtmpbuf = (char*) sendbuf;
1381 if( sendbuf == MPI_IN_PLACE ) {
1382 sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1383 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1386 rank = smpi_comm_rank(comm);
1387 size = smpi_comm_size(comm);
1388 //non commutative case, use a working algo from openmpi
1389 if(!smpi_op_is_commute(op)){
1390 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1391 datatype, op, root, comm);
1396 // Send buffer to root
1397 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1399 // FIXME: check for errors
1400 smpi_datatype_extent(datatype, &lb, &dataext);
1401 // Local copy from root
1402 if (sendtmpbuf && recvbuf)
1403 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1404 // Receive buffers from senders
1405 //TODO: make a MPI_barrier here ?
1406 requests = xbt_new(MPI_Request, size - 1);
1407 tmpbufs = xbt_new(void *, size - 1);
1409 for(src = 0; src < size; src++) {
1411 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1413 if (!smpi_process_get_replaying())
1414 tmpbufs[index] = xbt_malloc(count * dataext);
1416 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1418 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1423 // Wait for completion of irecv's.
1424 smpi_mpi_startall(size - 1, requests);
1425 for(src = 0; src < size - 1; src++) {
1426 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1427 XBT_DEBUG("finished waiting any request with index %d", index);
1428 if(index == MPI_UNDEFINED) {
1431 smpi_mpi_request_free(&requests[index]);
1433 if(op) /* op can be MPI_OP_NULL that does nothing */
1434 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1436 for(index = 0; index < size - 1; index++) {
1437 smpi_free_tmp_buffer(tmpbufs[index]);
1442 if( sendbuf == MPI_IN_PLACE ) {
1443 smpi_free_tmp_buffer(sendtmpbuf);
1448 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1449 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1451 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1452 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1455 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1456 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;
1464 rank = smpi_comm_rank(comm);
1465 size = smpi_comm_size(comm);
1467 // FIXME: check for errors
1468 smpi_datatype_extent(datatype, &lb, &dataext);
1470 // Local copy from self
1471 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1473 // Send/Recv buffers to/from others;
1474 requests = xbt_new(MPI_Request, size - 1);
1475 tmpbufs = xbt_new(void *, rank);
1477 for(other = 0; other < rank; other++) {
1478 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1480 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1482 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1486 for(other = rank + 1; other < size; other++) {
1488 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1491 // Wait for completion of all comms.
1492 smpi_mpi_startall(size - 1, requests);
1494 if(smpi_op_is_commute(op)){
1495 for(other = 0; other < size - 1; other++) {
1496 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1497 if(index == MPI_UNDEFINED) {
1501 // #Request is below rank: it's a irecv
1502 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1506 //non commutative case, wait in order
1507 for(other = 0; other < size - 1; other++) {
1508 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1510 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]);
1524 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1525 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1527 int system_tag = -888;
1528 int rank, size, other, index;
1529 MPI_Aint lb = 0, dataext = 0;
1530 MPI_Request *requests;
1532 int recvbuf_is_empty=1;
1533 rank = smpi_comm_rank(comm);
1534 size = smpi_comm_size(comm);
1536 // FIXME: check for errors
1537 smpi_datatype_extent(datatype, &lb, &dataext);
1539 // Send/Recv buffers to/from others;
1540 requests = xbt_new(MPI_Request, size - 1);
1541 tmpbufs = xbt_new(void *, rank);
1543 for(other = 0; other < rank; other++) {
1544 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1546 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1548 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1552 for(other = rank + 1; other < size; other++) {
1554 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1557 // Wait for completion of all comms.
1558 smpi_mpi_startall(size - 1, requests);
1559 if(smpi_op_is_commute(op)){
1560 for(other = 0; other < size - 1; other++) {
1561 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1562 if(index == MPI_UNDEFINED) {
1566 if(recvbuf_is_empty){
1567 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1570 // #Request is below rank: it's a irecv
1571 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1575 //non commutative case, wait in order
1576 for(other = 0; other < size - 1; other++) {
1577 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1579 if(recvbuf_is_empty){
1580 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1582 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1586 for(index = 0; index < rank; index++) {
1587 smpi_free_tmp_buffer(tmpbufs[index]);
1589 for(index = 0; index < size-1; index++) {
1590 smpi_mpi_request_free(&requests[index]);