1 /* Copyright (c) 2007, 2008, 2009, 2010. 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"
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 == 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
41 static int match_send(void* a, void* b,smx_action_t ignored) {
42 MPI_Request ref = (MPI_Request)a;
43 MPI_Request req = (MPI_Request)b;
44 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
45 xbt_assert(ref, "Cannot match send against null reference");
46 xbt_assert(req, "Cannot match send against null request");
48 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
49 && (req->tag == MPI_ANY_TAG || req->tag == ref->tag))
51 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
52 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
53 if(req->real_size < ref->real_size) req->truncated = 1;
55 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
63 typedef struct s_smpi_factor *smpi_factor_t;
64 typedef struct s_smpi_factor {
67 double values[4];//arbitrary set to 4
69 xbt_dynar_t smpi_os_values = NULL;
70 xbt_dynar_t smpi_or_values = NULL;
71 xbt_dynar_t smpi_ois_values = NULL;
73 // Methods used to parse and store the values for timing injections in smpi
74 // These are taken from surf/network.c and generalized to have more factors
75 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
77 static int factor_cmp(const void *pa, const void *pb)
79 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor);
83 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
86 unsigned int iter = 0;
89 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
91 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
92 radical_elements = xbt_str_split(smpi_coef_string, ";");
93 xbt_dynar_foreach(radical_elements, iter, value) {
95 radical_elements2 = xbt_str_split(value, ":");
96 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
97 xbt_die("Malformed radical for smpi factor!");
98 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
100 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
102 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
107 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
108 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
109 xbt_dynar_free(&radical_elements2);
111 xbt_dynar_free(&radical_elements);
113 xbt_dynar_sort(smpi_factor, &factor_cmp);
114 xbt_dynar_foreach(smpi_factor, iter, fact) {
115 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
120 static double smpi_os(double size)
122 if (!smpi_os_values) {
123 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
124 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
126 unsigned int iter = 0;
127 s_smpi_factor_t fact;
129 xbt_dynar_foreach(smpi_os_values, iter, fact) {
130 if (size <= fact.factor) {
131 XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
134 current=fact.values[0]+fact.values[1]*size;
137 XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
142 static double smpi_ois(double size)
144 if (!smpi_ois_values) {
145 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
146 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
148 unsigned int iter = 0;
149 s_smpi_factor_t fact;
151 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
152 if (size <= fact.factor) {
153 XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
156 current=fact.values[0]+fact.values[1]*size;
159 XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
164 static double smpi_or(double size)
166 if (!smpi_or_values) {
167 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
168 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
170 unsigned int iter = 0;
171 s_smpi_factor_t fact;
173 xbt_dynar_foreach(smpi_or_values, iter, fact) {
174 if (size <= fact.factor) {
175 XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
178 current=fact.values[0]+fact.values[1]*size;
180 XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
185 static MPI_Request build_request(void *buf, int count,
186 MPI_Datatype datatype, int src, int dst,
187 int tag, MPI_Comm comm, unsigned flags)
191 void *old_buf = NULL;
193 request = xbt_new(s_smpi_mpi_request_t, 1);
195 s_smpi_subtype_t *subtype = datatype->substruct;
197 if(datatype->has_subtype == 1){
198 // This part handles the problem of non-contiguous memory
200 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
202 subtype->serialize(old_buf, buf, count, datatype->substruct);
207 // This part handles the problem of non-contiguous memory (for the
208 // unserialisation at the reception)
209 request->old_buf = old_buf;
210 request->old_type = datatype;
212 request->size = smpi_datatype_size(datatype) * count;
216 request->comm = comm;
217 request->action = NULL;
218 request->flags = flags;
219 request->detached = 0;
220 request->detached_sender = NULL;
222 request->truncated = 0;
223 request->real_size = 0;
224 request->real_tag = 0;
231 if (flags & SEND) smpi_datatype_unuse(datatype);
237 void smpi_empty_status(MPI_Status * status)
239 if(status != MPI_STATUS_IGNORE) {
240 status->MPI_SOURCE = MPI_ANY_SOURCE;
241 status->MPI_TAG = MPI_ANY_TAG;
242 status->MPI_ERROR = MPI_SUCCESS;
247 void smpi_action_trace_run(char *path)
251 xbt_dict_cursor_t cursor;
255 action_fp = fopen(path, "r");
256 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
260 if (!xbt_dict_is_empty(action_queues)) {
262 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
265 xbt_dict_foreach(action_queues, cursor, name, todo) {
266 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
272 xbt_dict_free(&action_queues);
273 action_queues = xbt_dict_new_homogeneous(NULL);
276 static void smpi_mpi_request_free_voidp(void* request)
278 MPI_Request req = request;
279 smpi_mpi_request_free(&req);
282 /* MPI Low level calls */
283 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
284 int dst, int tag, MPI_Comm comm)
286 MPI_Request request =
287 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
288 comm, PERSISTENT | SEND);
293 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
294 int dst, int tag, MPI_Comm comm)
296 MPI_Request request =
297 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
298 comm, PERSISTENT | SSEND | SEND);
303 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
304 int src, int tag, MPI_Comm comm)
306 MPI_Request request =
307 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
308 comm, PERSISTENT | RECV);
313 void smpi_mpi_start(MPI_Request request)
317 xbt_assert(!request->action,
318 "Cannot (re)start a non-finished communication");
319 if(request->flags & RECV) {
320 print_request("New recv", request);
321 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
322 mailbox = smpi_process_mailbox_small();
324 mailbox = smpi_process_mailbox();
325 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
326 request->real_size=request->size;
327 smpi_datatype_use(request->old_type);
328 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
330 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
331 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
333 simcall_process_sleep(sleeptime);
334 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
340 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
343 int rank = smpi_process_index();
344 if (TRACE_smpi_view_internals()) {
345 TRACE_smpi_send(rank, rank, receiver);
348 /* if(receiver == MPI_UNDEFINED) {*/
349 /* XBT_WARN("Trying to send a message to a wrong rank");*/
352 print_request("New send", request);
353 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
354 mailbox = smpi_process_remote_mailbox_small(receiver);
356 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
357 mailbox = smpi_process_remote_mailbox(receiver);
359 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
361 request->detached = 1;
363 if(request->old_type->has_subtype == 0){
364 oldbuf = request->buf;
365 if (oldbuf && request->size!=0){
366 request->buf = xbt_malloc(request->size);
367 memcpy(request->buf,oldbuf,request->size);
370 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
373 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
374 request->real_size=request->size;
375 smpi_datatype_use(request->old_type);
377 //if we are giving back the control to the user without waiting for completion, we have to inject timings
378 double sleeptime =0.0;
379 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
380 //isend and send timings may be different
381 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
385 simcall_process_sleep(sleeptime);
386 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
390 simcall_comm_isend(mailbox, request->size, -1.0,
391 request->buf, request->real_size,
393 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
395 // detach if msg size < eager/rdv switch limit
399 /* FIXME: detached sends are not traceable (request->action == NULL) */
401 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
409 void smpi_mpi_startall(int count, MPI_Request * requests)
413 for(i = 0; i < count; i++) {
414 smpi_mpi_start(requests[i]);
418 void smpi_mpi_request_free(MPI_Request * request)
421 if((*request) != MPI_REQUEST_NULL){
422 (*request)->refcount--;
423 if((*request)->refcount<0) xbt_die("wrong refcount");
425 if((*request)->refcount==0){
426 print_request("Destroying", (*request));
428 *request = MPI_REQUEST_NULL;
430 print_request("Decrementing", (*request));
434 xbt_die("freeing an already free request");
438 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
439 int dst, int tag, MPI_Comm comm)
441 MPI_Request request =
442 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
443 comm, NON_PERSISTENT | SEND);
448 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
449 int dst, int tag, MPI_Comm comm)
451 MPI_Request request =
452 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
453 comm, NON_PERSISTENT | ISEND | SEND);
455 smpi_mpi_start(request);
459 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
460 int dst, int tag, MPI_Comm comm)
462 MPI_Request request =
463 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
464 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
465 smpi_mpi_start(request);
471 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
472 int src, int tag, MPI_Comm comm)
474 MPI_Request request =
475 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
476 comm, NON_PERSISTENT | RECV);
480 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
481 int src, int tag, MPI_Comm comm)
483 MPI_Request request =
484 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
485 comm, NON_PERSISTENT | RECV);
487 smpi_mpi_start(request);
491 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
492 int tag, MPI_Comm comm, MPI_Status * status)
495 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
496 smpi_mpi_wait(&request, status);
501 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
502 int tag, MPI_Comm comm)
504 MPI_Request request =
505 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
506 comm, NON_PERSISTENT | SEND);
507 smpi_mpi_start(request);
508 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
512 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
513 int dst, int tag, MPI_Comm comm)
515 MPI_Request request =
516 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
517 comm, NON_PERSISTENT | SSEND | SEND);
519 smpi_mpi_start(request);
520 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
523 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
524 int dst, int sendtag, void *recvbuf, int recvcount,
525 MPI_Datatype recvtype, int src, int recvtag,
526 MPI_Comm comm, MPI_Status * status)
528 MPI_Request requests[2];
532 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
534 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
535 smpi_mpi_startall(2, requests);
536 smpi_mpi_waitall(2, requests, stats);
537 if(status != MPI_STATUS_IGNORE) {
538 // Copy receive status
543 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
545 return status->count / smpi_datatype_size(datatype);
548 static void finish_wait(MPI_Request * request, MPI_Status * status)
550 MPI_Request req = *request;
551 if(status != MPI_STATUS_IGNORE)
552 smpi_empty_status(status);
554 if(!(req->detached && req->flags & SEND)){
555 if(status != MPI_STATUS_IGNORE) {
556 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
557 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
558 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
559 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
560 // this handles the case were size in receive differs from size in send
561 // FIXME: really this should just contain the count of receive-type blocks,
563 status->count = req->real_size;
566 print_request("Finishing", req);
567 MPI_Datatype datatype = req->old_type;
569 if(datatype->has_subtype == 1){
570 // This part handles the problem of non-contignous memory
571 // the unserialization at the reception
572 s_smpi_subtype_t *subtype = datatype->substruct;
573 if(req->flags & RECV) {
574 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
576 if(req->detached == 0) free(req->buf);
578 smpi_datatype_unuse(datatype);
583 if (TRACE_smpi_view_internals()) {
584 if(req->flags & RECV){
585 int rank = smpi_process_index();
586 int src_traced = smpi_group_index(smpi_comm_group(req->comm), req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
587 TRACE_smpi_recv(rank, src_traced, rank);
592 if(req->detached_sender!=NULL){
593 smpi_mpi_request_free(&(req->detached_sender));
596 if(req->flags & NON_PERSISTENT) {
597 smpi_mpi_request_free(request);
603 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
606 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
607 if ((*request)->action == NULL)
610 flag = simcall_comm_test((*request)->action);
612 finish_wait(request, status);
613 request=MPI_REQUEST_NULL;
615 smpi_empty_status(status);
620 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
627 *index = MPI_UNDEFINED;
630 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
631 map = xbt_new(int, count);
633 for(i = 0; i < count; i++) {
634 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
635 xbt_dynar_push(comms, &requests[i]->action);
641 i = simcall_comm_testany(comms);
642 // not MPI_UNDEFINED, as this is a simix return code
645 finish_wait(&requests[*index], status);
649 //all requests are null or inactive, return true
651 smpi_empty_status(status);
654 xbt_dynar_free(&comms);
661 int smpi_mpi_testall(int count, MPI_Request requests[],
665 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
668 for(i=0; i<count; i++){
669 if(requests[i]!= MPI_REQUEST_NULL){
670 if (smpi_mpi_test(&requests[i], pstat)!=1){
674 smpi_empty_status(pstat);
676 if(status != MPI_STATUSES_IGNORE) {
683 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
685 //FIXME find another wait to avoid busy waiting ?
686 // the issue here is that we have to wait on a nonexistent comm
688 smpi_mpi_iprobe(source, tag, comm, &flag, status);
689 XBT_DEBUG("Busy Waiting on probing : %d", flag);
693 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
695 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,
696 comm, NON_PERSISTENT | RECV);
698 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
699 double sleeptime= sg_cfg_get_double("smpi/iprobe");
700 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
701 static int nsleeps = 1;
703 simcall_process_sleep(sleeptime);
705 // behave like a receive, but don't do it
708 print_request("New iprobe", request);
709 // We have to test both mailboxes as we don't know if we will receive one one or another
710 if (sg_cfg_get_int("smpi/async_small_thres")>0){
711 mailbox = smpi_process_mailbox_small();
712 XBT_DEBUG("trying to probe the perm recv mailbox");
713 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
715 if (request->action==NULL){
716 mailbox = smpi_process_mailbox();
717 XBT_DEBUG("trying to probe the other mailbox");
718 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
722 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
724 if(status != MPI_STATUS_IGNORE) {
725 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
726 status->MPI_TAG = req->tag;
727 status->MPI_ERROR = MPI_SUCCESS;
728 status->count = req->real_size;
730 nsleeps=1;//reset the number of sleeps we will do next time
736 smpi_mpi_request_free(&request);
741 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
743 print_request("Waiting", *request);
744 if ((*request)->action != NULL) { // this is not a detached send
745 simcall_comm_wait((*request)->action, -1.0);
747 finish_wait(request, status);
749 // FIXME for a detached send, finish_wait is not called:
752 int smpi_mpi_waitany(int count, MPI_Request requests[],
759 index = MPI_UNDEFINED;
761 // Wait for a request to complete
762 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
763 map = xbt_new(int, count);
765 XBT_DEBUG("Wait for one of %d", count);
766 for(i = 0; i < count; i++) {
767 if(requests[i] != MPI_REQUEST_NULL) {
768 if (requests[i]->action != NULL) {
769 XBT_DEBUG("Waiting any %p ", requests[i]);
770 xbt_dynar_push(comms, &requests[i]->action);
774 //This is a finished detached request, let's return this one
775 size=0;//so we free the dynar but don't do the waitany call
777 finish_wait(&requests[i], status);//cleanup if refcount = 0
778 requests[i]=MPI_REQUEST_NULL;//set to null
784 i = simcall_comm_waitany(comms);
786 // not MPI_UNDEFINED, as this is a simix return code
789 finish_wait(&requests[index], status);
793 xbt_dynar_free(&comms);
796 if (index==MPI_UNDEFINED)
797 smpi_empty_status(status);
802 int smpi_mpi_waitall(int count, MPI_Request requests[],
807 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
808 int retvalue = MPI_SUCCESS;
809 //tag invalid requests in the set
810 if (status != MPI_STATUSES_IGNORE) {
811 for (c = 0; c < count; c++) {
812 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
813 smpi_empty_status(&status[c]);
814 } else if (requests[c]->src == MPI_PROC_NULL) {
815 smpi_empty_status(&status[c]);
816 status[c].MPI_SOURCE = MPI_PROC_NULL;
820 for(c = 0; c < count; c++) {
821 if (MC_is_active()) {
822 smpi_mpi_wait(&requests[c], pstat);
825 index = smpi_mpi_waitany(count, requests, pstat);
826 if (index == MPI_UNDEFINED)
828 requests[index]=MPI_REQUEST_NULL;
830 if (status != MPI_STATUSES_IGNORE) {
831 status[index] = *pstat;
832 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
833 retvalue = MPI_ERR_IN_STATUS;
840 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
845 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
848 for(i = 0; i < incount; i++)
850 index=smpi_mpi_waitany(incount, requests, pstat);
851 if(index!=MPI_UNDEFINED){
852 indices[count] = index;
854 if(status != MPI_STATUSES_IGNORE) {
855 status[index] = *pstat;
857 requests[index]=MPI_REQUEST_NULL;
859 return MPI_UNDEFINED;
865 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
868 int i, count, count_dead;
870 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
874 for(i = 0; i < incount; i++) {
875 if((requests[i] != MPI_REQUEST_NULL)) {
876 if(smpi_mpi_test(&requests[i], pstat)) {
879 if(status != MPI_STATUSES_IGNORE) {
882 requests[i]=MPI_REQUEST_NULL;
889 if(count_dead==incount)return MPI_UNDEFINED;
893 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
896 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
897 nary_tree_bcast(buf, count, datatype, root, comm, 4);
900 void smpi_mpi_barrier(MPI_Comm comm)
902 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
903 nary_tree_barrier(comm, 4);
906 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
907 void *recvbuf, int recvcount, MPI_Datatype recvtype,
908 int root, MPI_Comm comm)
910 int system_tag = 666;
911 int rank, size, src, index;
912 MPI_Aint lb = 0, recvext = 0;
913 MPI_Request *requests;
915 rank = smpi_comm_rank(comm);
916 size = smpi_comm_size(comm);
918 // Send buffer to root
919 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
921 // FIXME: check for errors
922 smpi_datatype_extent(recvtype, &lb, &recvext);
923 // Local copy from root
924 smpi_datatype_copy(sendbuf, sendcount, sendtype,
925 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
926 // Receive buffers from senders
927 requests = xbt_new(MPI_Request, size - 1);
929 for(src = 0; src < size; src++) {
931 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
933 src, system_tag, comm);
937 // Wait for completion of irecv's.
938 smpi_mpi_startall(size - 1, requests);
939 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
945 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
946 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
950 int rank = smpi_process_index();
951 /* arbitrarily choose root as rank 0 */
952 size = smpi_comm_size(comm);
954 displs = xbt_new(int, size);
955 for (i = 0; i < size; i++) {
957 count += recvcounts[i];
959 mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, 0, comm);
960 smpi_mpi_scatterv(recvbuf, recvcounts, displs, datatype, recvbuf,
961 recvcounts[rank], datatype, 0, comm);
965 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
966 void *recvbuf, int *recvcounts, int *displs,
967 MPI_Datatype recvtype, int root, MPI_Comm comm)
969 int system_tag = 666;
970 int rank, size, src, index;
971 MPI_Aint lb = 0, recvext = 0;
972 MPI_Request *requests;
974 rank = smpi_comm_rank(comm);
975 size = smpi_comm_size(comm);
977 // Send buffer to root
978 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
980 // FIXME: check for errors
981 smpi_datatype_extent(recvtype, &lb, &recvext);
982 // Local copy from root
983 smpi_datatype_copy(sendbuf, sendcount, sendtype,
984 (char *)recvbuf + displs[root] * recvext,
985 recvcounts[root], recvtype);
986 // Receive buffers from senders
987 requests = xbt_new(MPI_Request, size - 1);
989 for(src = 0; src < size; src++) {
992 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
993 recvcounts[src], recvtype, src, system_tag, comm);
997 // Wait for completion of irecv's.
998 smpi_mpi_startall(size - 1, requests);
999 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1004 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1005 MPI_Datatype sendtype, void *recvbuf,
1006 int recvcount, MPI_Datatype recvtype,
1009 int system_tag = 666;
1010 int rank, size, other, index;
1011 MPI_Aint lb = 0, recvext = 0;
1012 MPI_Request *requests;
1014 rank = smpi_comm_rank(comm);
1015 size = smpi_comm_size(comm);
1016 // FIXME: check for errors
1017 smpi_datatype_extent(recvtype, &lb, &recvext);
1018 // Local copy from self
1019 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1020 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1022 // Send/Recv buffers to/from others;
1023 requests = xbt_new(MPI_Request, 2 * (size - 1));
1025 for(other = 0; other < size; other++) {
1028 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1031 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1032 recvcount, recvtype, other,
1037 // Wait for completion of all comms.
1038 smpi_mpi_startall(2 * (size - 1), requests);
1039 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1043 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1044 MPI_Datatype sendtype, void *recvbuf,
1045 int *recvcounts, int *displs,
1046 MPI_Datatype recvtype, MPI_Comm comm)
1048 int system_tag = 666;
1049 int rank, size, other, index;
1050 MPI_Aint lb = 0, recvext = 0;
1051 MPI_Request *requests;
1053 rank = smpi_comm_rank(comm);
1054 size = smpi_comm_size(comm);
1055 // FIXME: check for errors
1056 smpi_datatype_extent(recvtype, &lb, &recvext);
1057 // Local copy from self
1058 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1059 (char *)recvbuf + displs[rank] * recvext,
1060 recvcounts[rank], recvtype);
1061 // Send buffers to others;
1062 requests = xbt_new(MPI_Request, 2 * (size - 1));
1064 for(other = 0; other < size; other++) {
1067 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1071 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1072 recvtype, other, system_tag, comm);
1076 // Wait for completion of all comms.
1077 smpi_mpi_startall(2 * (size - 1), requests);
1078 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1082 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1083 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1084 int root, MPI_Comm comm)
1086 int system_tag = 666;
1087 int rank, size, dst, index;
1088 MPI_Aint lb = 0, sendext = 0;
1089 MPI_Request *requests;
1091 rank = smpi_comm_rank(comm);
1092 size = smpi_comm_size(comm);
1094 // Recv buffer from root
1095 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1098 // FIXME: check for errors
1099 smpi_datatype_extent(sendtype, &lb, &sendext);
1100 // Local copy from root
1101 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1102 sendcount, sendtype, recvbuf, recvcount, recvtype);
1103 // Send buffers to receivers
1104 requests = xbt_new(MPI_Request, size - 1);
1106 for(dst = 0; dst < size; dst++) {
1108 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1109 sendcount, sendtype, dst,
1114 // Wait for completion of isend's.
1115 smpi_mpi_startall(size - 1, requests);
1116 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1121 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1122 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1123 MPI_Datatype recvtype, int root, MPI_Comm comm)
1125 int system_tag = 666;
1126 int rank, size, dst, index;
1127 MPI_Aint lb = 0, sendext = 0;
1128 MPI_Request *requests;
1130 rank = smpi_comm_rank(comm);
1131 size = smpi_comm_size(comm);
1133 // Recv buffer from root
1134 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1137 // FIXME: check for errors
1138 smpi_datatype_extent(sendtype, &lb, &sendext);
1139 // Local copy from root
1140 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1141 sendtype, recvbuf, recvcount, recvtype);
1142 // Send buffers to receivers
1143 requests = xbt_new(MPI_Request, size - 1);
1145 for(dst = 0; dst < size; dst++) {
1148 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1149 sendtype, dst, system_tag, comm);
1153 // Wait for completion of isend's.
1154 smpi_mpi_startall(size - 1, requests);
1155 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1160 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1161 MPI_Datatype datatype, MPI_Op op, int root,
1164 int system_tag = 666;
1165 int rank, size, src, index;
1166 MPI_Aint lb = 0, dataext = 0;
1167 MPI_Request *requests;
1170 rank = smpi_comm_rank(comm);
1171 size = smpi_comm_size(comm);
1173 // Send buffer to root
1174 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
1176 // FIXME: check for errors
1177 smpi_datatype_extent(datatype, &lb, &dataext);
1178 // Local copy from root
1179 if (sendbuf && recvbuf)
1180 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1181 // Receive buffers from senders
1182 //TODO: make a MPI_barrier here ?
1183 requests = xbt_new(MPI_Request, size - 1);
1184 tmpbufs = xbt_new(void *, size - 1);
1186 for(src = 0; src < size; src++) {
1188 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1190 tmpbufs[index] = xbt_malloc(count * dataext);
1192 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1197 // Wait for completion of irecv's.
1198 smpi_mpi_startall(size - 1, requests);
1199 for(src = 0; src < size - 1; src++) {
1200 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1201 XBT_DEBUG("finished waiting any request with index %d", index);
1202 if(index == MPI_UNDEFINED) {
1205 if(op) /* op can be MPI_OP_NULL that does nothing */
1206 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1208 for(index = 0; index < size - 1; index++) {
1209 xbt_free(tmpbufs[index]);
1216 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1217 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1219 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1220 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1223 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1224 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1226 int system_tag = 666;
1227 int rank, size, other, index;
1228 MPI_Aint lb = 0, dataext = 0;
1229 MPI_Request *requests;
1232 rank = smpi_comm_rank(comm);
1233 size = smpi_comm_size(comm);
1235 // FIXME: check for errors
1236 smpi_datatype_extent(datatype, &lb, &dataext);
1238 // Local copy from self
1239 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1241 // Send/Recv buffers to/from others;
1242 requests = xbt_new(MPI_Request, size - 1);
1243 tmpbufs = xbt_new(void *, rank);
1245 for(other = 0; other < rank; other++) {
1246 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1248 tmpbufs[index] = xbt_malloc(count * dataext);
1250 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1254 for(other = rank + 1; other < size; other++) {
1256 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1259 // Wait for completion of all comms.
1260 smpi_mpi_startall(size - 1, requests);
1261 for(other = 0; other < size - 1; other++) {
1262 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1263 if(index == MPI_UNDEFINED) {
1267 // #Request is below rank: it's a irecv
1268 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1271 for(index = 0; index < rank; index++) {
1272 xbt_free(tmpbufs[index]);