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)
124 parse_factor(sg_cfg_get_string("smpi/os"));
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)
146 parse_factor(sg_cfg_get_string("smpi/ois"));
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)
168 parse_factor(sg_cfg_get_string("smpi/or"));
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 = 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) {
238 if(status != MPI_STATUS_IGNORE) {
239 status->MPI_SOURCE=MPI_ANY_SOURCE;
240 status->MPI_TAG=MPI_ANY_TAG;
245 void smpi_action_trace_run(char *path)
249 xbt_dict_cursor_t cursor;
253 action_fp = fopen(path, "r");
254 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
258 if (!xbt_dict_is_empty(action_queues)) {
260 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
263 xbt_dict_foreach(action_queues, cursor, name, todo) {
264 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
270 xbt_dict_free(&action_queues);
271 action_queues = xbt_dict_new_homogeneous(NULL);
274 static void smpi_mpi_request_free_voidp(void* request)
276 MPI_Request req = request;
277 smpi_mpi_request_free(&req);
280 /* MPI Low level calls */
281 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
282 int dst, int tag, MPI_Comm comm)
284 MPI_Request request =
285 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
286 comm, PERSISTENT | SEND);
291 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
292 int dst, int tag, MPI_Comm comm)
294 MPI_Request request =
295 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
296 comm, PERSISTENT | SSEND | SEND);
301 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
302 int src, int tag, MPI_Comm comm)
304 MPI_Request request =
305 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
306 comm, PERSISTENT | RECV);
311 void smpi_mpi_start(MPI_Request request)
315 xbt_assert(!request->action,
316 "Cannot (re)start a non-finished communication");
317 if(request->flags & RECV) {
318 print_request("New recv", request);
319 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
320 mailbox = smpi_process_mailbox_small();
322 mailbox = smpi_process_mailbox();
323 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
324 request->real_size=request->size;
325 smpi_datatype_use(request->old_type);
326 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
328 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
329 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
331 simcall_process_sleep(sleeptime);
332 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
337 int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
338 /* if(receiver == MPI_UNDEFINED) {*/
339 /* XBT_WARN("Trying to send a message to a wrong rank");*/
342 print_request("New send", request);
343 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
344 mailbox = smpi_process_remote_mailbox_small(receiver);
346 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
347 mailbox = smpi_process_remote_mailbox(receiver);
349 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
351 request->detached = 1;
353 if(request->old_type->has_subtype == 0){
354 oldbuf = request->buf;
356 request->buf = xbt_malloc(request->size);
357 memcpy(request->buf,oldbuf,request->size);
360 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
362 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
363 request->real_size=request->size;
364 smpi_datatype_use(request->old_type);
366 //if we are giving back the control to the user without waiting for completion, we have to inject timings
367 double sleeptime =0.0;
368 if(request->detached){
369 //isend and send timings may be different
370 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
374 simcall_process_sleep(sleeptime);
375 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
379 simcall_comm_isend(mailbox, request->size, -1.0,
380 request->buf, request->real_size,
382 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
384 // detach if msg size < eager/rdv switch limit
388 /* FIXME: detached sends are not traceable (request->action == NULL) */
390 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
397 void smpi_mpi_startall(int count, MPI_Request * requests)
401 for(i = 0; i < count; i++) {
402 smpi_mpi_start(requests[i]);
406 void smpi_mpi_request_free(MPI_Request * request)
409 if((*request) != MPI_REQUEST_NULL){
410 (*request)->refcount--;
411 if((*request)->refcount<0) xbt_die("wrong refcount");
413 if((*request)->refcount==0){
415 *request = MPI_REQUEST_NULL;
418 xbt_die("freeing an already free request");
422 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
423 int dst, int tag, MPI_Comm comm)
425 MPI_Request request =
426 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
427 comm, NON_PERSISTENT | SEND);
432 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
433 int dst, int tag, MPI_Comm comm)
435 MPI_Request request =
436 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
437 comm, NON_PERSISTENT | ISEND | SEND);
439 smpi_mpi_start(request);
443 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
444 int dst, int tag, MPI_Comm comm)
446 MPI_Request request =
447 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
448 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
449 smpi_mpi_start(request);
455 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
456 int src, int tag, MPI_Comm comm)
458 MPI_Request request =
459 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
460 comm, NON_PERSISTENT | RECV);
464 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
465 int src, int tag, MPI_Comm comm)
467 MPI_Request request =
468 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
469 comm, NON_PERSISTENT | RECV);
471 smpi_mpi_start(request);
475 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
476 int tag, MPI_Comm comm, MPI_Status * status)
479 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
480 smpi_mpi_wait(&request, status);
485 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
486 int tag, MPI_Comm comm)
488 MPI_Request request =
489 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
490 comm, NON_PERSISTENT | SEND);
492 smpi_mpi_start(request);
493 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
497 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
498 int dst, int tag, MPI_Comm comm)
500 MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
501 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
504 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
505 int dst, int sendtag, void *recvbuf, int recvcount,
506 MPI_Datatype recvtype, int src, int recvtag,
507 MPI_Comm comm, MPI_Status * status)
509 MPI_Request requests[2];
513 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
515 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
516 smpi_mpi_startall(2, requests);
517 smpi_mpi_waitall(2, requests, stats);
518 if(status != MPI_STATUS_IGNORE) {
519 // Copy receive status
520 memcpy(status, &stats[1], sizeof(MPI_Status));
524 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
526 return status->count / smpi_datatype_size(datatype);
529 static void finish_wait(MPI_Request * request, MPI_Status * status)
531 MPI_Request req = *request;
532 if(!(req->detached && req->flags & SEND)){
533 if(status != MPI_STATUS_IGNORE) {
534 status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
535 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
537 status->MPI_ERROR = MPI_ERR_TRUNCATE;
538 else status->MPI_ERROR = MPI_SUCCESS ;
539 // this handles the case were size in receive differs from size in send
540 // FIXME: really this should just contain the count of receive-type blocks,
542 status->count = req->real_size;
545 print_request("Finishing", req);
546 MPI_Datatype datatype = req->old_type;
548 if(datatype->has_subtype == 1){
549 // This part handles the problem of non-contignous memory
550 // the unserialization at the reception
551 s_smpi_subtype_t *subtype = datatype->substruct;
552 if(req->flags & RECV) {
553 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
555 if(req->detached == 0) free(req->buf);
557 smpi_datatype_unuse(datatype);
560 if(req->detached_sender!=NULL){
561 smpi_mpi_request_free(&(req->detached_sender));
564 if(req->flags & NON_PERSISTENT) {
565 smpi_mpi_request_free(request);
571 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
574 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
575 if ((*request)->action == NULL)
578 flag = simcall_comm_test((*request)->action);
580 (*request)->refcount++;
581 finish_wait(request, status);
583 smpi_empty_status(status);
588 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
595 *index = MPI_UNDEFINED;
598 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
599 map = xbt_new(int, count);
601 for(i = 0; i < count; i++) {
602 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
603 xbt_dynar_push(comms, &requests[i]->action);
609 i = simcall_comm_testany(comms);
610 // not MPI_UNDEFINED, as this is a simix return code
613 finish_wait(&requests[*index], status);
617 //all requests are null or inactive, return true
619 smpi_empty_status(status);
622 xbt_dynar_free(&comms);
629 int smpi_mpi_testall(int count, MPI_Request requests[],
633 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
636 for(i=0; i<count; i++){
637 if(requests[i]!= MPI_REQUEST_NULL){
638 if (smpi_mpi_test(&requests[i], pstat)!=1){
642 smpi_empty_status(pstat);
644 if(status != MPI_STATUSES_IGNORE) {
645 memcpy(&status[i], pstat, sizeof(*pstat));
651 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
653 //FIXME find another wait to avoid busy waiting ?
654 // the issue here is that we have to wait on a nonexistent comm
656 smpi_mpi_iprobe(source, tag, comm, &flag, status);
657 XBT_DEBUG("Busy Waiting on probing : %d", flag);
659 simcall_process_sleep(0.0001);
664 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
665 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
666 comm, NON_PERSISTENT | RECV);
668 // behave like a receive, but don't do it
671 print_request("New iprobe", request);
672 // We have to test both mailboxes as we don't know if we will receive one one or another
673 if (sg_cfg_get_int("smpi/async_small_thres")>0){
674 mailbox = smpi_process_mailbox_small();
675 XBT_DEBUG("trying to probe the perm recv mailbox");
676 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
678 if (request->action==NULL){
679 mailbox = smpi_process_mailbox();
680 XBT_DEBUG("trying to probe the other mailbox");
681 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
685 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
687 if(status != MPI_STATUS_IGNORE) {
688 status->MPI_SOURCE = req->src;
689 status->MPI_TAG = req->tag;
690 status->MPI_ERROR = MPI_SUCCESS;
691 status->count = req->real_size;
695 smpi_mpi_request_free(&request);
700 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
702 print_request("Waiting", *request);
703 if ((*request)->action != NULL) { // this is not a detached send
704 simcall_comm_wait((*request)->action, -1.0);
706 finish_wait(request, status);
708 // FIXME for a detached send, finish_wait is not called:
711 int smpi_mpi_waitany(int count, MPI_Request requests[],
718 index = MPI_UNDEFINED;
720 // Wait for a request to complete
721 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
722 map = xbt_new(int, count);
724 XBT_DEBUG("Wait for one of %d", count);
725 for(i = 0; i < count; i++) {
726 if(requests[i] != MPI_REQUEST_NULL) {
727 if (requests[i]->action != NULL) {
728 XBT_DEBUG("Waiting any %p ", requests[i]);
729 xbt_dynar_push(comms, &requests[i]->action);
733 //This is a finished detached request, let's return this one
734 size=0;//so we free the dynar but don't do the waitany call
736 finish_wait(&requests[i], status);//cleanup if refcount = 0
737 requests[i]=MPI_REQUEST_NULL;//set to null
743 i = simcall_comm_waitany(comms);
745 // not MPI_UNDEFINED, as this is a simix return code
748 finish_wait(&requests[index], status);
752 xbt_dynar_free(&comms);
755 if (index==MPI_UNDEFINED)
756 smpi_empty_status(status);
761 int smpi_mpi_waitall(int count, MPI_Request requests[],
766 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
767 int retvalue=MPI_SUCCESS;
768 //tag invalid requests in the set
769 for(c = 0; c < count; c++) {
770 if(requests[c]==MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ){
771 if(status != MPI_STATUSES_IGNORE)
772 smpi_empty_status(&status[c]);
773 }else if(requests[c]->src == MPI_PROC_NULL ){
774 if(status != MPI_STATUSES_IGNORE) {
775 smpi_empty_status(&status[c]);
776 status[c].MPI_SOURCE=MPI_PROC_NULL;
780 for(c = 0; c < count; c++) {
782 smpi_mpi_wait(&requests[c], pstat);
785 index = smpi_mpi_waitany(count, requests, pstat);
786 if(index == MPI_UNDEFINED) {
789 if(status != MPI_STATUSES_IGNORE) {
790 memcpy(&status[index], pstat, sizeof(*pstat));
791 if(status[index].MPI_ERROR==MPI_ERR_TRUNCATE)retvalue=MPI_ERR_IN_STATUS;
800 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
805 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
808 for(i = 0; i < incount; i++)
810 index=smpi_mpi_waitany(incount, requests, pstat);
811 if(index!=MPI_UNDEFINED){
812 indices[count] = index;
814 if(status != MPI_STATUSES_IGNORE) {
815 memcpy(&status[index], pstat, sizeof(*pstat));
818 return MPI_UNDEFINED;
824 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
827 int i, count, count_dead;
829 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
833 for(i = 0; i < incount; i++) {
834 if((requests[i] != MPI_REQUEST_NULL)) {
835 if(smpi_mpi_test(&requests[i], pstat)) {
838 if(status != MPI_STATUSES_IGNORE) {
839 memcpy(&status[i], pstat, sizeof(*pstat));
846 if(count_dead==incount)return MPI_UNDEFINED;
850 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
853 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
854 nary_tree_bcast(buf, count, datatype, root, comm, 4);
857 void smpi_mpi_barrier(MPI_Comm comm)
859 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
860 nary_tree_barrier(comm, 4);
863 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
864 void *recvbuf, int recvcount, MPI_Datatype recvtype,
865 int root, MPI_Comm comm)
867 int system_tag = 666;
868 int rank, size, src, index;
869 MPI_Aint lb = 0, recvext = 0;
870 MPI_Request *requests;
872 rank = smpi_comm_rank(comm);
873 size = smpi_comm_size(comm);
875 // Send buffer to root
876 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
878 // FIXME: check for errors
879 smpi_datatype_extent(recvtype, &lb, &recvext);
880 // Local copy from root
881 smpi_datatype_copy(sendbuf, sendcount, sendtype,
882 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
883 // Receive buffers from senders
884 requests = xbt_new(MPI_Request, size - 1);
886 for(src = 0; src < size; src++) {
888 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
890 src, system_tag, comm);
894 // Wait for completion of irecv's.
895 smpi_mpi_startall(size - 1, requests);
896 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
901 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
902 void *recvbuf, int *recvcounts, int *displs,
903 MPI_Datatype recvtype, int root, MPI_Comm comm)
905 int system_tag = 666;
906 int rank, size, src, index;
907 MPI_Aint lb = 0, recvext = 0;
908 MPI_Request *requests;
910 rank = smpi_comm_rank(comm);
911 size = smpi_comm_size(comm);
913 // Send buffer to root
914 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
916 // FIXME: check for errors
917 smpi_datatype_extent(recvtype, &lb, &recvext);
918 // Local copy from root
919 smpi_datatype_copy(sendbuf, sendcount, sendtype,
920 (char *)recvbuf + displs[root] * recvext,
921 recvcounts[root], recvtype);
922 // Receive buffers from senders
923 requests = xbt_new(MPI_Request, size - 1);
925 for(src = 0; src < size; src++) {
928 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
929 recvcounts[src], recvtype, src, system_tag, comm);
933 // Wait for completion of irecv's.
934 smpi_mpi_startall(size - 1, requests);
935 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
940 void smpi_mpi_allgather(void *sendbuf, int sendcount,
941 MPI_Datatype sendtype, void *recvbuf,
942 int recvcount, MPI_Datatype recvtype,
945 int system_tag = 666;
946 int rank, size, other, index;
947 MPI_Aint lb = 0, recvext = 0;
948 MPI_Request *requests;
950 rank = smpi_comm_rank(comm);
951 size = smpi_comm_size(comm);
952 // FIXME: check for errors
953 smpi_datatype_extent(recvtype, &lb, &recvext);
954 // Local copy from self
955 smpi_datatype_copy(sendbuf, sendcount, sendtype,
956 (char *)recvbuf + rank * recvcount * recvext, recvcount,
958 // Send/Recv buffers to/from others;
959 requests = xbt_new(MPI_Request, 2 * (size - 1));
961 for(other = 0; other < size; other++) {
964 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
967 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
968 recvcount, recvtype, other,
973 // Wait for completion of all comms.
974 smpi_mpi_startall(2 * (size - 1), requests);
975 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
979 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
980 MPI_Datatype sendtype, void *recvbuf,
981 int *recvcounts, int *displs,
982 MPI_Datatype recvtype, MPI_Comm comm)
984 int system_tag = 666;
985 int rank, size, other, index;
986 MPI_Aint lb = 0, recvext = 0;
987 MPI_Request *requests;
989 rank = smpi_comm_rank(comm);
990 size = smpi_comm_size(comm);
991 // FIXME: check for errors
992 smpi_datatype_extent(recvtype, &lb, &recvext);
993 // Local copy from self
994 smpi_datatype_copy(sendbuf, sendcount, sendtype,
995 (char *)recvbuf + displs[rank] * recvext,
996 recvcounts[rank], recvtype);
997 // Send buffers to others;
998 requests = xbt_new(MPI_Request, 2 * (size - 1));
1000 for(other = 0; other < size; other++) {
1003 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1007 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1008 recvtype, other, system_tag, comm);
1012 // Wait for completion of all comms.
1013 smpi_mpi_startall(2 * (size - 1), requests);
1014 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1018 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1019 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1020 int root, MPI_Comm comm)
1022 int system_tag = 666;
1023 int rank, size, dst, index;
1024 MPI_Aint lb = 0, sendext = 0;
1025 MPI_Request *requests;
1027 rank = smpi_comm_rank(comm);
1028 size = smpi_comm_size(comm);
1030 // Recv buffer from root
1031 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1034 // FIXME: check for errors
1035 smpi_datatype_extent(sendtype, &lb, &sendext);
1036 // Local copy from root
1037 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1038 sendcount, sendtype, recvbuf, recvcount, recvtype);
1039 // Send buffers to receivers
1040 requests = xbt_new(MPI_Request, size - 1);
1042 for(dst = 0; dst < size; dst++) {
1044 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1045 sendcount, sendtype, dst,
1050 // Wait for completion of isend's.
1051 smpi_mpi_startall(size - 1, requests);
1052 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1057 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1058 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1059 MPI_Datatype recvtype, int root, MPI_Comm comm)
1061 int system_tag = 666;
1062 int rank, size, dst, index;
1063 MPI_Aint lb = 0, sendext = 0;
1064 MPI_Request *requests;
1066 rank = smpi_comm_rank(comm);
1067 size = smpi_comm_size(comm);
1069 // Recv buffer from root
1070 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1073 // FIXME: check for errors
1074 smpi_datatype_extent(sendtype, &lb, &sendext);
1075 // Local copy from root
1076 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1077 sendtype, recvbuf, recvcount, recvtype);
1078 // Send buffers to receivers
1079 requests = xbt_new(MPI_Request, size - 1);
1081 for(dst = 0; dst < size; dst++) {
1084 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1085 sendtype, dst, system_tag, comm);
1089 // Wait for completion of isend's.
1090 smpi_mpi_startall(size - 1, requests);
1091 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1096 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1097 MPI_Datatype datatype, MPI_Op op, int root,
1100 int system_tag = 666;
1101 int rank, size, src, index;
1102 MPI_Aint lb = 0, dataext = 0;
1103 MPI_Request *requests;
1106 rank = smpi_comm_rank(comm);
1107 size = smpi_comm_size(comm);
1109 // Send buffer to root
1110 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
1112 // FIXME: check for errors
1113 smpi_datatype_extent(datatype, &lb, &dataext);
1114 // Local copy from root
1115 if (sendbuf && recvbuf)
1116 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1117 // Receive buffers from senders
1118 //TODO: make a MPI_barrier here ?
1119 requests = xbt_new(MPI_Request, size - 1);
1120 tmpbufs = xbt_new(void *, size - 1);
1122 for(src = 0; src < size; src++) {
1124 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1126 tmpbufs[index] = xbt_malloc(count * dataext);
1128 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1133 // Wait for completion of irecv's.
1134 smpi_mpi_startall(size - 1, requests);
1135 for(src = 0; src < size - 1; src++) {
1136 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1137 XBT_DEBUG("finished waiting any request with index %d", index);
1138 if(index == MPI_UNDEFINED) {
1141 if(op) /* op can be MPI_OP_NULL that does nothing */
1142 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1144 for(index = 0; index < size - 1; index++) {
1145 xbt_free(tmpbufs[index]);
1152 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1153 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1155 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1156 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1159 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1160 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1162 int system_tag = 666;
1163 int rank, size, other, index;
1164 MPI_Aint lb = 0, dataext = 0;
1165 MPI_Request *requests;
1168 rank = smpi_comm_rank(comm);
1169 size = smpi_comm_size(comm);
1171 // FIXME: check for errors
1172 smpi_datatype_extent(datatype, &lb, &dataext);
1174 // Local copy from self
1175 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1177 // Send/Recv buffers to/from others;
1178 requests = xbt_new(MPI_Request, size - 1);
1179 tmpbufs = xbt_new(void *, rank);
1181 for(other = 0; other < rank; other++) {
1182 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1184 tmpbufs[index] = xbt_malloc(count * dataext);
1186 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1190 for(other = rank + 1; other < size; other++) {
1192 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1195 // Wait for completion of all comms.
1196 smpi_mpi_startall(size - 1, requests);
1197 for(other = 0; other < size - 1; other++) {
1198 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1199 if(index == MPI_UNDEFINED) {
1203 // #Request is below rank: it's a irecv
1204 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1207 for(index = 0; index < rank; index++) {
1208 xbt_free(tmpbufs[index]);