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 = 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_comm_rank(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_comm_rank(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, smpi_comm_rank(comm), 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) {
321 int rank = smpi_process_index();
322 if (TRACE_smpi_view_internals()) {
323 TRACE_smpi_computing_out(rank);
324 TRACE_smpi_ptp_in(rank, 0, rank, __FUNCTION__);
327 print_request("New recv", request);
328 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
329 mailbox = smpi_process_mailbox_small();
331 mailbox = smpi_process_mailbox();
332 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
333 request->real_size=request->size;
334 smpi_datatype_use(request->old_type);
335 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
337 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
338 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
340 simcall_process_sleep(sleeptime);
341 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
350 int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
353 int rank = smpi_process_index();
354 if (TRACE_smpi_view_internals()) {
355 TRACE_smpi_ptp_in(rank, rank, receiver, __FUNCTION__);
356 TRACE_smpi_send(rank, rank, receiver);
359 /* if(receiver == MPI_UNDEFINED) {*/
360 /* XBT_WARN("Trying to send a message to a wrong rank");*/
363 print_request("New send", request);
364 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
365 mailbox = smpi_process_remote_mailbox_small(receiver);
367 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
368 mailbox = smpi_process_remote_mailbox(receiver);
370 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
372 request->detached = 1;
374 if(request->old_type->has_subtype == 0){
375 oldbuf = request->buf;
377 request->buf = xbt_malloc(request->size);
378 memcpy(request->buf,oldbuf,request->size);
381 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
383 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
384 request->real_size=request->size;
385 smpi_datatype_use(request->old_type);
387 //if we are giving back the control to the user without waiting for completion, we have to inject timings
388 double sleeptime =0.0;
389 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
390 //isend and send timings may be different
391 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
395 simcall_process_sleep(sleeptime);
396 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
400 simcall_comm_isend(mailbox, request->size, -1.0,
401 request->buf, request->real_size,
403 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
405 // detach if msg size < eager/rdv switch limit
409 /* FIXME: detached sends are not traceable (request->action == NULL) */
411 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
419 void smpi_mpi_startall(int count, MPI_Request * requests)
423 for(i = 0; i < count; i++) {
424 smpi_mpi_start(requests[i]);
428 void smpi_mpi_request_free(MPI_Request * request)
431 if((*request) != MPI_REQUEST_NULL){
432 (*request)->refcount--;
433 if((*request)->refcount<0) xbt_die("wrong refcount");
435 if((*request)->refcount==0){
437 *request = MPI_REQUEST_NULL;
440 xbt_die("freeing an already free request");
444 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
445 int dst, int tag, MPI_Comm comm)
447 MPI_Request request =
448 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
449 comm, NON_PERSISTENT | SEND);
454 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
455 int dst, int tag, MPI_Comm comm)
457 MPI_Request request =
458 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
459 comm, NON_PERSISTENT | ISEND | SEND);
461 smpi_mpi_start(request);
465 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
466 int dst, int tag, MPI_Comm comm)
468 MPI_Request request =
469 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
470 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
471 smpi_mpi_start(request);
477 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
478 int src, int tag, MPI_Comm comm)
480 MPI_Request request =
481 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
482 comm, NON_PERSISTENT | RECV);
486 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
487 int src, int tag, MPI_Comm comm)
489 MPI_Request request =
490 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
491 comm, NON_PERSISTENT | RECV);
493 smpi_mpi_start(request);
497 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
498 int tag, MPI_Comm comm, MPI_Status * status)
501 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
502 smpi_mpi_wait(&request, status);
507 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
508 int tag, MPI_Comm comm)
510 MPI_Request request =
511 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
512 comm, NON_PERSISTENT | SEND);
514 smpi_mpi_start(request);
515 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
519 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
520 int dst, int tag, MPI_Comm comm)
522 MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
523 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
526 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
527 int dst, int sendtag, void *recvbuf, int recvcount,
528 MPI_Datatype recvtype, int src, int recvtag,
529 MPI_Comm comm, MPI_Status * status)
531 MPI_Request requests[2];
535 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
537 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
538 smpi_mpi_startall(2, requests);
539 smpi_mpi_waitall(2, requests, stats);
540 if(status != MPI_STATUS_IGNORE) {
541 // Copy receive status
546 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
548 return status->count / smpi_datatype_size(datatype);
551 static void finish_wait(MPI_Request * request, MPI_Status * status)
553 MPI_Request req = *request;
554 if(status != MPI_STATUS_IGNORE)
555 smpi_empty_status(status);
557 if(!(req->detached && req->flags & SEND)){
558 if(status != MPI_STATUS_IGNORE) {
559 status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
560 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
561 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
562 // this handles the case were size in receive differs from size in send
563 // FIXME: really this should just contain the count of receive-type blocks,
565 status->count = req->real_size;
568 print_request("Finishing", req);
569 MPI_Datatype datatype = req->old_type;
571 if(datatype->has_subtype == 1){
572 // This part handles the problem of non-contignous memory
573 // the unserialization at the reception
574 s_smpi_subtype_t *subtype = datatype->substruct;
575 if(req->flags & RECV) {
576 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
578 if(req->detached == 0) free(req->buf);
580 smpi_datatype_unuse(datatype);
583 if (TRACE_smpi_view_internals()) {
584 int rank = smpi_process_index();
585 TRACE_smpi_ptp_out(rank, rank, 0, __FUNCTION__);
592 if (TRACE_smpi_view_internals()) {
593 if(req->flags & RECV)
595 int rank = smpi_process_index();
597 //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
598 TRACE_smpi_ptp_out(rank, 0, rank, __FUNCTION__);
599 int src_traced = smpi_group_index(smpi_comm_group(req->comm), req->src);
600 TRACE_smpi_recv(rank, src_traced, rank);
601 TRACE_smpi_computing_in(rank);
606 if(req->detached_sender!=NULL){
607 smpi_mpi_request_free(&(req->detached_sender));
610 if(req->flags & NON_PERSISTENT) {
611 smpi_mpi_request_free(request);
617 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
620 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
621 if ((*request)->action == NULL)
624 flag = simcall_comm_test((*request)->action);
626 (*request)->refcount++;
627 finish_wait(request, status);
629 smpi_empty_status(status);
634 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
641 *index = MPI_UNDEFINED;
644 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
645 map = xbt_new(int, count);
647 for(i = 0; i < count; i++) {
648 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
649 xbt_dynar_push(comms, &requests[i]->action);
655 i = simcall_comm_testany(comms);
656 // not MPI_UNDEFINED, as this is a simix return code
659 finish_wait(&requests[*index], status);
663 //all requests are null or inactive, return true
665 smpi_empty_status(status);
668 xbt_dynar_free(&comms);
675 int smpi_mpi_testall(int count, MPI_Request requests[],
679 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
682 for(i=0; i<count; i++){
683 if(requests[i]!= MPI_REQUEST_NULL){
684 if (smpi_mpi_test(&requests[i], pstat)!=1){
688 smpi_empty_status(pstat);
690 if(status != MPI_STATUSES_IGNORE) {
697 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
699 //FIXME find another wait to avoid busy waiting ?
700 // the issue here is that we have to wait on a nonexistent comm
702 smpi_mpi_iprobe(source, tag, comm, &flag, status);
703 XBT_DEBUG("Busy Waiting on probing : %d", flag);
705 simcall_process_sleep(0.0001);
710 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
711 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
712 comm, NON_PERSISTENT | RECV);
714 // behave like a receive, but don't do it
717 print_request("New iprobe", request);
718 // We have to test both mailboxes as we don't know if we will receive one one or another
719 if (sg_cfg_get_int("smpi/async_small_thres")>0){
720 mailbox = smpi_process_mailbox_small();
721 XBT_DEBUG("trying to probe the perm recv mailbox");
722 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
724 if (request->action==NULL){
725 mailbox = smpi_process_mailbox();
726 XBT_DEBUG("trying to probe the other mailbox");
727 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
731 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
733 if(status != MPI_STATUS_IGNORE) {
734 status->MPI_SOURCE = req->src;
735 status->MPI_TAG = req->tag;
736 status->MPI_ERROR = MPI_SUCCESS;
737 status->count = req->real_size;
741 smpi_mpi_request_free(&request);
746 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
748 print_request("Waiting", *request);
749 if ((*request)->action != NULL) { // this is not a detached send
750 simcall_comm_wait((*request)->action, -1.0);
752 finish_wait(request, status);
754 // FIXME for a detached send, finish_wait is not called:
757 int smpi_mpi_waitany(int count, MPI_Request requests[],
764 index = MPI_UNDEFINED;
766 // Wait for a request to complete
767 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
768 map = xbt_new(int, count);
770 XBT_DEBUG("Wait for one of %d", count);
771 for(i = 0; i < count; i++) {
772 if(requests[i] != MPI_REQUEST_NULL) {
773 if (requests[i]->action != NULL) {
774 XBT_DEBUG("Waiting any %p ", requests[i]);
775 xbt_dynar_push(comms, &requests[i]->action);
779 //This is a finished detached request, let's return this one
780 size=0;//so we free the dynar but don't do the waitany call
782 finish_wait(&requests[i], status);//cleanup if refcount = 0
783 requests[i]=MPI_REQUEST_NULL;//set to null
789 i = simcall_comm_waitany(comms);
791 // not MPI_UNDEFINED, as this is a simix return code
794 finish_wait(&requests[index], status);
798 xbt_dynar_free(&comms);
801 if (index==MPI_UNDEFINED)
802 smpi_empty_status(status);
807 int smpi_mpi_waitall(int count, MPI_Request requests[],
812 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
813 int retvalue = MPI_SUCCESS;
814 //tag invalid requests in the set
815 if (status != MPI_STATUSES_IGNORE) {
816 for (c = 0; c < count; c++) {
817 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
818 smpi_empty_status(&status[c]);
819 } else if (requests[c]->src == MPI_PROC_NULL) {
820 smpi_empty_status(&status[c]);
821 status[c].MPI_SOURCE = MPI_PROC_NULL;
825 for(c = 0; c < count; c++) {
826 if (MC_is_active()) {
827 smpi_mpi_wait(&requests[c], pstat);
830 index = smpi_mpi_waitany(count, requests, pstat);
831 if (index == MPI_UNDEFINED)
834 if (status != MPI_STATUSES_IGNORE) {
835 status[index] = *pstat;
836 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
837 retvalue = MPI_ERR_IN_STATUS;
844 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
849 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
852 for(i = 0; i < incount; i++)
854 index=smpi_mpi_waitany(incount, requests, pstat);
855 if(index!=MPI_UNDEFINED){
856 indices[count] = index;
858 if(status != MPI_STATUSES_IGNORE) {
859 status[index] = *pstat;
862 return MPI_UNDEFINED;
868 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
871 int i, count, count_dead;
873 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
877 for(i = 0; i < incount; i++) {
878 if((requests[i] != MPI_REQUEST_NULL)) {
879 if(smpi_mpi_test(&requests[i], pstat)) {
882 if(status != MPI_STATUSES_IGNORE) {
890 if(count_dead==incount)return MPI_UNDEFINED;
894 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
897 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
898 nary_tree_bcast(buf, count, datatype, root, comm, 4);
901 void smpi_mpi_barrier(MPI_Comm comm)
903 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
904 nary_tree_barrier(comm, 4);
907 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
908 void *recvbuf, int recvcount, MPI_Datatype recvtype,
909 int root, MPI_Comm comm)
911 int system_tag = 666;
912 int rank, size, src, index;
913 MPI_Aint lb = 0, recvext = 0;
914 MPI_Request *requests;
916 rank = smpi_comm_rank(comm);
917 size = smpi_comm_size(comm);
919 // Send buffer to root
920 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
922 // FIXME: check for errors
923 smpi_datatype_extent(recvtype, &lb, &recvext);
924 // Local copy from root
925 smpi_datatype_copy(sendbuf, sendcount, sendtype,
926 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
927 // Receive buffers from senders
928 requests = xbt_new(MPI_Request, size - 1);
930 for(src = 0; src < size; src++) {
932 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
934 src, system_tag, comm);
938 // Wait for completion of irecv's.
939 smpi_mpi_startall(size - 1, requests);
940 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
945 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
946 void *recvbuf, int *recvcounts, int *displs,
947 MPI_Datatype recvtype, int root, MPI_Comm comm)
949 int system_tag = 666;
950 int rank, size, src, index;
951 MPI_Aint lb = 0, recvext = 0;
952 MPI_Request *requests;
954 rank = smpi_comm_rank(comm);
955 size = smpi_comm_size(comm);
957 // Send buffer to root
958 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
960 // FIXME: check for errors
961 smpi_datatype_extent(recvtype, &lb, &recvext);
962 // Local copy from root
963 smpi_datatype_copy(sendbuf, sendcount, sendtype,
964 (char *)recvbuf + displs[root] * recvext,
965 recvcounts[root], recvtype);
966 // Receive buffers from senders
967 requests = xbt_new(MPI_Request, size - 1);
969 for(src = 0; src < size; src++) {
972 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
973 recvcounts[src], recvtype, src, system_tag, comm);
977 // Wait for completion of irecv's.
978 smpi_mpi_startall(size - 1, requests);
979 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
984 void smpi_mpi_allgather(void *sendbuf, int sendcount,
985 MPI_Datatype sendtype, void *recvbuf,
986 int recvcount, MPI_Datatype recvtype,
989 int system_tag = 666;
990 int rank, size, other, index;
991 MPI_Aint lb = 0, recvext = 0;
992 MPI_Request *requests;
994 rank = smpi_comm_rank(comm);
995 size = smpi_comm_size(comm);
996 // FIXME: check for errors
997 smpi_datatype_extent(recvtype, &lb, &recvext);
998 // Local copy from self
999 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1000 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1002 // Send/Recv buffers to/from others;
1003 requests = xbt_new(MPI_Request, 2 * (size - 1));
1005 for(other = 0; other < size; other++) {
1008 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1011 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1012 recvcount, recvtype, other,
1017 // Wait for completion of all comms.
1018 smpi_mpi_startall(2 * (size - 1), requests);
1019 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1023 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1024 MPI_Datatype sendtype, void *recvbuf,
1025 int *recvcounts, int *displs,
1026 MPI_Datatype recvtype, MPI_Comm comm)
1028 int system_tag = 666;
1029 int rank, size, other, index;
1030 MPI_Aint lb = 0, recvext = 0;
1031 MPI_Request *requests;
1033 rank = smpi_comm_rank(comm);
1034 size = smpi_comm_size(comm);
1035 // FIXME: check for errors
1036 smpi_datatype_extent(recvtype, &lb, &recvext);
1037 // Local copy from self
1038 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1039 (char *)recvbuf + displs[rank] * recvext,
1040 recvcounts[rank], recvtype);
1041 // Send buffers to others;
1042 requests = xbt_new(MPI_Request, 2 * (size - 1));
1044 for(other = 0; other < size; other++) {
1047 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1051 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1052 recvtype, other, system_tag, comm);
1056 // Wait for completion of all comms.
1057 smpi_mpi_startall(2 * (size - 1), requests);
1058 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1062 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1063 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1064 int root, MPI_Comm comm)
1066 int system_tag = 666;
1067 int rank, size, dst, index;
1068 MPI_Aint lb = 0, sendext = 0;
1069 MPI_Request *requests;
1071 rank = smpi_comm_rank(comm);
1072 size = smpi_comm_size(comm);
1074 // Recv buffer from root
1075 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1078 // FIXME: check for errors
1079 smpi_datatype_extent(sendtype, &lb, &sendext);
1080 // Local copy from root
1081 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1082 sendcount, sendtype, recvbuf, recvcount, recvtype);
1083 // Send buffers to receivers
1084 requests = xbt_new(MPI_Request, size - 1);
1086 for(dst = 0; dst < size; dst++) {
1088 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1089 sendcount, sendtype, dst,
1094 // Wait for completion of isend's.
1095 smpi_mpi_startall(size - 1, requests);
1096 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1101 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1102 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1103 MPI_Datatype recvtype, int root, MPI_Comm comm)
1105 int system_tag = 666;
1106 int rank, size, dst, index;
1107 MPI_Aint lb = 0, sendext = 0;
1108 MPI_Request *requests;
1110 rank = smpi_comm_rank(comm);
1111 size = smpi_comm_size(comm);
1113 // Recv buffer from root
1114 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1117 // FIXME: check for errors
1118 smpi_datatype_extent(sendtype, &lb, &sendext);
1119 // Local copy from root
1120 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1121 sendtype, recvbuf, recvcount, recvtype);
1122 // Send buffers to receivers
1123 requests = xbt_new(MPI_Request, size - 1);
1125 for(dst = 0; dst < size; dst++) {
1128 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1129 sendtype, dst, system_tag, comm);
1133 // Wait for completion of isend's.
1134 smpi_mpi_startall(size - 1, requests);
1135 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1140 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1141 MPI_Datatype datatype, MPI_Op op, int root,
1144 int system_tag = 666;
1145 int rank, size, src, index;
1146 MPI_Aint lb = 0, dataext = 0;
1147 MPI_Request *requests;
1150 rank = smpi_comm_rank(comm);
1151 size = smpi_comm_size(comm);
1153 // Send buffer to root
1154 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
1156 // FIXME: check for errors
1157 smpi_datatype_extent(datatype, &lb, &dataext);
1158 // Local copy from root
1159 if (sendbuf && recvbuf)
1160 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1161 // Receive buffers from senders
1162 //TODO: make a MPI_barrier here ?
1163 requests = xbt_new(MPI_Request, size - 1);
1164 tmpbufs = xbt_new(void *, size - 1);
1166 for(src = 0; src < size; src++) {
1168 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1170 tmpbufs[index] = xbt_malloc(count * dataext);
1172 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1177 // Wait for completion of irecv's.
1178 smpi_mpi_startall(size - 1, requests);
1179 for(src = 0; src < size - 1; src++) {
1180 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1181 XBT_DEBUG("finished waiting any request with index %d", index);
1182 if(index == MPI_UNDEFINED) {
1185 if(op) /* op can be MPI_OP_NULL that does nothing */
1186 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1188 for(index = 0; index < size - 1; index++) {
1189 xbt_free(tmpbufs[index]);
1196 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1197 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1199 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1200 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1203 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1204 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1206 int system_tag = 666;
1207 int rank, size, other, index;
1208 MPI_Aint lb = 0, dataext = 0;
1209 MPI_Request *requests;
1212 rank = smpi_comm_rank(comm);
1213 size = smpi_comm_size(comm);
1215 // FIXME: check for errors
1216 smpi_datatype_extent(datatype, &lb, &dataext);
1218 // Local copy from self
1219 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1221 // Send/Recv buffers to/from others;
1222 requests = xbt_new(MPI_Request, size - 1);
1223 tmpbufs = xbt_new(void *, rank);
1225 for(other = 0; other < rank; other++) {
1226 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1228 tmpbufs[index] = xbt_malloc(count * dataext);
1230 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1234 for(other = rank + 1; other < size; other++) {
1236 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1239 // Wait for completion of all comms.
1240 smpi_mpi_startall(size - 1, requests);
1241 for(other = 0; other < size - 1; other++) {
1242 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1243 if(index == MPI_UNDEFINED) {
1247 // #Request is below rank: it's a irecv
1248 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1251 for(index = 0; index < rank; index++) {
1252 xbt_free(tmpbufs[index]);