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;
255 if (flags & SEND) smpi_datatype_unuse(datatype);
261 void smpi_empty_status(MPI_Status * status)
263 if(status != MPI_STATUS_IGNORE) {
264 status->MPI_SOURCE = MPI_ANY_SOURCE;
265 status->MPI_TAG = MPI_ANY_TAG;
266 status->MPI_ERROR = MPI_SUCCESS;
271 void smpi_action_trace_run(char *path)
275 xbt_dict_cursor_t cursor;
279 action_fp = fopen(path, "r");
280 if (action_fp == NULL)
281 xbt_die("Cannot open %s: %s", path, strerror(errno));
284 if (!xbt_dict_is_empty(action_queues)) {
286 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
289 xbt_dict_foreach(action_queues, cursor, name, todo) {
290 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
296 xbt_dict_free(&action_queues);
297 action_queues = xbt_dict_new_homogeneous(NULL);
300 static void smpi_mpi_request_free_voidp(void* request)
302 MPI_Request req = request;
303 smpi_mpi_request_free(&req);
306 /* MPI Low level calls */
307 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
308 int dst, int tag, MPI_Comm comm)
310 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
311 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
312 comm, PERSISTENT | SEND | PREPARED);
316 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
317 int dst, int tag, MPI_Comm comm)
319 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
320 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
321 comm, PERSISTENT | SSEND | SEND | PREPARED);
325 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
326 int src, int tag, MPI_Comm comm)
328 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
329 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,
330 comm, PERSISTENT | RECV | PREPARED);
334 void smpi_mpi_start(MPI_Request request)
338 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
339 request->flags &= ~PREPARED;
340 request->flags &= ~FINISHED;
343 if (request->flags & RECV) {
344 print_request("New recv", request);
346 xbt_mutex_t mut=smpi_process_mailboxes_mutex();
347 xbt_mutex_acquire(mut);
349 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
350 //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
351 mailbox = smpi_process_mailbox_small();
352 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
353 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
356 mailbox = smpi_process_mailbox();
357 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
358 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
360 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
361 mailbox = smpi_process_mailbox_small();
364 XBT_DEBUG("yes there was something for us in the large mailbox");
367 mailbox = smpi_process_mailbox_small();
368 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
369 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
372 XBT_DEBUG("No, nothing in the permanent receive mailbox");
373 mailbox = smpi_process_mailbox();
375 XBT_DEBUG("yes there was something for us in the small mailbox");
379 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
380 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
382 simcall_process_sleep(sleeptime);
383 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
386 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
387 request->real_size=request->size;
388 smpi_datatype_use(request->old_type);
389 smpi_comm_use(request->comm);
390 request->action = simcall_comm_irecv(mailbox, request->buf,
391 &request->real_size, &match_recv,
392 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
393 : &smpi_comm_null_copy_buffer_callback,
395 XBT_DEBUG("recv simcall posted");
397 xbt_mutex_release(mut);
401 int receiver = request->dst;
403 int rank = request->src;
404 if (TRACE_smpi_view_internals()) {
405 TRACE_smpi_send(rank, rank, receiver,request->size);
407 print_request("New send", request);
409 //if we are giving back the control to the user without waiting for completion, we have to inject timings
410 double sleeptime = 0.0;
411 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
412 //isend and send timings may be different
413 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
416 if(sleeptime != 0.0){
417 simcall_process_sleep(sleeptime);
418 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
421 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
422 xbt_mutex_acquire(mut);
424 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
425 mailbox = smpi_process_remote_mailbox(receiver);
426 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
427 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
429 if (! (request->flags & SSEND)){
430 mailbox = smpi_process_remote_mailbox_small(receiver);
431 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
433 mailbox = smpi_process_remote_mailbox_small(receiver);
434 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
435 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
437 XBT_DEBUG("No, we are first, send to large mailbox");
438 mailbox = smpi_process_remote_mailbox(receiver);
442 XBT_DEBUG("Yes there was something for us in the large mailbox");
445 mailbox = smpi_process_remote_mailbox(receiver);
446 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
449 void* buf = request->buf;
450 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
452 request->detached = 1;
453 XBT_DEBUG("Send request %p is detached", request);
455 if(request->old_type->has_subtype == 0){
456 oldbuf = request->buf;
457 if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
458 if((smpi_privatize_global_variables)
459 && ((char*) request->buf >= smpi_start_data_exe)
460 && ((char*)request->buf < smpi_start_data_exe + smpi_size_data_exe )){
461 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
462 smpi_switch_data_segment(request->src);
464 buf = xbt_malloc(request->size);
465 memcpy(buf,oldbuf,request->size);
466 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
471 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
472 request->real_size=request->size;
473 smpi_datatype_use(request->old_type);
474 smpi_comm_use(request->comm);
476 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
477 buf, request->real_size,
479 &xbt_free_f, // how to free the userdata if a detached send fails
480 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
481 : &smpi_comm_null_copy_buffer_callback,
483 // detach if msg size < eager/rdv switch limit
485 XBT_DEBUG("send simcall posted");
489 /* FIXME: detached sends are not traceable (request->action == NULL) */
491 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
493 xbt_mutex_release(mut);
498 void smpi_mpi_startall(int count, MPI_Request * requests)
501 if(requests==NULL) return;
503 for(i = 0; i < count; i++) {
504 smpi_mpi_start(requests[i]);
508 void smpi_mpi_request_free(MPI_Request * request)
510 if((*request) != MPI_REQUEST_NULL){
511 (*request)->refcount--;
512 if((*request)->refcount<0) xbt_die("wrong refcount");
514 if((*request)->refcount==0){
515 print_request("Destroying", (*request));
517 *request = MPI_REQUEST_NULL;
519 print_request("Decrementing", (*request));
522 xbt_die("freeing an already free request");
527 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
528 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
530 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
532 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
533 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
535 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
536 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
542 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
543 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
545 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
547 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
548 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
550 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
551 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
558 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
559 int dst, int tag, MPI_Comm comm)
561 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
562 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
563 comm, PERSISTENT | ISEND | SEND | PREPARED);
567 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
568 int dst, int tag, MPI_Comm comm)
570 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
571 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
572 comm, NON_PERSISTENT | ISEND | SEND);
573 smpi_mpi_start(request);
577 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
578 int dst, int tag, MPI_Comm comm)
580 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
581 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
582 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
583 smpi_mpi_start(request);
587 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
588 int src, int tag, MPI_Comm comm)
590 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
591 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,
592 comm, PERSISTENT | RECV | PREPARED);
596 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
597 int src, int tag, MPI_Comm comm)
599 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
600 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,
601 comm, NON_PERSISTENT | RECV);
602 smpi_mpi_start(request);
606 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
607 int tag, MPI_Comm comm, MPI_Status * status)
609 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
610 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
611 smpi_mpi_wait(&request, status);
617 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
618 int tag, MPI_Comm comm)
620 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
621 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
622 comm, NON_PERSISTENT | SEND);
624 smpi_mpi_start(request);
625 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
629 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
630 int dst, int tag, MPI_Comm comm)
632 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
633 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
634 comm, NON_PERSISTENT | SSEND | SEND);
636 smpi_mpi_start(request);
637 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
641 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
642 int dst, int sendtag, void *recvbuf, int recvcount,
643 MPI_Datatype recvtype, int src, int recvtag,
644 MPI_Comm comm, MPI_Status * status)
646 MPI_Request requests[2];
648 int myid=smpi_process_index();
649 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
650 smpi_datatype_copy(sendbuf, sendcount, sendtype,
651 recvbuf, recvcount, recvtype);
655 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
657 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
658 smpi_mpi_startall(2, requests);
659 smpi_mpi_waitall(2, requests, stats);
660 smpi_mpi_request_free(&requests[0]);
661 smpi_mpi_request_free(&requests[1]);
662 if(status != MPI_STATUS_IGNORE) {
663 // Copy receive status
668 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
670 return status->count / smpi_datatype_size(datatype);
673 static void finish_wait(MPI_Request * request, MPI_Status * status)
675 MPI_Request req = *request;
676 smpi_empty_status(status);
678 if(!(req->detached && req->flags & SEND)
679 && !(req->flags & PREPARED)){
680 if(status != MPI_STATUS_IGNORE) {
681 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
682 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
683 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
684 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
685 // this handles the case were size in receive differs from size in send
686 // FIXME: really this should just contain the count of receive-type blocks,
688 status->count = req->real_size;
691 print_request("Finishing", req);
692 MPI_Datatype datatype = req->old_type;
694 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
695 if (!smpi_process_get_replaying()){
696 if( smpi_privatize_global_variables
697 && ((char*)req->old_buf >= smpi_start_data_exe)
698 && ((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->has_subtype == 1){
706 // This part handles the problem of non-contignous memory
707 // the unserialization at the reception
708 s_smpi_subtype_t *subtype = datatype->substruct;
709 if(req->flags & RECV)
710 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
711 if(req->detached == 0) free(req->buf);
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);
717 smpi_comm_unuse(req->comm);
718 smpi_datatype_unuse(datatype);
722 if (TRACE_smpi_view_internals()) {
723 if(req->flags & RECV){
724 int rank = smpi_process_index();
725 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
726 TRACE_smpi_recv(rank, src_traced, rank);
730 if(req->detached_sender!=NULL){
731 smpi_mpi_request_free(&(req->detached_sender));
733 if(req->flags & PERSISTENT)
735 req->flags |= FINISHED;
737 smpi_mpi_request_free(request);
741 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
744 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
747 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
748 static int nsleeps = 1;
749 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
751 smpi_empty_status(status);
753 if (!((*request)->flags & PREPARED)) {
754 if ((*request)->action != NULL)
755 flag = simcall_comm_test((*request)->action);
757 finish_wait(request, status);
758 nsleeps=1;//reset the number of sleeps we will do next time
759 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
760 *request = MPI_REQUEST_NULL;
768 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
775 *index = MPI_UNDEFINED;
777 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
778 map = xbt_new(int, count);
780 for(i = 0; i < count; i++) {
781 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
782 !(requests[i]->flags & PREPARED)) {
783 xbt_dynar_push(comms, &requests[i]->action);
789 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
790 static int nsleeps = 1;
791 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
793 i = simcall_comm_testany(comms);
794 // not MPI_UNDEFINED, as this is a simix return code
797 finish_wait(&requests[*index], status);
798 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
799 requests[*index] = MPI_REQUEST_NULL;
806 //all requests are null or inactive, return true
808 smpi_empty_status(status);
811 xbt_dynar_free(&comms);
817 int smpi_mpi_testall(int count, MPI_Request requests[],
821 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
824 for(i=0; i<count; i++){
825 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
826 if (smpi_mpi_test(&requests[i], pstat)!=1){
829 requests[i]=MPI_REQUEST_NULL;
832 smpi_empty_status(pstat);
834 if(status != MPI_STATUSES_IGNORE) {
841 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
843 //FIXME find another wait to avoid busy waiting ?
844 // the issue here is that we have to wait on a nonexistent comm
846 smpi_mpi_iprobe(source, tag, comm, &flag, status);
847 XBT_DEBUG("Busy Waiting on probing : %d", flag);
851 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
853 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,
854 comm, PERSISTENT | RECV);
856 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
857 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
858 static int nsleeps = 1;
859 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
860 // behave like a receive, but don't do it
863 print_request("New iprobe", request);
864 // We have to test both mailboxes as we don't know if we will receive one one or another
865 if (sg_cfg_get_int("smpi/async_small_thres")>0){
866 mailbox = smpi_process_mailbox_small();
867 XBT_DEBUG("trying to probe the perm recv mailbox");
868 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
870 if (request->action==NULL){
871 mailbox = smpi_process_mailbox();
872 XBT_DEBUG("trying to probe the other mailbox");
873 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
877 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
879 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
880 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
881 status->MPI_TAG = req->tag;
882 status->MPI_ERROR = MPI_SUCCESS;
883 status->count = req->real_size;
885 nsleeps=1;//reset the number of sleeps we will do next time
891 smpi_mpi_request_free(&request);
896 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
898 print_request("Waiting", *request);
899 if ((*request)->flags & PREPARED) {
900 smpi_empty_status(status);
904 if ((*request)->action != NULL) { // this is not a detached send
905 simcall_comm_wait((*request)->action, -1.0);
907 if((MC_is_active() || MC_record_replay_is_active()) && (*request)->action)
908 (*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
911 finish_wait(request, status);
912 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
913 *request = MPI_REQUEST_NULL;
914 // FIXME for a detached send, finish_wait is not called:
917 int smpi_mpi_waitany(int count, MPI_Request requests[],
924 index = MPI_UNDEFINED;
926 // Wait for a request to complete
927 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
928 map = xbt_new(int, count);
930 XBT_DEBUG("Wait for one of %d", count);
931 for(i = 0; i < count; i++) {
932 if (requests[i] != MPI_REQUEST_NULL
933 && !(requests[i]->flags & PREPARED)
934 && !(requests[i]->flags & FINISHED)) {
935 if (requests[i]->action != NULL) {
936 XBT_DEBUG("Waiting any %p ", requests[i]);
937 xbt_dynar_push(comms, &requests[i]->action);
941 //This is a finished detached request, let's return this one
942 size=0;//so we free the dynar but don't do the waitany call
944 finish_wait(&requests[i], status);//cleanup if refcount = 0
945 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
946 requests[i]=MPI_REQUEST_NULL;//set to null
952 i = simcall_comm_waitany(comms);
954 // not MPI_UNDEFINED, as this is a simix return code
957 finish_wait(&requests[index], status);
958 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
959 requests[index] = MPI_REQUEST_NULL;
963 xbt_dynar_free(&comms);
966 if (index==MPI_UNDEFINED)
967 smpi_empty_status(status);
972 int smpi_mpi_waitall(int count, MPI_Request requests[],
977 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
978 int retvalue = MPI_SUCCESS;
979 //tag invalid requests in the set
980 if (status != MPI_STATUSES_IGNORE) {
981 for (c = 0; c < count; c++) {
982 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
983 (requests[c]->flags & PREPARED)) {
984 smpi_empty_status(&status[c]);
985 } else if (requests[c]->src == MPI_PROC_NULL) {
986 smpi_empty_status(&status[c]);
987 status[c].MPI_SOURCE = MPI_PROC_NULL;
991 for(c = 0; c < count; c++) {
993 if (MC_is_active() || MC_record_replay_is_active()) {
994 smpi_mpi_wait(&requests[c], pstat);
997 index = smpi_mpi_waitany(count, requests, pstat);
998 if (index == MPI_UNDEFINED)
1000 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1001 requests[index]=MPI_REQUEST_NULL;
1003 if (status != MPI_STATUSES_IGNORE) {
1004 status[index] = *pstat;
1005 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1006 retvalue = MPI_ERR_IN_STATUS;
1013 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
1014 MPI_Status status[])
1016 int i, count, index;
1018 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1021 for(i = 0; i < incount; i++)
1023 index=smpi_mpi_waitany(incount, requests, pstat);
1024 if(index!=MPI_UNDEFINED){
1025 indices[count] = index;
1027 if(status != MPI_STATUSES_IGNORE) {
1028 status[index] = *pstat;
1030 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1031 requests[index]=MPI_REQUEST_NULL;
1033 return MPI_UNDEFINED;
1039 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
1040 MPI_Status status[])
1042 int i, count, count_dead;
1044 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1048 for(i = 0; i < incount; i++) {
1049 if((requests[i] != MPI_REQUEST_NULL)) {
1050 if(smpi_mpi_test(&requests[i], pstat)) {
1053 if(status != MPI_STATUSES_IGNORE) {
1056 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1057 requests[i]=MPI_REQUEST_NULL;
1063 if(count_dead==incount)return MPI_UNDEFINED;
1067 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1070 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1071 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1074 void smpi_mpi_barrier(MPI_Comm comm)
1076 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1077 nary_tree_barrier(comm, 4);
1080 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1081 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1082 int root, MPI_Comm comm)
1084 int system_tag = COLL_TAG_GATHER;
1085 int rank, size, src, index;
1086 MPI_Aint lb = 0, recvext = 0;
1087 MPI_Request *requests;
1089 rank = smpi_comm_rank(comm);
1090 size = smpi_comm_size(comm);
1092 // Send buffer to root
1093 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1095 // FIXME: check for errors
1096 smpi_datatype_extent(recvtype, &lb, &recvext);
1097 // Local copy from root
1098 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1099 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1100 // Receive buffers from senders
1101 requests = xbt_new(MPI_Request, size - 1);
1103 for(src = 0; src < size; src++) {
1105 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1106 recvcount, recvtype,
1107 src, system_tag, comm);
1111 // Wait for completion of irecv's.
1112 smpi_mpi_startall(size - 1, requests);
1113 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1114 for(src = 0; src < size-1; src++) {
1115 smpi_mpi_request_free(&requests[src]);
1122 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1123 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1127 int rank = smpi_process_index();
1130 /* arbitrarily choose root as rank 0 */
1131 size = smpi_comm_size(comm);
1133 displs = xbt_new(int, size);
1134 for (i = 0; i < size; i++) {
1136 count += recvcounts[i];
1138 tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1140 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1141 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1142 recvcounts[rank], datatype, 0, comm);
1144 smpi_free_tmp_buffer(tmpbuf);
1147 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1148 void *recvbuf, int *recvcounts, int *displs,
1149 MPI_Datatype recvtype, int root, MPI_Comm comm)
1151 int system_tag = COLL_TAG_GATHERV;
1152 int rank, size, src, index;
1153 MPI_Aint lb = 0, recvext = 0;
1154 MPI_Request *requests;
1156 rank = smpi_comm_rank(comm);
1157 size = smpi_comm_size(comm);
1159 // Send buffer to root
1160 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1162 // FIXME: check for errors
1163 smpi_datatype_extent(recvtype, &lb, &recvext);
1164 // Local copy from root
1165 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1166 (char *)recvbuf + displs[root] * recvext,
1167 recvcounts[root], recvtype);
1168 // Receive buffers from senders
1169 requests = xbt_new(MPI_Request, size - 1);
1171 for(src = 0; src < size; src++) {
1174 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1175 recvcounts[src], recvtype, src, system_tag, comm);
1179 // Wait for completion of irecv's.
1180 smpi_mpi_startall(size - 1, requests);
1181 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1182 for(src = 0; src < size-1; src++) {
1183 smpi_mpi_request_free(&requests[src]);
1189 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1190 MPI_Datatype sendtype, void *recvbuf,
1191 int recvcount, MPI_Datatype recvtype,
1194 int system_tag = COLL_TAG_ALLGATHER;
1195 int rank, size, other, index;
1196 MPI_Aint lb = 0, recvext = 0;
1197 MPI_Request *requests;
1199 rank = smpi_comm_rank(comm);
1200 size = smpi_comm_size(comm);
1201 // FIXME: check for errors
1202 smpi_datatype_extent(recvtype, &lb, &recvext);
1203 // Local copy from self
1204 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1205 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1207 // Send/Recv buffers to/from others;
1208 requests = xbt_new(MPI_Request, 2 * (size - 1));
1210 for(other = 0; other < size; other++) {
1213 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1216 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1217 recvcount, recvtype, other,
1222 // Wait for completion of all comms.
1223 smpi_mpi_startall(2 * (size - 1), requests);
1224 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1225 for(other = 0; other < 2*(size-1); other++) {
1226 smpi_mpi_request_free(&requests[other]);
1231 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1232 MPI_Datatype sendtype, void *recvbuf,
1233 int *recvcounts, int *displs,
1234 MPI_Datatype recvtype, MPI_Comm comm)
1236 int system_tag = COLL_TAG_ALLGATHERV;
1237 int rank, size, other, index;
1238 MPI_Aint lb = 0, recvext = 0;
1239 MPI_Request *requests;
1241 rank = smpi_comm_rank(comm);
1242 size = smpi_comm_size(comm);
1243 // FIXME: check for errors
1244 smpi_datatype_extent(recvtype, &lb, &recvext);
1245 // Local copy from self
1246 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1247 (char *)recvbuf + displs[rank] * recvext,
1248 recvcounts[rank], recvtype);
1249 // Send buffers to others;
1250 requests = xbt_new(MPI_Request, 2 * (size - 1));
1252 for(other = 0; other < size; other++) {
1255 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1259 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1260 recvtype, other, system_tag, comm);
1264 // Wait for completion of all comms.
1265 smpi_mpi_startall(2 * (size - 1), requests);
1266 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1267 for(other = 0; other < 2*(size-1); other++) {
1268 smpi_mpi_request_free(&requests[other]);
1273 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1274 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1275 int root, MPI_Comm comm)
1277 int system_tag = COLL_TAG_SCATTER;
1278 int rank, size, dst, index;
1279 MPI_Aint lb = 0, sendext = 0;
1280 MPI_Request *requests;
1282 rank = smpi_comm_rank(comm);
1283 size = smpi_comm_size(comm);
1285 // Recv buffer from root
1286 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1289 // FIXME: check for errors
1290 smpi_datatype_extent(sendtype, &lb, &sendext);
1291 // Local copy from root
1292 if(recvbuf!=MPI_IN_PLACE){
1293 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1294 sendcount, sendtype, recvbuf, recvcount, recvtype);
1296 // Send buffers to receivers
1297 requests = xbt_new(MPI_Request, size - 1);
1299 for(dst = 0; dst < size; dst++) {
1301 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1302 sendcount, sendtype, dst,
1307 // Wait for completion of isend's.
1308 smpi_mpi_startall(size - 1, requests);
1309 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1310 for(dst = 0; dst < size-1; dst++) {
1311 smpi_mpi_request_free(&requests[dst]);
1317 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1318 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1319 MPI_Datatype recvtype, int root, MPI_Comm comm)
1321 int system_tag = COLL_TAG_SCATTERV;
1322 int rank, size, dst, index;
1323 MPI_Aint lb = 0, sendext = 0;
1324 MPI_Request *requests;
1326 rank = smpi_comm_rank(comm);
1327 size = smpi_comm_size(comm);
1329 // Recv buffer from root
1330 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1333 // FIXME: check for errors
1334 smpi_datatype_extent(sendtype, &lb, &sendext);
1335 // Local copy from root
1336 if(recvbuf!=MPI_IN_PLACE){
1337 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1338 sendtype, recvbuf, recvcount, recvtype);
1340 // Send buffers to receivers
1341 requests = xbt_new(MPI_Request, size - 1);
1343 for(dst = 0; dst < size; dst++) {
1346 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1347 sendtype, dst, system_tag, comm);
1351 // Wait for completion of isend's.
1352 smpi_mpi_startall(size - 1, requests);
1353 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1354 for(dst = 0; dst < size-1; dst++) {
1355 smpi_mpi_request_free(&requests[dst]);
1361 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1362 MPI_Datatype datatype, MPI_Op op, int root,
1365 int system_tag = COLL_TAG_REDUCE;
1366 int rank, size, src, index;
1367 MPI_Aint lb = 0, dataext = 0;
1368 MPI_Request *requests;
1372 char* sendtmpbuf = (char*) sendbuf;
1373 if( sendbuf == MPI_IN_PLACE ) {
1374 sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1375 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1378 rank = smpi_comm_rank(comm);
1379 size = smpi_comm_size(comm);
1380 //non commutative case, use a working algo from openmpi
1381 if(!smpi_op_is_commute(op)){
1382 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1383 datatype, op, root, comm);
1388 // Send buffer to root
1389 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1391 // FIXME: check for errors
1392 smpi_datatype_extent(datatype, &lb, &dataext);
1393 // Local copy from root
1394 if (sendtmpbuf && recvbuf)
1395 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1396 // Receive buffers from senders
1397 //TODO: make a MPI_barrier here ?
1398 requests = xbt_new(MPI_Request, size - 1);
1399 tmpbufs = xbt_new(void *, size - 1);
1401 for(src = 0; src < size; src++) {
1403 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1405 if (!smpi_process_get_replaying())
1406 tmpbufs[index] = xbt_malloc(count * dataext);
1408 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1410 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1415 // Wait for completion of irecv's.
1416 smpi_mpi_startall(size - 1, requests);
1417 for(src = 0; src < size - 1; src++) {
1418 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1419 XBT_DEBUG("finished waiting any request with index %d", index);
1420 if(index == MPI_UNDEFINED) {
1423 smpi_mpi_request_free(&requests[index]);
1425 if(op) /* op can be MPI_OP_NULL that does nothing */
1426 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1428 for(index = 0; index < size - 1; index++) {
1429 smpi_free_tmp_buffer(tmpbufs[index]);
1434 if( sendbuf == MPI_IN_PLACE ) {
1435 smpi_free_tmp_buffer(sendtmpbuf);
1440 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1441 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1443 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1444 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1447 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1448 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1450 int system_tag = -888;
1451 int rank, size, other, index;
1452 MPI_Aint lb = 0, dataext = 0;
1453 MPI_Request *requests;
1456 rank = smpi_comm_rank(comm);
1457 size = smpi_comm_size(comm);
1459 // FIXME: check for errors
1460 smpi_datatype_extent(datatype, &lb, &dataext);
1462 // Local copy from self
1463 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1465 // Send/Recv buffers to/from others;
1466 requests = xbt_new(MPI_Request, size - 1);
1467 tmpbufs = xbt_new(void *, rank);
1469 for(other = 0; other < rank; other++) {
1470 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1472 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1474 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1478 for(other = rank + 1; other < size; other++) {
1480 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1483 // Wait for completion of all comms.
1484 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 // #Request is below rank: it's a irecv
1494 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1498 //non commutative case, wait in order
1499 for(other = 0; other < size - 1; other++) {
1500 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1502 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1506 for(index = 0; index < rank; index++) {
1507 smpi_free_tmp_buffer(tmpbufs[index]);
1509 for(index = 0; index < size-1; index++) {
1510 smpi_mpi_request_free(&requests[index]);
1516 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1517 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1519 int system_tag = -888;
1520 int rank, size, other, index;
1521 MPI_Aint lb = 0, dataext = 0;
1522 MPI_Request *requests;
1524 int recvbuf_is_empty=1;
1525 rank = smpi_comm_rank(comm);
1526 size = smpi_comm_size(comm);
1528 // FIXME: check for errors
1529 smpi_datatype_extent(datatype, &lb, &dataext);
1531 // Send/Recv buffers to/from others;
1532 requests = xbt_new(MPI_Request, size - 1);
1533 tmpbufs = xbt_new(void *, rank);
1535 for(other = 0; other < rank; other++) {
1536 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1538 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1540 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1544 for(other = rank + 1; other < size; other++) {
1546 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1549 // Wait for completion of all comms.
1550 smpi_mpi_startall(size - 1, requests);
1551 if(smpi_op_is_commute(op)){
1552 for(other = 0; other < size - 1; other++) {
1553 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1554 if(index == MPI_UNDEFINED) {
1558 if(recvbuf_is_empty){
1559 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1562 // #Request is below rank: it's a irecv
1563 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1567 //non commutative case, wait in order
1568 for(other = 0; other < size - 1; other++) {
1569 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1571 if(recvbuf_is_empty){
1572 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1574 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1578 for(index = 0; index < rank; index++) {
1579 smpi_free_tmp_buffer(tmpbufs[index]);
1581 for(index = 0; index < size-1; index++) {
1582 smpi_mpi_request_free(&requests[index]);