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"
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);
84 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
87 unsigned int iter = 0;
90 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
92 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
93 radical_elements = xbt_str_split(smpi_coef_string, ";");
94 xbt_dynar_foreach(radical_elements, iter, value) {
96 radical_elements2 = xbt_str_split(value, ":");
97 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
98 xbt_die("Malformed radical for smpi factor!");
99 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
101 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
103 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
108 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
109 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
110 xbt_dynar_free(&radical_elements2);
112 xbt_dynar_free(&radical_elements);
114 xbt_dynar_sort(smpi_factor, &factor_cmp);
115 xbt_dynar_foreach(smpi_factor, iter, fact) {
116 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
121 static double smpi_os(double size)
123 if (!smpi_os_values) {
124 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
125 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
127 unsigned int iter = 0;
128 s_smpi_factor_t fact;
130 xbt_dynar_foreach(smpi_os_values, iter, fact) {
131 if (size <= fact.factor) {
132 XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
135 current=fact.values[0]+fact.values[1]*size;
138 XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
143 static double smpi_ois(double size)
145 if (!smpi_ois_values) {
146 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
147 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
149 unsigned int iter = 0;
150 s_smpi_factor_t fact;
152 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
153 if (size <= fact.factor) {
154 XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
157 current=fact.values[0]+fact.values[1]*size;
160 XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
165 static double smpi_or(double size)
167 if (!smpi_or_values) {
168 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
169 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
171 unsigned int iter = 0;
172 s_smpi_factor_t fact;
174 xbt_dynar_foreach(smpi_or_values, iter, fact) {
175 if (size <= fact.factor) {
176 XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
179 current=fact.values[0]+fact.values[1]*size;
181 XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
186 static MPI_Request build_request(void *buf, int count,
187 MPI_Datatype datatype, int src, int dst,
188 int tag, MPI_Comm comm, unsigned flags)
192 void *old_buf = NULL;
194 request = xbt_new(s_smpi_mpi_request_t, 1);
196 s_smpi_subtype_t *subtype = datatype->substruct;
198 if(datatype->has_subtype == 1){
199 // This part handles the problem of non-contiguous memory
201 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
203 subtype->serialize(old_buf, buf, count, datatype->substruct);
208 // This part handles the problem of non-contiguous memory (for the
209 // unserialisation at the reception)
210 request->old_buf = old_buf;
211 request->old_type = datatype;
213 request->size = smpi_datatype_size(datatype) * count;
217 request->comm = comm;
218 request->action = NULL;
219 request->flags = flags;
220 request->detached = 0;
221 request->detached_sender = NULL;
223 request->truncated = 0;
224 request->real_size = 0;
225 request->real_tag = 0;
232 if (flags & SEND) smpi_datatype_unuse(datatype);
238 void smpi_empty_status(MPI_Status * status)
240 if(status != MPI_STATUS_IGNORE) {
241 status->MPI_SOURCE = MPI_ANY_SOURCE;
242 status->MPI_TAG = MPI_ANY_TAG;
243 status->MPI_ERROR = MPI_SUCCESS;
248 void smpi_action_trace_run(char *path)
252 xbt_dict_cursor_t cursor;
256 action_fp = fopen(path, "r");
257 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
261 if (!xbt_dict_is_empty(action_queues)) {
263 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
266 xbt_dict_foreach(action_queues, cursor, name, todo) {
267 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
273 xbt_dict_free(&action_queues);
274 action_queues = xbt_dict_new_homogeneous(NULL);
277 static void smpi_mpi_request_free_voidp(void* request)
279 MPI_Request req = request;
280 smpi_mpi_request_free(&req);
283 /* MPI Low level calls */
284 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
285 int dst, int tag, MPI_Comm comm)
287 MPI_Request request =
288 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
289 comm, PERSISTENT | SEND);
294 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
295 int dst, int tag, MPI_Comm comm)
297 MPI_Request request =
298 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
299 comm, PERSISTENT | SSEND | SEND);
304 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
305 int src, int tag, MPI_Comm comm)
307 MPI_Request request =
308 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
309 comm, PERSISTENT | RECV);
314 void smpi_mpi_start(MPI_Request request)
318 xbt_assert(!request->action,
319 "Cannot (re)start a non-finished communication");
320 if(request->flags & RECV) {
321 print_request("New recv", request);
322 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
323 mailbox = smpi_process_mailbox_small();
325 mailbox = smpi_process_mailbox();
326 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
327 request->real_size=request->size;
328 smpi_datatype_use(request->old_type);
329 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
331 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
332 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
334 simcall_process_sleep(sleeptime);
335 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
341 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
344 int rank = smpi_process_index();
345 if (TRACE_smpi_view_internals()) {
346 TRACE_smpi_send(rank, rank, receiver);
349 /* if(receiver == MPI_UNDEFINED) {*/
350 /* XBT_WARN("Trying to send a message to a wrong rank");*/
353 print_request("New send", request);
354 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
355 mailbox = smpi_process_remote_mailbox_small(receiver);
357 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
358 mailbox = smpi_process_remote_mailbox(receiver);
360 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
362 request->detached = 1;
364 if(request->old_type->has_subtype == 0){
365 oldbuf = request->buf;
366 if (oldbuf && request->size!=0){
367 request->buf = xbt_malloc(request->size);
368 memcpy(request->buf,oldbuf,request->size);
371 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
374 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
375 request->real_size=request->size;
376 smpi_datatype_use(request->old_type);
378 //if we are giving back the control to the user without waiting for completion, we have to inject timings
379 double sleeptime =0.0;
380 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
381 //isend and send timings may be different
382 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
386 simcall_process_sleep(sleeptime);
387 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
391 simcall_comm_isend(mailbox, request->size, -1.0,
392 request->buf, request->real_size,
394 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
396 // detach if msg size < eager/rdv switch limit
400 /* FIXME: detached sends are not traceable (request->action == NULL) */
402 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
410 void smpi_mpi_startall(int count, MPI_Request * requests)
414 for(i = 0; i < count; i++) {
415 smpi_mpi_start(requests[i]);
419 void smpi_mpi_request_free(MPI_Request * request)
422 if((*request) != MPI_REQUEST_NULL){
423 (*request)->refcount--;
424 if((*request)->refcount<0) xbt_die("wrong refcount");
426 if((*request)->refcount==0){
427 print_request("Destroying", (*request));
429 *request = MPI_REQUEST_NULL;
431 print_request("Decrementing", (*request));
435 xbt_die("freeing an already free request");
439 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
440 int dst, int tag, MPI_Comm comm)
442 MPI_Request request =
443 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
444 comm, NON_PERSISTENT | SEND);
449 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
450 int dst, int tag, MPI_Comm comm)
452 MPI_Request request =
453 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
454 comm, NON_PERSISTENT | ISEND | SEND);
456 smpi_mpi_start(request);
460 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
461 int dst, int tag, MPI_Comm comm)
463 MPI_Request request =
464 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
465 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
466 smpi_mpi_start(request);
472 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
473 int src, int tag, MPI_Comm comm)
475 MPI_Request request =
476 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
477 comm, NON_PERSISTENT | RECV);
481 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
482 int src, int tag, MPI_Comm comm)
484 MPI_Request request =
485 build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
486 comm, NON_PERSISTENT | RECV);
488 smpi_mpi_start(request);
492 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
493 int tag, MPI_Comm comm, MPI_Status * status)
496 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
497 smpi_mpi_wait(&request, status);
502 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
503 int tag, MPI_Comm comm)
505 MPI_Request request =
506 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
507 comm, NON_PERSISTENT | SEND);
508 smpi_mpi_start(request);
509 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
513 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
514 int dst, int tag, MPI_Comm comm)
516 MPI_Request request =
517 build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
518 comm, NON_PERSISTENT | SSEND | SEND);
520 smpi_mpi_start(request);
521 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
524 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
525 int dst, int sendtag, void *recvbuf, int recvcount,
526 MPI_Datatype recvtype, int src, int recvtag,
527 MPI_Comm comm, MPI_Status * status)
529 MPI_Request requests[2];
531 int myid=smpi_process_index();
532 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
533 smpi_datatype_copy(sendbuf, sendcount, sendtype,
534 recvbuf, recvcount, recvtype);
538 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
540 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
541 smpi_mpi_startall(2, requests);
542 smpi_mpi_waitall(2, requests, stats);
543 if(status != MPI_STATUS_IGNORE) {
544 // Copy receive status
549 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
551 return status->count / smpi_datatype_size(datatype);
554 static void finish_wait(MPI_Request * request, MPI_Status * status)
556 MPI_Request req = *request;
557 if(status != MPI_STATUS_IGNORE)
558 smpi_empty_status(status);
560 if(!(req->detached && req->flags & SEND)){
561 if(status != MPI_STATUS_IGNORE) {
562 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
563 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
564 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
565 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
566 // this handles the case were size in receive differs from size in send
567 // FIXME: really this should just contain the count of receive-type blocks,
569 status->count = req->real_size;
572 print_request("Finishing", req);
573 MPI_Datatype datatype = req->old_type;
575 if(datatype->has_subtype == 1){
576 // This part handles the problem of non-contignous memory
577 // the unserialization at the reception
578 s_smpi_subtype_t *subtype = datatype->substruct;
579 if(req->flags & RECV) {
580 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
582 if(req->detached == 0) free(req->buf);
584 smpi_datatype_unuse(datatype);
589 if (TRACE_smpi_view_internals()) {
590 if(req->flags & RECV){
591 int rank = smpi_process_index();
592 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
593 TRACE_smpi_recv(rank, src_traced, rank);
598 if(req->detached_sender!=NULL){
599 smpi_mpi_request_free(&(req->detached_sender));
602 if(req->flags & NON_PERSISTENT) {
603 smpi_mpi_request_free(request);
609 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
612 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
613 if ((*request)->action == NULL)
616 flag = simcall_comm_test((*request)->action);
618 finish_wait(request, status);
619 request=MPI_REQUEST_NULL;
621 smpi_empty_status(status);
626 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
633 *index = MPI_UNDEFINED;
636 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
637 map = xbt_new(int, count);
639 for(i = 0; i < count; i++) {
640 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
641 xbt_dynar_push(comms, &requests[i]->action);
647 i = simcall_comm_testany(comms);
648 // not MPI_UNDEFINED, as this is a simix return code
651 finish_wait(&requests[*index], status);
655 //all requests are null or inactive, return true
657 smpi_empty_status(status);
660 xbt_dynar_free(&comms);
667 int smpi_mpi_testall(int count, MPI_Request requests[],
671 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
674 for(i=0; i<count; i++){
675 if(requests[i]!= MPI_REQUEST_NULL){
676 if (smpi_mpi_test(&requests[i], pstat)!=1){
680 smpi_empty_status(pstat);
682 if(status != MPI_STATUSES_IGNORE) {
689 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
691 //FIXME find another wait to avoid busy waiting ?
692 // the issue here is that we have to wait on a nonexistent comm
694 smpi_mpi_iprobe(source, tag, comm, &flag, status);
695 XBT_DEBUG("Busy Waiting on probing : %d", flag);
699 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
701 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,
702 comm, NON_PERSISTENT | RECV);
704 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
705 double sleeptime= sg_cfg_get_double("smpi/iprobe");
706 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
707 static int nsleeps = 1;
709 simcall_process_sleep(sleeptime);
711 // behave like a receive, but don't do it
714 print_request("New iprobe", request);
715 // We have to test both mailboxes as we don't know if we will receive one one or another
716 if (sg_cfg_get_int("smpi/async_small_thres")>0){
717 mailbox = smpi_process_mailbox_small();
718 XBT_DEBUG("trying to probe the perm recv mailbox");
719 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
721 if (request->action==NULL){
722 mailbox = smpi_process_mailbox();
723 XBT_DEBUG("trying to probe the other mailbox");
724 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
728 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
730 if(status != MPI_STATUS_IGNORE) {
731 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
732 status->MPI_TAG = req->tag;
733 status->MPI_ERROR = MPI_SUCCESS;
734 status->count = req->real_size;
736 nsleeps=1;//reset the number of sleeps we will do next time
742 smpi_mpi_request_free(&request);
747 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
749 print_request("Waiting", *request);
750 if ((*request)->action != NULL) { // this is not a detached send
751 simcall_comm_wait((*request)->action, -1.0);
753 finish_wait(request, status);
755 // FIXME for a detached send, finish_wait is not called:
758 int smpi_mpi_waitany(int count, MPI_Request requests[],
765 index = MPI_UNDEFINED;
767 // Wait for a request to complete
768 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
769 map = xbt_new(int, count);
771 XBT_DEBUG("Wait for one of %d", count);
772 for(i = 0; i < count; i++) {
773 if(requests[i] != MPI_REQUEST_NULL) {
774 if (requests[i]->action != NULL) {
775 XBT_DEBUG("Waiting any %p ", requests[i]);
776 xbt_dynar_push(comms, &requests[i]->action);
780 //This is a finished detached request, let's return this one
781 size=0;//so we free the dynar but don't do the waitany call
783 finish_wait(&requests[i], status);//cleanup if refcount = 0
784 requests[i]=MPI_REQUEST_NULL;//set to null
790 i = simcall_comm_waitany(comms);
792 // not MPI_UNDEFINED, as this is a simix return code
795 finish_wait(&requests[index], status);
799 xbt_dynar_free(&comms);
802 if (index==MPI_UNDEFINED)
803 smpi_empty_status(status);
808 int smpi_mpi_waitall(int count, MPI_Request requests[],
813 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
814 int retvalue = MPI_SUCCESS;
815 //tag invalid requests in the set
816 if (status != MPI_STATUSES_IGNORE) {
817 for (c = 0; c < count; c++) {
818 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
819 smpi_empty_status(&status[c]);
820 } else if (requests[c]->src == MPI_PROC_NULL) {
821 smpi_empty_status(&status[c]);
822 status[c].MPI_SOURCE = MPI_PROC_NULL;
826 for(c = 0; c < count; c++) {
827 if (MC_is_active()) {
828 smpi_mpi_wait(&requests[c], pstat);
831 index = smpi_mpi_waitany(count, requests, pstat);
832 if (index == MPI_UNDEFINED)
834 requests[index]=MPI_REQUEST_NULL;
836 if (status != MPI_STATUSES_IGNORE) {
837 status[index] = *pstat;
838 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
839 retvalue = MPI_ERR_IN_STATUS;
846 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
851 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
854 for(i = 0; i < incount; i++)
856 index=smpi_mpi_waitany(incount, requests, pstat);
857 if(index!=MPI_UNDEFINED){
858 indices[count] = index;
860 if(status != MPI_STATUSES_IGNORE) {
861 status[index] = *pstat;
863 requests[index]=MPI_REQUEST_NULL;
865 return MPI_UNDEFINED;
871 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
874 int i, count, count_dead;
876 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
880 for(i = 0; i < incount; i++) {
881 if((requests[i] != MPI_REQUEST_NULL)) {
882 if(smpi_mpi_test(&requests[i], pstat)) {
885 if(status != MPI_STATUSES_IGNORE) {
888 requests[i]=MPI_REQUEST_NULL;
895 if(count_dead==incount)return MPI_UNDEFINED;
899 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
902 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
903 nary_tree_bcast(buf, count, datatype, root, comm, 4);
906 void smpi_mpi_barrier(MPI_Comm comm)
908 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
909 nary_tree_barrier(comm, 4);
912 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
913 void *recvbuf, int recvcount, MPI_Datatype recvtype,
914 int root, MPI_Comm comm)
916 int system_tag = COLL_TAG_GATHER;
917 int rank, size, src, index;
918 MPI_Aint lb = 0, recvext = 0;
919 MPI_Request *requests;
921 rank = smpi_comm_rank(comm);
922 size = smpi_comm_size(comm);
924 // Send buffer to root
925 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
927 // FIXME: check for errors
928 smpi_datatype_extent(recvtype, &lb, &recvext);
929 // Local copy from root
930 smpi_datatype_copy(sendbuf, sendcount, sendtype,
931 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
932 // Receive buffers from senders
933 requests = xbt_new(MPI_Request, size - 1);
935 for(src = 0; src < size; src++) {
937 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
939 src, system_tag, comm);
943 // Wait for completion of irecv's.
944 smpi_mpi_startall(size - 1, requests);
945 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
951 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
952 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
956 int rank = smpi_process_index();
959 /* arbitrarily choose root as rank 0 */
960 size = smpi_comm_size(comm);
962 displs = xbt_new(int, size);
963 for (i = 0; i < size; i++) {
965 count += recvcounts[i];
967 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
968 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
969 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
970 recvcounts[rank], datatype, 0, comm);
975 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
976 void *recvbuf, int *recvcounts, int *displs,
977 MPI_Datatype recvtype, int root, MPI_Comm comm)
979 int system_tag = COLL_TAG_GATHERV;
980 int rank, size, src, index;
981 MPI_Aint lb = 0, recvext = 0;
982 MPI_Request *requests;
984 rank = smpi_comm_rank(comm);
985 size = smpi_comm_size(comm);
987 // Send buffer to root
988 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
990 // FIXME: check for errors
991 smpi_datatype_extent(recvtype, &lb, &recvext);
992 // Local copy from root
993 smpi_datatype_copy(sendbuf, sendcount, sendtype,
994 (char *)recvbuf + displs[root] * recvext,
995 recvcounts[root], recvtype);
996 // Receive buffers from senders
997 requests = xbt_new(MPI_Request, size - 1);
999 for(src = 0; src < size; src++) {
1002 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1003 recvcounts[src], recvtype, src, system_tag, comm);
1007 // Wait for completion of irecv's.
1008 smpi_mpi_startall(size - 1, requests);
1009 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1014 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1015 MPI_Datatype sendtype, void *recvbuf,
1016 int recvcount, MPI_Datatype recvtype,
1019 int system_tag = COLL_TAG_ALLGATHER;
1020 int rank, size, other, index;
1021 MPI_Aint lb = 0, recvext = 0;
1022 MPI_Request *requests;
1024 rank = smpi_comm_rank(comm);
1025 size = smpi_comm_size(comm);
1026 // FIXME: check for errors
1027 smpi_datatype_extent(recvtype, &lb, &recvext);
1028 // Local copy from self
1029 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1030 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1032 // Send/Recv buffers to/from others;
1033 requests = xbt_new(MPI_Request, 2 * (size - 1));
1035 for(other = 0; other < size; other++) {
1038 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1041 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1042 recvcount, recvtype, other,
1047 // Wait for completion of all comms.
1048 smpi_mpi_startall(2 * (size - 1), requests);
1049 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1053 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1054 MPI_Datatype sendtype, void *recvbuf,
1055 int *recvcounts, int *displs,
1056 MPI_Datatype recvtype, MPI_Comm comm)
1058 int system_tag = COLL_TAG_ALLGATHERV;
1059 int rank, size, other, index;
1060 MPI_Aint lb = 0, recvext = 0;
1061 MPI_Request *requests;
1063 rank = smpi_comm_rank(comm);
1064 size = smpi_comm_size(comm);
1065 // FIXME: check for errors
1066 smpi_datatype_extent(recvtype, &lb, &recvext);
1067 // Local copy from self
1068 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1069 (char *)recvbuf + displs[rank] * recvext,
1070 recvcounts[rank], recvtype);
1071 // Send buffers to others;
1072 requests = xbt_new(MPI_Request, 2 * (size - 1));
1074 for(other = 0; other < size; other++) {
1077 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1081 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1082 recvtype, other, system_tag, comm);
1086 // Wait for completion of all comms.
1087 smpi_mpi_startall(2 * (size - 1), requests);
1088 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1092 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1093 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1094 int root, MPI_Comm comm)
1096 int system_tag = COLL_TAG_SCATTER;
1097 int rank, size, dst, index;
1098 MPI_Aint lb = 0, sendext = 0;
1099 MPI_Request *requests;
1101 rank = smpi_comm_rank(comm);
1102 size = smpi_comm_size(comm);
1104 // Recv buffer from root
1105 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1108 // FIXME: check for errors
1109 smpi_datatype_extent(sendtype, &lb, &sendext);
1110 // Local copy from root
1111 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1112 sendcount, sendtype, recvbuf, recvcount, recvtype);
1113 // Send buffers to receivers
1114 requests = xbt_new(MPI_Request, size - 1);
1116 for(dst = 0; dst < size; dst++) {
1118 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1119 sendcount, sendtype, dst,
1124 // Wait for completion of isend's.
1125 smpi_mpi_startall(size - 1, requests);
1126 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1131 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1132 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1133 MPI_Datatype recvtype, int root, MPI_Comm comm)
1135 int system_tag = COLL_TAG_SCATTERV;
1136 int rank, size, dst, index;
1137 MPI_Aint lb = 0, sendext = 0;
1138 MPI_Request *requests;
1140 rank = smpi_comm_rank(comm);
1141 size = smpi_comm_size(comm);
1143 // Recv buffer from root
1144 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1147 // FIXME: check for errors
1148 smpi_datatype_extent(sendtype, &lb, &sendext);
1149 // Local copy from root
1150 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1151 sendtype, recvbuf, recvcount, recvtype);
1152 // Send buffers to receivers
1153 requests = xbt_new(MPI_Request, size - 1);
1155 for(dst = 0; dst < size; dst++) {
1158 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1159 sendtype, dst, system_tag, comm);
1163 // Wait for completion of isend's.
1164 smpi_mpi_startall(size - 1, requests);
1165 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1170 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1171 MPI_Datatype datatype, MPI_Op op, int root,
1174 int system_tag = COLL_TAG_REDUCE;
1175 int rank, size, src, index;
1176 MPI_Aint lb = 0, dataext = 0;
1177 MPI_Request *requests;
1180 char* sendtmpbuf = (char*) sendbuf;
1181 if( sendbuf == MPI_IN_PLACE ) {
1182 sendtmpbuf = (char *)recvbuf;
1185 rank = smpi_comm_rank(comm);
1186 size = smpi_comm_size(comm);
1187 //non commutative case, use a working algo from openmpi
1188 if(!smpi_op_is_commute(op)){
1189 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1190 datatype, op, root, comm);
1195 // Send buffer to root
1196 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1198 // FIXME: check for errors
1199 smpi_datatype_extent(datatype, &lb, &dataext);
1200 // Local copy from root
1201 if (sendtmpbuf && recvbuf)
1202 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1203 // Receive buffers from senders
1204 //TODO: make a MPI_barrier here ?
1205 requests = xbt_new(MPI_Request, size - 1);
1206 tmpbufs = xbt_new(void *, size - 1);
1208 for(src = 0; src < size; src++) {
1210 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1212 tmpbufs[index] = xbt_malloc(count * dataext);
1214 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1219 // Wait for completion of irecv's.
1220 smpi_mpi_startall(size - 1, requests);
1221 for(src = 0; src < size - 1; src++) {
1222 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1223 XBT_DEBUG("finished waiting any request with index %d", index);
1224 if(index == MPI_UNDEFINED) {
1227 if(op) /* op can be MPI_OP_NULL that does nothing */
1228 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1230 for(index = 0; index < size - 1; index++) {
1231 xbt_free(tmpbufs[index]);
1238 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1239 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1241 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1242 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1245 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1246 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1248 int system_tag = 888;
1249 int rank, size, other, index;
1250 MPI_Aint lb = 0, dataext = 0;
1251 MPI_Request *requests;
1254 rank = smpi_comm_rank(comm);
1255 size = smpi_comm_size(comm);
1257 // FIXME: check for errors
1258 smpi_datatype_extent(datatype, &lb, &dataext);
1260 // Local copy from self
1261 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1263 // Send/Recv buffers to/from others;
1264 requests = xbt_new(MPI_Request, size - 1);
1265 tmpbufs = xbt_new(void *, rank);
1267 for(other = 0; other < rank; other++) {
1268 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1270 tmpbufs[index] = xbt_malloc(count * dataext);
1272 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1276 for(other = rank + 1; other < size; other++) {
1278 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1281 // Wait for completion of all comms.
1282 smpi_mpi_startall(size - 1, requests);
1283 for(other = 0; other < size - 1; other++) {
1284 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1285 if(index == MPI_UNDEFINED) {
1289 // #Request is below rank: it's a irecv
1290 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1293 for(index = 0; index < rank; index++) {
1294 xbt_free(tmpbufs[index]);