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 // Methods used to parse and store the values for timing injections in smpi
75 // These are taken from surf/network.c and generalized to have more factors
76 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
78 static int factor_cmp(const void *pa, const void *pb)
80 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
81 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
85 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
88 unsigned int iter = 0;
92 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
94 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
95 radical_elements = xbt_str_split(smpi_coef_string, ";");
96 xbt_dynar_foreach(radical_elements, iter, value) {
97 memset(&fact, 0, sizeof(s_smpi_factor_t));
98 radical_elements2 = xbt_str_split(value, ":");
99 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
100 xbt_die("Malformed radical for smpi factor!");
101 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
103 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
105 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
110 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
111 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
112 xbt_dynar_free(&radical_elements2);
114 xbt_dynar_free(&radical_elements);
116 xbt_dynar_sort(smpi_factor, &factor_cmp);
117 xbt_dynar_foreach(smpi_factor, iter, fact) {
118 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
123 static double smpi_os(double size)
125 if (!smpi_os_values) {
126 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
127 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
129 unsigned int iter = 0;
130 s_smpi_factor_t fact;
132 xbt_dynar_foreach(smpi_os_values, iter, fact) {
133 if (size <= fact.factor) {
134 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
137 current=fact.values[0]+fact.values[1]*size;
140 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
145 static double smpi_ois(double size)
147 if (!smpi_ois_values) {
148 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
149 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
151 unsigned int iter = 0;
152 s_smpi_factor_t fact;
154 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
155 if (size <= fact.factor) {
156 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
159 current=fact.values[0]+fact.values[1]*size;
162 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
167 static double smpi_or(double size)
169 if (!smpi_or_values) {
170 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
171 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
173 unsigned int iter = 0;
174 s_smpi_factor_t fact;
176 xbt_dynar_foreach(smpi_or_values, iter, fact) {
177 if (size <= fact.factor) {
178 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
181 current=fact.values[0]+fact.values[1]*size;
183 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
190 if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
192 time = SIMIX_get_clock();
193 //to avoid deadlocks if called too many times
194 double sleeptime= sg_cfg_get_double("smpi/wtime");
195 simcall_process_sleep(sleeptime);
198 time = SIMIX_get_clock();
203 static MPI_Request build_request(void *buf, int count,
204 MPI_Datatype datatype, int src, int dst,
205 int tag, MPI_Comm comm, unsigned flags)
207 MPI_Request request = NULL;
209 void *old_buf = NULL;
211 request = xbt_new(s_smpi_mpi_request_t, 1);
213 s_smpi_subtype_t *subtype = datatype->substruct;
215 if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
216 // This part handles the problem of non-contiguous memory
218 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
219 if ((datatype->has_subtype == 1) && (flags & SEND)) {
220 subtype->serialize(old_buf, buf, count, datatype->substruct);
225 // This part handles the problem of non-contiguous memory (for the
226 // unserialisation at the reception)
227 request->old_buf = old_buf;
228 request->old_type = datatype;
230 request->size = smpi_datatype_size(datatype) * count;
234 request->comm = comm;
235 request->action = NULL;
236 request->flags = flags;
237 request->detached = 0;
238 request->detached_sender = NULL;
239 request->real_src = 0;
241 request->truncated = 0;
242 request->real_size = 0;
243 request->real_tag = 0;
244 if(flags & PERSISTENT)
245 request->refcount = 1;
247 request->refcount = 0;
248 request->op = MPI_REPLACE;
253 if (flags & SEND) smpi_datatype_unuse(datatype);
259 void smpi_empty_status(MPI_Status * status)
261 if(status != MPI_STATUS_IGNORE) {
262 status->MPI_SOURCE = MPI_ANY_SOURCE;
263 status->MPI_TAG = MPI_ANY_TAG;
264 status->MPI_ERROR = MPI_SUCCESS;
269 void smpi_action_trace_run(char *path)
273 xbt_dict_cursor_t cursor;
277 action_fp = fopen(path, "r");
278 if (action_fp == NULL)
279 xbt_die("Cannot open %s: %s", path, strerror(errno));
282 if (!xbt_dict_is_empty(action_queues)) {
284 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
287 xbt_dict_foreach(action_queues, cursor, name, todo) {
288 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
294 xbt_dict_free(&action_queues);
295 action_queues = xbt_dict_new_homogeneous(NULL);
298 static void smpi_mpi_request_free_voidp(void* request)
300 MPI_Request req = request;
301 smpi_mpi_request_free(&req);
304 /* MPI Low level calls */
305 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
306 int dst, int tag, MPI_Comm comm)
308 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
309 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
310 comm, PERSISTENT | SEND | PREPARED);
314 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
315 int dst, int tag, MPI_Comm comm)
317 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
318 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
319 comm, PERSISTENT | SSEND | SEND | PREPARED);
323 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
324 int src, int tag, MPI_Comm comm)
326 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
327 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,
328 comm, PERSISTENT | RECV | PREPARED);
332 void smpi_mpi_start(MPI_Request request)
336 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
337 request->flags &= ~PREPARED;
338 request->flags &= ~FINISHED;
341 if (request->flags & RECV) {
342 print_request("New recv", request);
343 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
344 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres"))
345 mailbox = smpi_process_mailbox_small();
347 mailbox = smpi_process_mailbox();
348 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
349 request->real_size=request->size;
350 smpi_datatype_use(request->old_type);
351 smpi_comm_use(request->comm);
352 request->action = simcall_comm_irecv(mailbox, request->buf,
353 &request->real_size, &match_recv,
354 &smpi_comm_copy_buffer_callback,
357 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
358 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
360 simcall_process_sleep(sleeptime);
361 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
367 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
370 int rank = request->src;
371 if (TRACE_smpi_view_internals()) {
372 TRACE_smpi_send(rank, rank, receiver,request->size);
375 /* if(receiver == MPI_UNDEFINED) {*/
376 /* XBT_WARN("Trying to send a message to a wrong rank");*/
379 print_request("New send", request);
380 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
381 mailbox = smpi_process_remote_mailbox_small(receiver);
383 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
384 mailbox = smpi_process_remote_mailbox(receiver);
387 void* buf = request->buf;
388 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
390 request->detached = 1;
391 XBT_DEBUG("Send request %p is detached", request);
393 if(request->old_type->has_subtype == 0){
394 oldbuf = request->buf;
395 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
396 if((smpi_privatize_global_variables)
397 && ((char*)request->buf >= start_data_exe)
398 && ((char*)request->buf < start_data_exe + size_data_exe )){
399 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
400 switch_data_segment(request->src);
402 buf = xbt_malloc(request->size);
403 memcpy(buf,oldbuf,request->size);
404 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
409 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
410 request->real_size=request->size;
411 smpi_datatype_use(request->old_type);
412 smpi_comm_use(request->comm);
414 //if we are giving back the control to the user without waiting for completion, we have to inject timings
415 double sleeptime = 0.0;
416 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
417 //isend and send timings may be different
418 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
421 if(sleeptime != 0.0){
422 simcall_process_sleep(sleeptime);
423 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
426 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
427 buf, request->real_size,
429 &xbt_free_f, // how to free the userdata if a detached send fails
430 &smpi_comm_copy_buffer_callback,
432 // detach if msg size < eager/rdv switch limit
436 /* FIXME: detached sends are not traceable (request->action == NULL) */
438 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
446 void smpi_mpi_startall(int count, MPI_Request * requests)
449 if(requests==NULL) return;
451 for(i = 0; i < count; i++) {
452 smpi_mpi_start(requests[i]);
456 void smpi_mpi_request_free(MPI_Request * request)
458 if((*request) != MPI_REQUEST_NULL){
459 (*request)->refcount--;
460 if((*request)->refcount<0) xbt_die("wrong refcount");
462 if((*request)->refcount==0){
463 print_request("Destroying", (*request));
465 *request = MPI_REQUEST_NULL;
467 print_request("Decrementing", (*request));
470 xbt_die("freeing an already free request");
475 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
476 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
478 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
480 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
481 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
483 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
484 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
490 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
491 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
493 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
495 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
496 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
498 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
499 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
506 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
507 int dst, int tag, MPI_Comm comm)
509 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
510 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
511 comm, PERSISTENT | ISEND | SEND | PREPARED);
515 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
516 int dst, int tag, MPI_Comm comm)
518 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
519 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
520 comm, NON_PERSISTENT | ISEND | SEND);
521 smpi_mpi_start(request);
525 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
526 int dst, int tag, MPI_Comm comm)
528 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
529 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
530 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
531 smpi_mpi_start(request);
535 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
536 int src, int tag, MPI_Comm comm)
538 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
539 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,
540 comm, PERSISTENT | RECV | PREPARED);
544 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
545 int src, int tag, MPI_Comm comm)
547 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
548 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,
549 comm, NON_PERSISTENT | RECV);
550 smpi_mpi_start(request);
554 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
555 int tag, MPI_Comm comm, MPI_Status * status)
557 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
558 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
559 smpi_mpi_wait(&request, status);
565 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
566 int tag, MPI_Comm comm)
568 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
569 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
570 comm, NON_PERSISTENT | SEND);
572 smpi_mpi_start(request);
573 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
577 void smpi_mpi_ssend(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 | SSEND | SEND);
584 smpi_mpi_start(request);
585 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
589 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
590 int dst, int sendtag, void *recvbuf, int recvcount,
591 MPI_Datatype recvtype, int src, int recvtag,
592 MPI_Comm comm, MPI_Status * status)
594 MPI_Request requests[2];
596 int myid=smpi_process_index();
597 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
598 smpi_datatype_copy(sendbuf, sendcount, sendtype,
599 recvbuf, recvcount, recvtype);
603 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
605 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
606 smpi_mpi_startall(2, requests);
607 smpi_mpi_waitall(2, requests, stats);
608 smpi_mpi_request_free(&requests[0]);
609 smpi_mpi_request_free(&requests[1]);
610 if(status != MPI_STATUS_IGNORE) {
611 // Copy receive status
616 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
618 return status->count / smpi_datatype_size(datatype);
621 static void finish_wait(MPI_Request * request, MPI_Status * status)
623 MPI_Request req = *request;
624 smpi_empty_status(status);
626 if(!(req->detached && req->flags & SEND)
627 && !(req->flags & PREPARED)){
628 if(status != MPI_STATUS_IGNORE) {
629 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
630 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
631 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
632 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
633 // this handles the case were size in receive differs from size in send
634 // FIXME: really this should just contain the count of receive-type blocks,
636 status->count = req->real_size;
639 print_request("Finishing", req);
640 MPI_Datatype datatype = req->old_type;
642 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
643 if (!_xbt_replay_is_active()){
644 if( smpi_privatize_global_variables
645 && ((char*)req->old_buf >= start_data_exe)
646 && ((char*)req->old_buf < start_data_exe + size_data_exe )
648 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
649 switch_data_segment(smpi_process_index());
653 if(datatype->has_subtype == 1){
654 // This part handles the problem of non-contignous memory
655 // the unserialization at the reception
656 s_smpi_subtype_t *subtype = datatype->substruct;
657 if(req->flags & RECV)
658 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
659 if(req->detached == 0) free(req->buf);
660 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
661 int n =req->real_size/smpi_datatype_size(datatype);
662 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
665 smpi_comm_unuse(req->comm);
666 smpi_datatype_unuse(datatype);
671 if (TRACE_smpi_view_internals()) {
672 if(req->flags & RECV){
673 int rank = smpi_process_index();
674 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
675 TRACE_smpi_recv(rank, src_traced, rank);
680 if(req->detached_sender!=NULL){
681 smpi_mpi_request_free(&(req->detached_sender));
683 if(req->flags & PERSISTENT)
685 req->flags |= FINISHED;
687 smpi_mpi_request_free(request);
691 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
694 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
697 double sleeptime= sg_cfg_get_double("smpi/test");
698 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
699 static int nsleeps = 1;
700 simcall_process_sleep(nsleeps*sleeptime);
702 smpi_empty_status(status);
704 if (!((*request)->flags & PREPARED)) {
705 if ((*request)->action != NULL)
706 flag = simcall_comm_test((*request)->action);
708 finish_wait(request, status);
709 nsleeps=1;//reset the number of sleeps we will do next time
710 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
711 *request = MPI_REQUEST_NULL;
719 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
726 *index = MPI_UNDEFINED;
728 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
729 map = xbt_new(int, count);
731 for(i = 0; i < count; i++) {
732 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
733 !(requests[i]->flags & PREPARED)) {
734 xbt_dynar_push(comms, &requests[i]->action);
741 double sleeptime= sg_cfg_get_double("smpi/test");
742 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
743 static int nsleeps = 1;
744 simcall_process_sleep(nsleeps*sleeptime);
746 i = simcall_comm_testany(comms);
747 // not MPI_UNDEFINED, as this is a simix return code
750 finish_wait(&requests[*index], status);
751 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
752 requests[*index] = MPI_REQUEST_NULL;
759 //all requests are null or inactive, return true
761 smpi_empty_status(status);
764 xbt_dynar_free(&comms);
770 int smpi_mpi_testall(int count, MPI_Request requests[],
774 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
777 for(i=0; i<count; i++){
778 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
779 if (smpi_mpi_test(&requests[i], pstat)!=1){
782 requests[i]=MPI_REQUEST_NULL;
785 smpi_empty_status(pstat);
787 if(status != MPI_STATUSES_IGNORE) {
794 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
796 //FIXME find another wait to avoid busy waiting ?
797 // the issue here is that we have to wait on a nonexistent comm
799 smpi_mpi_iprobe(source, tag, comm, &flag, status);
800 XBT_DEBUG("Busy Waiting on probing : %d", flag);
804 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
806 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,
807 comm, PERSISTENT | RECV);
809 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
810 double sleeptime= sg_cfg_get_double("smpi/iprobe");
811 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
812 static int nsleeps = 1;
814 simcall_process_sleep(nsleeps*sleeptime);
816 // behave like a receive, but don't do it
819 print_request("New iprobe", request);
820 // We have to test both mailboxes as we don't know if we will receive one one or another
821 if (sg_cfg_get_int("smpi/async_small_thres")>0){
822 mailbox = smpi_process_mailbox_small();
823 XBT_DEBUG("trying to probe the perm recv mailbox");
824 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
826 if (request->action==NULL){
827 mailbox = smpi_process_mailbox();
828 XBT_DEBUG("trying to probe the other mailbox");
829 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
833 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
835 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
836 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
837 status->MPI_TAG = req->tag;
838 status->MPI_ERROR = MPI_SUCCESS;
839 status->count = req->real_size;
841 nsleeps=1;//reset the number of sleeps we will do next time
847 smpi_mpi_request_free(&request);
852 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
854 print_request("Waiting", *request);
855 if ((*request)->flags & PREPARED) {
856 smpi_empty_status(status);
860 if ((*request)->action != NULL) { // this is not a detached send
861 simcall_comm_wait((*request)->action, -1.0);
863 if(MC_is_active() && (*request)->action)
864 (*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
868 finish_wait(request, status);
869 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
870 *request = MPI_REQUEST_NULL;
871 // FIXME for a detached send, finish_wait is not called:
874 int smpi_mpi_waitany(int count, MPI_Request requests[],
881 index = MPI_UNDEFINED;
883 // Wait for a request to complete
884 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
885 map = xbt_new(int, count);
887 XBT_DEBUG("Wait for one of %d", count);
888 for(i = 0; i < count; i++) {
889 if (requests[i] != MPI_REQUEST_NULL
890 && !(requests[i]->flags & PREPARED)
891 && !(requests[i]->flags & FINISHED)) {
892 if (requests[i]->action != NULL) {
893 XBT_DEBUG("Waiting any %p ", requests[i]);
894 xbt_dynar_push(comms, &requests[i]->action);
898 //This is a finished detached request, let's return this one
899 size=0;//so we free the dynar but don't do the waitany call
901 finish_wait(&requests[i], status);//cleanup if refcount = 0
902 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
903 requests[i]=MPI_REQUEST_NULL;//set to null
909 i = simcall_comm_waitany(comms);
911 // not MPI_UNDEFINED, as this is a simix return code
914 finish_wait(&requests[index], status);
915 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
916 requests[index] = MPI_REQUEST_NULL;
920 xbt_dynar_free(&comms);
923 if (index==MPI_UNDEFINED)
924 smpi_empty_status(status);
929 int smpi_mpi_waitall(int count, MPI_Request requests[],
934 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
935 int retvalue = MPI_SUCCESS;
936 //tag invalid requests in the set
937 if (status != MPI_STATUSES_IGNORE) {
938 for (c = 0; c < count; c++) {
939 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
940 (requests[c]->flags & PREPARED)) {
941 smpi_empty_status(&status[c]);
942 } else if (requests[c]->src == MPI_PROC_NULL) {
943 smpi_empty_status(&status[c]);
944 status[c].MPI_SOURCE = MPI_PROC_NULL;
948 for(c = 0; c < count; c++) {
950 if (MC_is_active()) {
951 smpi_mpi_wait(&requests[c], pstat);
954 index = smpi_mpi_waitany(count, requests, pstat);
955 if (index == MPI_UNDEFINED)
957 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
958 requests[index]=MPI_REQUEST_NULL;
960 if (status != MPI_STATUSES_IGNORE) {
961 status[index] = *pstat;
962 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
963 retvalue = MPI_ERR_IN_STATUS;
970 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
975 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
978 for(i = 0; i < incount; i++)
980 index=smpi_mpi_waitany(incount, requests, pstat);
981 if(index!=MPI_UNDEFINED){
982 indices[count] = index;
984 if(status != MPI_STATUSES_IGNORE) {
985 status[index] = *pstat;
987 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
988 requests[index]=MPI_REQUEST_NULL;
990 return MPI_UNDEFINED;
996 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
999 int i, count, count_dead;
1001 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1005 for(i = 0; i < incount; i++) {
1006 if((requests[i] != MPI_REQUEST_NULL)) {
1007 if(smpi_mpi_test(&requests[i], pstat)) {
1010 if(status != MPI_STATUSES_IGNORE) {
1013 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1014 requests[i]=MPI_REQUEST_NULL;
1020 if(count_dead==incount)return MPI_UNDEFINED;
1024 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1027 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1028 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1031 void smpi_mpi_barrier(MPI_Comm comm)
1033 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1034 nary_tree_barrier(comm, 4);
1037 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1038 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1039 int root, MPI_Comm comm)
1041 int system_tag = COLL_TAG_GATHER;
1042 int rank, size, src, index;
1043 MPI_Aint lb = 0, recvext = 0;
1044 MPI_Request *requests;
1046 rank = smpi_comm_rank(comm);
1047 size = smpi_comm_size(comm);
1049 // Send buffer to root
1050 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1052 // FIXME: check for errors
1053 smpi_datatype_extent(recvtype, &lb, &recvext);
1054 // Local copy from root
1055 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1056 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1057 // Receive buffers from senders
1058 requests = xbt_new(MPI_Request, size - 1);
1060 for(src = 0; src < size; src++) {
1062 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1063 recvcount, recvtype,
1064 src, system_tag, comm);
1068 // Wait for completion of irecv's.
1069 smpi_mpi_startall(size - 1, requests);
1070 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1071 for(src = 0; src < size-1; src++) {
1072 smpi_mpi_request_free(&requests[src]);
1079 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1080 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1084 int rank = smpi_process_index();
1087 /* arbitrarily choose root as rank 0 */
1088 size = smpi_comm_size(comm);
1090 displs = xbt_new(int, size);
1091 for (i = 0; i < size; i++) {
1093 count += recvcounts[i];
1095 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1096 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1097 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1098 recvcounts[rank], datatype, 0, comm);
1103 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1104 void *recvbuf, int *recvcounts, int *displs,
1105 MPI_Datatype recvtype, int root, MPI_Comm comm)
1107 int system_tag = COLL_TAG_GATHERV;
1108 int rank, size, src, index;
1109 MPI_Aint lb = 0, recvext = 0;
1110 MPI_Request *requests;
1112 rank = smpi_comm_rank(comm);
1113 size = smpi_comm_size(comm);
1115 // Send buffer to root
1116 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1118 // FIXME: check for errors
1119 smpi_datatype_extent(recvtype, &lb, &recvext);
1120 // Local copy from root
1121 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1122 (char *)recvbuf + displs[root] * recvext,
1123 recvcounts[root], recvtype);
1124 // Receive buffers from senders
1125 requests = xbt_new(MPI_Request, size - 1);
1127 for(src = 0; src < size; src++) {
1130 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1131 recvcounts[src], recvtype, src, system_tag, comm);
1135 // Wait for completion of irecv's.
1136 smpi_mpi_startall(size - 1, requests);
1137 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1138 for(src = 0; src < size-1; src++) {
1139 smpi_mpi_request_free(&requests[src]);
1145 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1146 MPI_Datatype sendtype, void *recvbuf,
1147 int recvcount, MPI_Datatype recvtype,
1150 int system_tag = COLL_TAG_ALLGATHER;
1151 int rank, size, other, index;
1152 MPI_Aint lb = 0, recvext = 0;
1153 MPI_Request *requests;
1155 rank = smpi_comm_rank(comm);
1156 size = smpi_comm_size(comm);
1157 // FIXME: check for errors
1158 smpi_datatype_extent(recvtype, &lb, &recvext);
1159 // Local copy from self
1160 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1161 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1163 // Send/Recv buffers to/from others;
1164 requests = xbt_new(MPI_Request, 2 * (size - 1));
1166 for(other = 0; other < size; other++) {
1169 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1172 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1173 recvcount, recvtype, other,
1178 // Wait for completion of all comms.
1179 smpi_mpi_startall(2 * (size - 1), requests);
1180 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1181 for(other = 0; other < 2*(size-1); other++) {
1182 smpi_mpi_request_free(&requests[other]);
1187 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1188 MPI_Datatype sendtype, void *recvbuf,
1189 int *recvcounts, int *displs,
1190 MPI_Datatype recvtype, MPI_Comm comm)
1192 int system_tag = COLL_TAG_ALLGATHERV;
1193 int rank, size, other, index;
1194 MPI_Aint lb = 0, recvext = 0;
1195 MPI_Request *requests;
1197 rank = smpi_comm_rank(comm);
1198 size = smpi_comm_size(comm);
1199 // FIXME: check for errors
1200 smpi_datatype_extent(recvtype, &lb, &recvext);
1201 // Local copy from self
1202 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1203 (char *)recvbuf + displs[rank] * recvext,
1204 recvcounts[rank], recvtype);
1205 // Send buffers to others;
1206 requests = xbt_new(MPI_Request, 2 * (size - 1));
1208 for(other = 0; other < size; other++) {
1211 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1215 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1216 recvtype, other, system_tag, comm);
1220 // Wait for completion of all comms.
1221 smpi_mpi_startall(2 * (size - 1), requests);
1222 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1223 for(other = 0; other < 2*(size-1); other++) {
1224 smpi_mpi_request_free(&requests[other]);
1229 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1230 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1231 int root, MPI_Comm comm)
1233 int system_tag = COLL_TAG_SCATTER;
1234 int rank, size, dst, index;
1235 MPI_Aint lb = 0, sendext = 0;
1236 MPI_Request *requests;
1238 rank = smpi_comm_rank(comm);
1239 size = smpi_comm_size(comm);
1241 // Recv buffer from root
1242 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1245 // FIXME: check for errors
1246 smpi_datatype_extent(sendtype, &lb, &sendext);
1247 // Local copy from root
1248 if(recvbuf!=MPI_IN_PLACE){
1249 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1250 sendcount, sendtype, recvbuf, recvcount, recvtype);
1252 // Send buffers to receivers
1253 requests = xbt_new(MPI_Request, size - 1);
1255 for(dst = 0; dst < size; dst++) {
1257 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1258 sendcount, sendtype, dst,
1263 // Wait for completion of isend's.
1264 smpi_mpi_startall(size - 1, requests);
1265 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1266 for(dst = 0; dst < size-1; dst++) {
1267 smpi_mpi_request_free(&requests[dst]);
1273 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1274 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1275 MPI_Datatype recvtype, int root, MPI_Comm comm)
1277 int system_tag = COLL_TAG_SCATTERV;
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 + displs[root] * sendext, sendcounts[root],
1294 sendtype, recvbuf, recvcount, recvtype);
1296 // Send buffers to receivers
1297 requests = xbt_new(MPI_Request, size - 1);
1299 for(dst = 0; dst < size; dst++) {
1302 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1303 sendtype, dst, system_tag, comm);
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_reduce(void *sendbuf, void *recvbuf, int count,
1318 MPI_Datatype datatype, MPI_Op op, int root,
1321 int system_tag = COLL_TAG_REDUCE;
1322 int rank, size, src, index;
1323 MPI_Aint lb = 0, dataext = 0;
1324 MPI_Request *requests;
1328 char* sendtmpbuf = (char*) sendbuf;
1329 if( sendbuf == MPI_IN_PLACE ) {
1330 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1331 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1334 rank = smpi_comm_rank(comm);
1335 size = smpi_comm_size(comm);
1336 //non commutative case, use a working algo from openmpi
1337 if(!smpi_op_is_commute(op)){
1338 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1339 datatype, op, root, comm);
1344 // Send buffer to root
1345 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1347 // FIXME: check for errors
1348 smpi_datatype_extent(datatype, &lb, &dataext);
1349 // Local copy from root
1350 if (sendtmpbuf && recvbuf)
1351 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1352 // Receive buffers from senders
1353 //TODO: make a MPI_barrier here ?
1354 requests = xbt_new(MPI_Request, size - 1);
1355 tmpbufs = xbt_new(void *, size - 1);
1357 for(src = 0; src < size; src++) {
1359 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1361 tmpbufs[index] = xbt_malloc(count * dataext);
1363 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1368 // Wait for completion of irecv's.
1369 smpi_mpi_startall(size - 1, requests);
1370 for(src = 0; src < size - 1; src++) {
1371 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1372 XBT_DEBUG("finished waiting any request with index %d", index);
1373 if(index == MPI_UNDEFINED) {
1376 smpi_mpi_request_free(&requests[index]);
1378 if(op) /* op can be MPI_OP_NULL that does nothing */
1379 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1381 for(index = 0; index < size - 1; index++) {
1382 xbt_free(tmpbufs[index]);
1387 if( sendbuf == MPI_IN_PLACE ) {
1388 xbt_free(sendtmpbuf);
1393 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1394 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1396 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1397 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1400 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1401 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1403 int system_tag = -888;
1404 int rank, size, other, index;
1405 MPI_Aint lb = 0, dataext = 0;
1406 MPI_Request *requests;
1409 rank = smpi_comm_rank(comm);
1410 size = smpi_comm_size(comm);
1412 // FIXME: check for errors
1413 smpi_datatype_extent(datatype, &lb, &dataext);
1415 // Local copy from self
1416 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1418 // Send/Recv buffers to/from others;
1419 requests = xbt_new(MPI_Request, size - 1);
1420 tmpbufs = xbt_new(void *, rank);
1422 for(other = 0; other < rank; other++) {
1423 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1425 tmpbufs[index] = xbt_malloc(count * dataext);
1427 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1431 for(other = rank + 1; other < size; other++) {
1433 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1436 // Wait for completion of all comms.
1437 smpi_mpi_startall(size - 1, requests);
1439 if(smpi_op_is_commute(op)){
1440 for(other = 0; other < size - 1; other++) {
1441 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1442 if(index == MPI_UNDEFINED) {
1446 // #Request is below rank: it's a irecv
1447 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1451 //non commutative case, wait in order
1452 for(other = 0; other < size - 1; other++) {
1453 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1455 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1459 for(index = 0; index < rank; index++) {
1460 xbt_free(tmpbufs[index]);
1462 for(index = 0; index < size-1; index++) {
1463 smpi_mpi_request_free(&requests[index]);
1469 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1470 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1472 int system_tag = -888;
1473 int rank, size, other, index;
1474 MPI_Aint lb = 0, dataext = 0;
1475 MPI_Request *requests;
1477 int recvbuf_is_empty=1;
1478 rank = smpi_comm_rank(comm);
1479 size = smpi_comm_size(comm);
1481 // FIXME: check for errors
1482 smpi_datatype_extent(datatype, &lb, &dataext);
1484 // Send/Recv buffers to/from others;
1485 requests = xbt_new(MPI_Request, size - 1);
1486 tmpbufs = xbt_new(void *, rank);
1488 for(other = 0; other < rank; other++) {
1489 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1491 tmpbufs[index] = xbt_malloc(count * dataext);
1493 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1497 for(other = rank + 1; other < size; other++) {
1499 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1502 // Wait for completion of all comms.
1503 smpi_mpi_startall(size - 1, requests);
1504 if(smpi_op_is_commute(op)){
1505 for(other = 0; other < size - 1; other++) {
1506 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1507 if(index == MPI_UNDEFINED) {
1511 if(recvbuf_is_empty){
1512 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1515 // #Request is below rank: it's a irecv
1516 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1520 //non commutative case, wait in order
1521 for(other = 0; other < size - 1; other++) {
1522 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1524 if(recvbuf_is_empty){
1525 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1527 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1531 for(index = 0; index < rank; index++) {
1532 xbt_free(tmpbufs[index]);
1534 for(index = 0; index < size-1; index++) {
1535 smpi_mpi_request_free(&requests[index]);