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_action_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_action_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 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
349 //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
350 mailbox = smpi_process_mailbox_small();
351 XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox);
352 smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
355 mailbox = smpi_process_mailbox();
356 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
357 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
359 XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox);
360 mailbox = smpi_process_mailbox_small();
363 XBT_DEBUG("yes there was something for us in the large mailbox");
366 mailbox = smpi_process_mailbox_small();
367 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
368 smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
371 XBT_DEBUG("No, nothing in the permanent receive mailbox");
372 mailbox = smpi_process_mailbox();
374 XBT_DEBUG("yes there was something for us in the small mailbox");
378 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
379 request->real_size=request->size;
380 smpi_datatype_use(request->old_type);
381 smpi_comm_use(request->comm);
382 request->action = simcall_comm_irecv(mailbox, request->buf,
383 &request->real_size, &match_recv,
384 &smpi_comm_copy_buffer_callback,
386 XBT_DEBUG("recv simcall posted");
387 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
388 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
390 simcall_process_sleep(sleeptime);
391 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
397 int receiver = request->dst;
400 int rank = request->src;
401 if (TRACE_smpi_view_internals()) {
402 TRACE_smpi_send(rank, rank, receiver,request->size);
405 print_request("New send", request);
406 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
407 mailbox = smpi_process_remote_mailbox(receiver);
408 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
409 smx_action_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
411 if (! (request->flags & SSEND)){
412 mailbox = smpi_process_remote_mailbox_small(receiver);
413 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
415 mailbox = smpi_process_remote_mailbox_small(receiver);
416 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
417 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
419 XBT_DEBUG("No, we are first, send to large mailbox");
420 mailbox = smpi_process_remote_mailbox(receiver);
424 XBT_DEBUG("Yes there was something for us in the large mailbox");
427 mailbox = smpi_process_remote_mailbox(receiver);
428 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
431 void* buf = request->buf;
432 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
434 request->detached = 1;
435 XBT_DEBUG("Send request %p is detached", request);
437 if(request->old_type->has_subtype == 0){
438 oldbuf = request->buf;
439 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
440 if((smpi_privatize_global_variables)
441 && ((char*)request->buf >= start_data_exe)
442 && ((char*)request->buf < start_data_exe + size_data_exe )){
443 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
444 smpi_switch_data_segment(request->src);
446 buf = xbt_malloc(request->size);
447 memcpy(buf,oldbuf,request->size);
448 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
453 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
454 request->real_size=request->size;
455 smpi_datatype_use(request->old_type);
456 smpi_comm_use(request->comm);
458 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
459 buf, request->real_size,
461 &xbt_free_f, // how to free the userdata if a detached send fails
462 &smpi_comm_copy_buffer_callback,
464 // detach if msg size < eager/rdv switch limit
466 XBT_DEBUG("send simcall posted");
467 //if we are giving back the control to the user without waiting for completion, we have to inject timings
468 double sleeptime = 0.0;
469 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
470 //isend and send timings may be different
471 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
474 if(sleeptime != 0.0){
475 simcall_process_sleep(sleeptime);
476 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
481 /* FIXME: detached sends are not traceable (request->action == NULL) */
483 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
491 void smpi_mpi_startall(int count, MPI_Request * requests)
494 if(requests==NULL) return;
496 for(i = 0; i < count; i++) {
497 smpi_mpi_start(requests[i]);
501 void smpi_mpi_request_free(MPI_Request * request)
503 if((*request) != MPI_REQUEST_NULL){
504 (*request)->refcount--;
505 if((*request)->refcount<0) xbt_die("wrong refcount");
507 if((*request)->refcount==0){
508 print_request("Destroying", (*request));
510 *request = MPI_REQUEST_NULL;
512 print_request("Decrementing", (*request));
515 xbt_die("freeing an already free request");
520 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
521 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
523 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
525 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
526 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
528 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
529 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
535 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
536 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
538 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
540 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
541 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
543 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
544 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
551 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
552 int dst, int tag, MPI_Comm comm)
554 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
555 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
556 comm, PERSISTENT | ISEND | SEND | PREPARED);
560 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
561 int dst, int tag, MPI_Comm comm)
563 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
564 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
565 comm, NON_PERSISTENT | ISEND | SEND);
566 smpi_mpi_start(request);
570 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
571 int dst, int tag, MPI_Comm comm)
573 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
574 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
575 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
576 smpi_mpi_start(request);
580 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
581 int src, int tag, MPI_Comm comm)
583 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
584 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,
585 comm, PERSISTENT | RECV | PREPARED);
589 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
590 int src, int tag, MPI_Comm comm)
592 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
593 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,
594 comm, NON_PERSISTENT | RECV);
595 smpi_mpi_start(request);
599 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
600 int tag, MPI_Comm comm, MPI_Status * status)
602 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
603 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
604 smpi_mpi_wait(&request, status);
610 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
611 int tag, MPI_Comm comm)
613 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
614 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
615 comm, NON_PERSISTENT | SEND);
617 smpi_mpi_start(request);
618 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
622 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
623 int dst, int tag, MPI_Comm comm)
625 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
626 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
627 comm, NON_PERSISTENT | SSEND | SEND);
629 smpi_mpi_start(request);
630 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
634 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
635 int dst, int sendtag, void *recvbuf, int recvcount,
636 MPI_Datatype recvtype, int src, int recvtag,
637 MPI_Comm comm, MPI_Status * status)
639 MPI_Request requests[2];
641 int myid=smpi_process_index();
642 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
643 smpi_datatype_copy(sendbuf, sendcount, sendtype,
644 recvbuf, recvcount, recvtype);
648 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
650 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
651 smpi_mpi_startall(2, requests);
652 smpi_mpi_waitall(2, requests, stats);
653 smpi_mpi_request_free(&requests[0]);
654 smpi_mpi_request_free(&requests[1]);
655 if(status != MPI_STATUS_IGNORE) {
656 // Copy receive status
661 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
663 return status->count / smpi_datatype_size(datatype);
666 static void finish_wait(MPI_Request * request, MPI_Status * status)
668 MPI_Request req = *request;
669 smpi_empty_status(status);
671 if(!(req->detached && req->flags & SEND)
672 && !(req->flags & PREPARED)){
673 if(status != MPI_STATUS_IGNORE) {
674 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
675 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
676 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
677 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
678 // this handles the case were size in receive differs from size in send
679 // FIXME: really this should just contain the count of receive-type blocks,
681 status->count = req->real_size;
684 print_request("Finishing", req);
685 MPI_Datatype datatype = req->old_type;
687 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
688 if (!_xbt_replay_is_active()){
689 if( smpi_privatize_global_variables
690 && ((char*)req->old_buf >= start_data_exe)
691 && ((char*)req->old_buf < start_data_exe + size_data_exe )
693 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
694 smpi_switch_data_segment(smpi_process_index());
698 if(datatype->has_subtype == 1){
699 // This part handles the problem of non-contignous memory
700 // the unserialization at the reception
701 s_smpi_subtype_t *subtype = datatype->substruct;
702 if(req->flags & RECV)
703 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
704 if(req->detached == 0) free(req->buf);
705 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
706 int n =req->real_size/smpi_datatype_size(datatype);
707 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
710 smpi_comm_unuse(req->comm);
711 smpi_datatype_unuse(datatype);
716 if (TRACE_smpi_view_internals()) {
717 if(req->flags & RECV){
718 int rank = smpi_process_index();
719 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
720 TRACE_smpi_recv(rank, src_traced, rank);
725 if(req->detached_sender!=NULL){
726 smpi_mpi_request_free(&(req->detached_sender));
728 if(req->flags & PERSISTENT)
730 req->flags |= FINISHED;
732 smpi_mpi_request_free(request);
736 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
739 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
742 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
743 static int nsleeps = 1;
744 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
746 smpi_empty_status(status);
748 if (!((*request)->flags & PREPARED)) {
749 if ((*request)->action != NULL)
750 flag = simcall_comm_test((*request)->action);
752 finish_wait(request, status);
753 nsleeps=1;//reset the number of sleeps we will do next time
754 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
755 *request = MPI_REQUEST_NULL;
763 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
770 *index = MPI_UNDEFINED;
772 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
773 map = xbt_new(int, count);
775 for(i = 0; i < count; i++) {
776 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
777 !(requests[i]->flags & PREPARED)) {
778 xbt_dynar_push(comms, &requests[i]->action);
784 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
785 static int nsleeps = 1;
786 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
788 i = simcall_comm_testany(comms);
789 // not MPI_UNDEFINED, as this is a simix return code
792 finish_wait(&requests[*index], status);
793 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
794 requests[*index] = MPI_REQUEST_NULL;
801 //all requests are null or inactive, return true
803 smpi_empty_status(status);
806 xbt_dynar_free(&comms);
812 int smpi_mpi_testall(int count, MPI_Request requests[],
816 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
819 for(i=0; i<count; i++){
820 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
821 if (smpi_mpi_test(&requests[i], pstat)!=1){
824 requests[i]=MPI_REQUEST_NULL;
827 smpi_empty_status(pstat);
829 if(status != MPI_STATUSES_IGNORE) {
836 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
838 //FIXME find another wait to avoid busy waiting ?
839 // the issue here is that we have to wait on a nonexistent comm
841 smpi_mpi_iprobe(source, tag, comm, &flag, status);
842 XBT_DEBUG("Busy Waiting on probing : %d", flag);
846 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
848 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
849 comm, PERSISTENT | RECV);
851 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
852 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
853 static int nsleeps = 1;
854 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
855 // behave like a receive, but don't do it
858 print_request("New iprobe", request);
859 // We have to test both mailboxes as we don't know if we will receive one one or another
860 if (sg_cfg_get_int("smpi/async_small_thres")>0){
861 mailbox = smpi_process_mailbox_small();
862 XBT_DEBUG("trying to probe the perm recv mailbox");
863 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
865 if (request->action==NULL){
866 mailbox = smpi_process_mailbox();
867 XBT_DEBUG("trying to probe the other mailbox");
868 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
872 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
874 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
875 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
876 status->MPI_TAG = req->tag;
877 status->MPI_ERROR = MPI_SUCCESS;
878 status->count = req->real_size;
880 nsleeps=1;//reset the number of sleeps we will do next time
886 smpi_mpi_request_free(&request);
891 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
893 print_request("Waiting", *request);
894 if ((*request)->flags & PREPARED) {
895 smpi_empty_status(status);
899 if ((*request)->action != NULL) { // this is not a detached send
900 simcall_comm_wait((*request)->action, -1.0);
902 if(MC_is_active() && (*request)->action)
903 (*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
907 finish_wait(request, status);
908 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
909 *request = MPI_REQUEST_NULL;
910 // FIXME for a detached send, finish_wait is not called:
913 int smpi_mpi_waitany(int count, MPI_Request requests[],
920 index = MPI_UNDEFINED;
922 // Wait for a request to complete
923 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
924 map = xbt_new(int, count);
926 XBT_DEBUG("Wait for one of %d", count);
927 for(i = 0; i < count; i++) {
928 if (requests[i] != MPI_REQUEST_NULL
929 && !(requests[i]->flags & PREPARED)
930 && !(requests[i]->flags & FINISHED)) {
931 if (requests[i]->action != NULL) {
932 XBT_DEBUG("Waiting any %p ", requests[i]);
933 xbt_dynar_push(comms, &requests[i]->action);
937 //This is a finished detached request, let's return this one
938 size=0;//so we free the dynar but don't do the waitany call
940 finish_wait(&requests[i], status);//cleanup if refcount = 0
941 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
942 requests[i]=MPI_REQUEST_NULL;//set to null
948 i = simcall_comm_waitany(comms);
950 // not MPI_UNDEFINED, as this is a simix return code
953 finish_wait(&requests[index], status);
954 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
955 requests[index] = MPI_REQUEST_NULL;
959 xbt_dynar_free(&comms);
962 if (index==MPI_UNDEFINED)
963 smpi_empty_status(status);
968 int smpi_mpi_waitall(int count, MPI_Request requests[],
973 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
974 int retvalue = MPI_SUCCESS;
975 //tag invalid requests in the set
976 if (status != MPI_STATUSES_IGNORE) {
977 for (c = 0; c < count; c++) {
978 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
979 (requests[c]->flags & PREPARED)) {
980 smpi_empty_status(&status[c]);
981 } else if (requests[c]->src == MPI_PROC_NULL) {
982 smpi_empty_status(&status[c]);
983 status[c].MPI_SOURCE = MPI_PROC_NULL;
987 for(c = 0; c < count; c++) {
989 if (MC_is_active()) {
990 smpi_mpi_wait(&requests[c], pstat);
993 index = smpi_mpi_waitany(count, requests, pstat);
994 if (index == MPI_UNDEFINED)
996 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
997 requests[index]=MPI_REQUEST_NULL;
999 if (status != MPI_STATUSES_IGNORE) {
1000 status[index] = *pstat;
1001 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1002 retvalue = MPI_ERR_IN_STATUS;
1009 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
1010 MPI_Status status[])
1012 int i, count, index;
1014 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1017 for(i = 0; i < incount; i++)
1019 index=smpi_mpi_waitany(incount, requests, pstat);
1020 if(index!=MPI_UNDEFINED){
1021 indices[count] = index;
1023 if(status != MPI_STATUSES_IGNORE) {
1024 status[index] = *pstat;
1026 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1027 requests[index]=MPI_REQUEST_NULL;
1029 return MPI_UNDEFINED;
1035 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
1036 MPI_Status status[])
1038 int i, count, count_dead;
1040 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1044 for(i = 0; i < incount; i++) {
1045 if((requests[i] != MPI_REQUEST_NULL)) {
1046 if(smpi_mpi_test(&requests[i], pstat)) {
1049 if(status != MPI_STATUSES_IGNORE) {
1052 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1053 requests[i]=MPI_REQUEST_NULL;
1059 if(count_dead==incount)return MPI_UNDEFINED;
1063 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1066 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1067 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1070 void smpi_mpi_barrier(MPI_Comm comm)
1072 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1073 nary_tree_barrier(comm, 4);
1076 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1077 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1078 int root, MPI_Comm comm)
1080 int system_tag = COLL_TAG_GATHER;
1081 int rank, size, src, index;
1082 MPI_Aint lb = 0, recvext = 0;
1083 MPI_Request *requests;
1085 rank = smpi_comm_rank(comm);
1086 size = smpi_comm_size(comm);
1088 // Send buffer to root
1089 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1091 // FIXME: check for errors
1092 smpi_datatype_extent(recvtype, &lb, &recvext);
1093 // Local copy from root
1094 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1095 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1096 // Receive buffers from senders
1097 requests = xbt_new(MPI_Request, size - 1);
1099 for(src = 0; src < size; src++) {
1101 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1102 recvcount, recvtype,
1103 src, system_tag, comm);
1107 // Wait for completion of irecv's.
1108 smpi_mpi_startall(size - 1, requests);
1109 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1110 for(src = 0; src < size-1; src++) {
1111 smpi_mpi_request_free(&requests[src]);
1118 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1119 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1123 int rank = smpi_process_index();
1126 /* arbitrarily choose root as rank 0 */
1127 size = smpi_comm_size(comm);
1129 displs = xbt_new(int, size);
1130 for (i = 0; i < size; i++) {
1132 count += recvcounts[i];
1134 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1135 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1136 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1137 recvcounts[rank], datatype, 0, comm);
1142 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1143 void *recvbuf, int *recvcounts, int *displs,
1144 MPI_Datatype recvtype, int root, MPI_Comm comm)
1146 int system_tag = COLL_TAG_GATHERV;
1147 int rank, size, src, index;
1148 MPI_Aint lb = 0, recvext = 0;
1149 MPI_Request *requests;
1151 rank = smpi_comm_rank(comm);
1152 size = smpi_comm_size(comm);
1154 // Send buffer to root
1155 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1157 // FIXME: check for errors
1158 smpi_datatype_extent(recvtype, &lb, &recvext);
1159 // Local copy from root
1160 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1161 (char *)recvbuf + displs[root] * recvext,
1162 recvcounts[root], recvtype);
1163 // Receive buffers from senders
1164 requests = xbt_new(MPI_Request, size - 1);
1166 for(src = 0; src < size; src++) {
1169 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1170 recvcounts[src], recvtype, src, system_tag, comm);
1174 // Wait for completion of irecv's.
1175 smpi_mpi_startall(size - 1, requests);
1176 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1177 for(src = 0; src < size-1; src++) {
1178 smpi_mpi_request_free(&requests[src]);
1184 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1185 MPI_Datatype sendtype, void *recvbuf,
1186 int recvcount, MPI_Datatype recvtype,
1189 int system_tag = COLL_TAG_ALLGATHER;
1190 int rank, size, other, index;
1191 MPI_Aint lb = 0, recvext = 0;
1192 MPI_Request *requests;
1194 rank = smpi_comm_rank(comm);
1195 size = smpi_comm_size(comm);
1196 // FIXME: check for errors
1197 smpi_datatype_extent(recvtype, &lb, &recvext);
1198 // Local copy from self
1199 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1200 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1202 // Send/Recv buffers to/from others;
1203 requests = xbt_new(MPI_Request, 2 * (size - 1));
1205 for(other = 0; other < size; other++) {
1208 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1211 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1212 recvcount, recvtype, other,
1217 // Wait for completion of all comms.
1218 smpi_mpi_startall(2 * (size - 1), requests);
1219 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1220 for(other = 0; other < 2*(size-1); other++) {
1221 smpi_mpi_request_free(&requests[other]);
1226 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1227 MPI_Datatype sendtype, void *recvbuf,
1228 int *recvcounts, int *displs,
1229 MPI_Datatype recvtype, MPI_Comm comm)
1231 int system_tag = COLL_TAG_ALLGATHERV;
1232 int rank, size, other, index;
1233 MPI_Aint lb = 0, recvext = 0;
1234 MPI_Request *requests;
1236 rank = smpi_comm_rank(comm);
1237 size = smpi_comm_size(comm);
1238 // FIXME: check for errors
1239 smpi_datatype_extent(recvtype, &lb, &recvext);
1240 // Local copy from self
1241 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1242 (char *)recvbuf + displs[rank] * recvext,
1243 recvcounts[rank], recvtype);
1244 // Send buffers to others;
1245 requests = xbt_new(MPI_Request, 2 * (size - 1));
1247 for(other = 0; other < size; other++) {
1250 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1254 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1255 recvtype, other, system_tag, comm);
1259 // Wait for completion of all comms.
1260 smpi_mpi_startall(2 * (size - 1), requests);
1261 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1262 for(other = 0; other < 2*(size-1); other++) {
1263 smpi_mpi_request_free(&requests[other]);
1268 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1269 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1270 int root, MPI_Comm comm)
1272 int system_tag = COLL_TAG_SCATTER;
1273 int rank, size, dst, index;
1274 MPI_Aint lb = 0, sendext = 0;
1275 MPI_Request *requests;
1277 rank = smpi_comm_rank(comm);
1278 size = smpi_comm_size(comm);
1280 // Recv buffer from root
1281 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1284 // FIXME: check for errors
1285 smpi_datatype_extent(sendtype, &lb, &sendext);
1286 // Local copy from root
1287 if(recvbuf!=MPI_IN_PLACE){
1288 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1289 sendcount, sendtype, recvbuf, recvcount, recvtype);
1291 // Send buffers to receivers
1292 requests = xbt_new(MPI_Request, size - 1);
1294 for(dst = 0; dst < size; dst++) {
1296 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1297 sendcount, sendtype, dst,
1302 // Wait for completion of isend's.
1303 smpi_mpi_startall(size - 1, requests);
1304 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1305 for(dst = 0; dst < size-1; dst++) {
1306 smpi_mpi_request_free(&requests[dst]);
1312 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1313 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1314 MPI_Datatype recvtype, int root, MPI_Comm comm)
1316 int system_tag = COLL_TAG_SCATTERV;
1317 int rank, size, dst, index;
1318 MPI_Aint lb = 0, sendext = 0;
1319 MPI_Request *requests;
1321 rank = smpi_comm_rank(comm);
1322 size = smpi_comm_size(comm);
1324 // Recv buffer from root
1325 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1328 // FIXME: check for errors
1329 smpi_datatype_extent(sendtype, &lb, &sendext);
1330 // Local copy from root
1331 if(recvbuf!=MPI_IN_PLACE){
1332 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1333 sendtype, recvbuf, recvcount, recvtype);
1335 // Send buffers to receivers
1336 requests = xbt_new(MPI_Request, size - 1);
1338 for(dst = 0; dst < size; dst++) {
1341 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1342 sendtype, dst, system_tag, comm);
1346 // Wait for completion of isend's.
1347 smpi_mpi_startall(size - 1, requests);
1348 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1349 for(dst = 0; dst < size-1; dst++) {
1350 smpi_mpi_request_free(&requests[dst]);
1356 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1357 MPI_Datatype datatype, MPI_Op op, int root,
1360 int system_tag = COLL_TAG_REDUCE;
1361 int rank, size, src, index;
1362 MPI_Aint lb = 0, dataext = 0;
1363 MPI_Request *requests;
1367 char* sendtmpbuf = (char*) sendbuf;
1368 if( sendbuf == MPI_IN_PLACE ) {
1369 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1370 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1373 rank = smpi_comm_rank(comm);
1374 size = smpi_comm_size(comm);
1375 //non commutative case, use a working algo from openmpi
1376 if(!smpi_op_is_commute(op)){
1377 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1378 datatype, op, root, comm);
1383 // Send buffer to root
1384 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1386 // FIXME: check for errors
1387 smpi_datatype_extent(datatype, &lb, &dataext);
1388 // Local copy from root
1389 if (sendtmpbuf && recvbuf)
1390 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1391 // Receive buffers from senders
1392 //TODO: make a MPI_barrier here ?
1393 requests = xbt_new(MPI_Request, size - 1);
1394 tmpbufs = xbt_new(void *, size - 1);
1396 for(src = 0; src < size; src++) {
1398 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1400 tmpbufs[index] = xbt_malloc(count * dataext);
1402 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1407 // Wait for completion of irecv's.
1408 smpi_mpi_startall(size - 1, requests);
1409 for(src = 0; src < size - 1; src++) {
1410 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1411 XBT_DEBUG("finished waiting any request with index %d", index);
1412 if(index == MPI_UNDEFINED) {
1415 smpi_mpi_request_free(&requests[index]);
1417 if(op) /* op can be MPI_OP_NULL that does nothing */
1418 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1420 for(index = 0; index < size - 1; index++) {
1421 xbt_free(tmpbufs[index]);
1426 if( sendbuf == MPI_IN_PLACE ) {
1427 xbt_free(sendtmpbuf);
1432 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1433 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1435 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1436 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1439 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1440 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1442 int system_tag = -888;
1443 int rank, size, other, index;
1444 MPI_Aint lb = 0, dataext = 0;
1445 MPI_Request *requests;
1448 rank = smpi_comm_rank(comm);
1449 size = smpi_comm_size(comm);
1451 // FIXME: check for errors
1452 smpi_datatype_extent(datatype, &lb, &dataext);
1454 // Local copy from self
1455 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1457 // Send/Recv buffers to/from others;
1458 requests = xbt_new(MPI_Request, size - 1);
1459 tmpbufs = xbt_new(void *, rank);
1461 for(other = 0; other < rank; other++) {
1462 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1464 tmpbufs[index] = xbt_malloc(count * dataext);
1466 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1470 for(other = rank + 1; other < size; other++) {
1472 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1475 // Wait for completion of all comms.
1476 smpi_mpi_startall(size - 1, requests);
1478 if(smpi_op_is_commute(op)){
1479 for(other = 0; other < size - 1; other++) {
1480 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1481 if(index == MPI_UNDEFINED) {
1485 // #Request is below rank: it's a irecv
1486 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1490 //non commutative case, wait in order
1491 for(other = 0; other < size - 1; other++) {
1492 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1494 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1498 for(index = 0; index < rank; index++) {
1499 xbt_free(tmpbufs[index]);
1501 for(index = 0; index < size-1; index++) {
1502 smpi_mpi_request_free(&requests[index]);
1508 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1509 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1511 int system_tag = -888;
1512 int rank, size, other, index;
1513 MPI_Aint lb = 0, dataext = 0;
1514 MPI_Request *requests;
1516 int recvbuf_is_empty=1;
1517 rank = smpi_comm_rank(comm);
1518 size = smpi_comm_size(comm);
1520 // FIXME: check for errors
1521 smpi_datatype_extent(datatype, &lb, &dataext);
1523 // Send/Recv buffers to/from others;
1524 requests = xbt_new(MPI_Request, size - 1);
1525 tmpbufs = xbt_new(void *, rank);
1527 for(other = 0; other < rank; other++) {
1528 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1530 tmpbufs[index] = xbt_malloc(count * dataext);
1532 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1536 for(other = rank + 1; other < size; other++) {
1538 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1541 // Wait for completion of all comms.
1542 smpi_mpi_startall(size - 1, requests);
1543 if(smpi_op_is_commute(op)){
1544 for(other = 0; other < size - 1; other++) {
1545 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1546 if(index == MPI_UNDEFINED) {
1550 if(recvbuf_is_empty){
1551 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1554 // #Request is below rank: it's a irecv
1555 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1559 //non commutative case, wait in order
1560 for(other = 0; other < size - 1; other++) {
1561 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1563 if(recvbuf_is_empty){
1564 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1566 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1570 for(index = 0; index < rank; index++) {
1571 xbt_free(tmpbufs[index]);
1573 for(index = 0; index < size-1; index++) {
1574 smpi_mpi_request_free(&requests[index]);