1 /* Copyright (c) 2007, 2008, 2009, 2010. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include "xbt/replay.h"
12 #include "simix/smx_private.h"
13 #include "surf/surf.h"
14 #include "simgrid/sg_config.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
20 static int match_recv(void* a, void* b, smx_action_t ignored) {
21 MPI_Request ref = (MPI_Request)a;
22 MPI_Request req = (MPI_Request)b;
23 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
25 xbt_assert(ref, "Cannot match recv against null reference");
26 xbt_assert(req, "Cannot match recv against null request");
27 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
28 && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag)){
29 //we match, we can transfer some values
30 // FIXME : move this to the copy function ?
31 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
32 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
33 if(ref->real_size < req->real_size) ref->truncated = 1;
35 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
41 static int match_send(void* a, void* b,smx_action_t ignored) {
42 MPI_Request ref = (MPI_Request)a;
43 MPI_Request req = (MPI_Request)b;
44 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
45 xbt_assert(ref, "Cannot match send against null reference");
46 xbt_assert(req, "Cannot match send against null request");
48 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
49 && (req->tag == MPI_ANY_TAG || req->tag == ref->tag))
51 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
52 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
53 if(req->real_size < ref->real_size) req->truncated = 1;
55 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
63 typedef struct s_smpi_factor *smpi_factor_t;
64 typedef struct s_smpi_factor {
67 double values[4];//arbitrary set to 4
69 xbt_dynar_t smpi_os_values = NULL;
70 xbt_dynar_t smpi_or_values = NULL;
71 xbt_dynar_t smpi_ois_values = NULL;
73 // Methods used to parse and store the values for timing injections in smpi
74 // These are taken from surf/network.c and generalized to have more factors
75 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
77 static int factor_cmp(const void *pa, const void *pb)
79 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor);
83 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
86 unsigned int iter = 0;
89 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
91 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
92 radical_elements = xbt_str_split(smpi_coef_string, ";");
93 xbt_dynar_foreach(radical_elements, iter, value) {
95 radical_elements2 = xbt_str_split(value, ":");
96 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
97 xbt_die("Malformed radical for smpi factor!");
98 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
100 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
102 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
107 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
108 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
109 xbt_dynar_free(&radical_elements2);
111 xbt_dynar_free(&radical_elements);
113 xbt_dynar_sort(smpi_factor, &factor_cmp);
114 xbt_dynar_foreach(smpi_factor, iter, fact) {
115 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
120 static double smpi_os(double size)
122 if (!smpi_os_values) {
123 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
124 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
126 unsigned int iter = 0;
127 s_smpi_factor_t fact;
129 xbt_dynar_foreach(smpi_os_values, iter, fact) {
130 if (size <= fact.factor) {
131 XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
134 current=fact.values[0]+fact.values[1]*size;
137 XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
142 static double smpi_ois(double size)
144 if (!smpi_ois_values) {
145 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
146 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
148 unsigned int iter = 0;
149 s_smpi_factor_t fact;
151 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
152 if (size <= fact.factor) {
153 XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
156 current=fact.values[0]+fact.values[1]*size;
159 XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
164 static double smpi_or(double size)
166 if (!smpi_or_values) {
167 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
168 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
170 unsigned int iter = 0;
171 s_smpi_factor_t fact;
173 xbt_dynar_foreach(smpi_or_values, iter, fact) {
174 if (size <= fact.factor) {
175 XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
178 current=fact.values[0]+fact.values[1]*size;
180 XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
185 static MPI_Request build_request(void *buf, int count,
186 MPI_Datatype datatype, int src, int dst,
187 int tag, MPI_Comm comm, unsigned flags)
191 void *old_buf = NULL;
193 request = xbt_new(s_smpi_mpi_request_t, 1);
195 s_smpi_subtype_t *subtype = datatype->substruct;
197 if(datatype->has_subtype == 1){
198 // This part handles the problem of non-contiguous memory
200 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
202 subtype->serialize(old_buf, buf, count, datatype->substruct);
207 // This part handles the problem of non-contiguous memory (for the
208 // unserialisation at the reception)
209 request->old_buf = old_buf;
210 request->old_type = datatype;
212 request->size = smpi_datatype_size(datatype) * count;
216 request->comm = comm;
217 request->action = NULL;
218 request->flags = flags;
219 request->detached = 0;
220 request->detached_sender = NULL;
222 request->truncated = 0;
223 request->real_size = 0;
224 request->real_tag = 0;
231 if (flags & SEND) smpi_datatype_unuse(datatype);
237 void smpi_empty_status(MPI_Status * status)
239 if(status != MPI_STATUS_IGNORE) {
240 status->MPI_SOURCE = MPI_ANY_SOURCE;
241 status->MPI_TAG = MPI_ANY_TAG;
242 status->MPI_ERROR = MPI_SUCCESS;
247 void smpi_action_trace_run(char *path)
251 xbt_dict_cursor_t cursor;
255 action_fp = fopen(path, "r");
256 xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
260 if (!xbt_dict_is_empty(action_queues)) {
262 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
265 xbt_dict_foreach(action_queues, cursor, name, todo) {
266 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
272 xbt_dict_free(&action_queues);
273 action_queues = xbt_dict_new_homogeneous(NULL);
276 static void smpi_mpi_request_free_voidp(void* request)
278 MPI_Request req = request;
279 smpi_mpi_request_free(&req);
282 /* MPI Low level calls */
283 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
284 int dst, int tag, MPI_Comm comm)
286 MPI_Request request =
287 build_request(buf, count, datatype, smpi_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) {
320 print_request("New recv", request);
321 if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
322 mailbox = smpi_process_mailbox_small();
324 mailbox = smpi_process_mailbox();
325 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
326 request->real_size=request->size;
327 smpi_datatype_use(request->old_type);
328 request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
330 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
331 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
333 simcall_process_sleep(sleeptime);
334 XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
340 int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
343 int rank = smpi_process_index();
344 if (TRACE_smpi_view_internals()) {
345 TRACE_smpi_send(rank, rank, receiver);
348 /* if(receiver == MPI_UNDEFINED) {*/
349 /* XBT_WARN("Trying to send a message to a wrong rank");*/
352 print_request("New send", request);
353 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
354 mailbox = smpi_process_remote_mailbox_small(receiver);
356 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
357 mailbox = smpi_process_remote_mailbox(receiver);
359 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
361 request->detached = 1;
363 if(request->old_type->has_subtype == 0){
364 oldbuf = request->buf;
365 if (oldbuf && request->size!=0){
366 request->buf = xbt_malloc(request->size);
367 memcpy(request->buf,oldbuf,request->size);
370 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
373 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
374 request->real_size=request->size;
375 smpi_datatype_use(request->old_type);
377 //if we are giving back the control to the user without waiting for completion, we have to inject timings
378 double sleeptime =0.0;
379 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
380 //isend and send timings may be different
381 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
385 simcall_process_sleep(sleeptime);
386 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
390 simcall_comm_isend(mailbox, request->size, -1.0,
391 request->buf, request->real_size,
393 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
395 // detach if msg size < eager/rdv switch limit
399 /* FIXME: detached sends are not traceable (request->action == NULL) */
401 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
409 void smpi_mpi_startall(int count, MPI_Request * requests)
413 for(i = 0; i < count; i++) {
414 smpi_mpi_start(requests[i]);
418 void smpi_mpi_request_free(MPI_Request * request)
421 if((*request) != MPI_REQUEST_NULL){
422 (*request)->refcount--;
423 if((*request)->refcount<0) xbt_die("wrong refcount");
425 if((*request)->refcount==0){
427 *request = MPI_REQUEST_NULL;
430 xbt_die("freeing an already free request");
434 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
435 int dst, int tag, MPI_Comm comm)
437 MPI_Request request =
438 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
439 comm, NON_PERSISTENT | SEND);
444 MPI_Request smpi_mpi_isend(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 | ISEND | SEND);
451 smpi_mpi_start(request);
455 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
456 int dst, int tag, MPI_Comm comm)
458 MPI_Request request =
459 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
460 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
461 smpi_mpi_start(request);
467 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
468 int src, int tag, MPI_Comm comm)
470 MPI_Request request =
471 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
472 comm, NON_PERSISTENT | RECV);
476 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
477 int src, int tag, MPI_Comm comm)
479 MPI_Request request =
480 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
481 comm, NON_PERSISTENT | RECV);
483 smpi_mpi_start(request);
487 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
488 int tag, MPI_Comm comm, MPI_Status * status)
491 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
492 smpi_mpi_wait(&request, status);
497 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
498 int tag, MPI_Comm comm)
500 MPI_Request request =
501 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
502 comm, NON_PERSISTENT | SEND);
504 smpi_mpi_start(request);
505 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
509 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
510 int dst, int tag, MPI_Comm comm)
512 MPI_Request request =
513 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
514 comm, NON_PERSISTENT | SSEND | SEND);
516 smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
519 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
520 int dst, int sendtag, void *recvbuf, int recvcount,
521 MPI_Datatype recvtype, int src, int recvtag,
522 MPI_Comm comm, MPI_Status * status)
524 MPI_Request requests[2];
528 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
530 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
531 smpi_mpi_startall(2, requests);
532 smpi_mpi_waitall(2, requests, stats);
533 if(status != MPI_STATUS_IGNORE) {
534 // Copy receive status
539 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
541 return status->count / smpi_datatype_size(datatype);
544 static void finish_wait(MPI_Request * request, MPI_Status * status)
546 MPI_Request req = *request;
547 if(status != MPI_STATUS_IGNORE)
548 smpi_empty_status(status);
550 if(!(req->detached && req->flags & SEND)){
551 if(status != MPI_STATUS_IGNORE) {
552 status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
553 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
554 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
555 // this handles the case were size in receive differs from size in send
556 // FIXME: really this should just contain the count of receive-type blocks,
558 status->count = req->real_size;
561 print_request("Finishing", req);
562 MPI_Datatype datatype = req->old_type;
564 if(datatype->has_subtype == 1){
565 // This part handles the problem of non-contignous memory
566 // the unserialization at the reception
567 s_smpi_subtype_t *subtype = datatype->substruct;
568 if(req->flags & RECV) {
569 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
571 if(req->detached == 0) free(req->buf);
573 smpi_datatype_unuse(datatype);
578 if (TRACE_smpi_view_internals()) {
579 if(req->flags & RECV){
580 int rank = smpi_process_index();
581 int src_traced = smpi_group_index(smpi_comm_group(req->comm), req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
582 TRACE_smpi_recv(rank, src_traced, rank);
587 if(req->detached_sender!=NULL){
588 smpi_mpi_request_free(&(req->detached_sender));
591 if(req->flags & NON_PERSISTENT) {
592 smpi_mpi_request_free(request);
598 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
601 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
602 if ((*request)->action == NULL)
605 flag = simcall_comm_test((*request)->action);
607 finish_wait(request, status);
608 request=MPI_REQUEST_NULL;
610 smpi_empty_status(status);
615 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
622 *index = MPI_UNDEFINED;
625 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
626 map = xbt_new(int, count);
628 for(i = 0; i < count; i++) {
629 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
630 xbt_dynar_push(comms, &requests[i]->action);
636 i = simcall_comm_testany(comms);
637 // not MPI_UNDEFINED, as this is a simix return code
640 finish_wait(&requests[*index], status);
644 //all requests are null or inactive, return true
646 smpi_empty_status(status);
649 xbt_dynar_free(&comms);
656 int smpi_mpi_testall(int count, MPI_Request requests[],
660 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
663 for(i=0; i<count; i++){
664 if(requests[i]!= MPI_REQUEST_NULL){
665 if (smpi_mpi_test(&requests[i], pstat)!=1){
669 smpi_empty_status(pstat);
671 if(status != MPI_STATUSES_IGNORE) {
678 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
680 //FIXME find another wait to avoid busy waiting ?
681 // the issue here is that we have to wait on a nonexistent comm
683 smpi_mpi_iprobe(source, tag, comm, &flag, status);
684 XBT_DEBUG("Busy Waiting on probing : %d", flag);
686 simcall_process_sleep(0.0001);
691 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
692 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
693 comm, NON_PERSISTENT | RECV);
695 // behave like a receive, but don't do it
698 print_request("New iprobe", request);
699 // We have to test both mailboxes as we don't know if we will receive one one or another
700 if (sg_cfg_get_int("smpi/async_small_thres")>0){
701 mailbox = smpi_process_mailbox_small();
702 XBT_DEBUG("trying to probe the perm recv mailbox");
703 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
705 if (request->action==NULL){
706 mailbox = smpi_process_mailbox();
707 XBT_DEBUG("trying to probe the other mailbox");
708 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
712 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
714 if(status != MPI_STATUS_IGNORE) {
715 status->MPI_SOURCE = req->src;
716 status->MPI_TAG = req->tag;
717 status->MPI_ERROR = MPI_SUCCESS;
718 status->count = req->real_size;
722 smpi_mpi_request_free(&request);
727 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
729 print_request("Waiting", *request);
730 if ((*request)->action != NULL) { // this is not a detached send
731 simcall_comm_wait((*request)->action, -1.0);
733 finish_wait(request, status);
735 // FIXME for a detached send, finish_wait is not called:
738 int smpi_mpi_waitany(int count, MPI_Request requests[],
745 index = MPI_UNDEFINED;
747 // Wait for a request to complete
748 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
749 map = xbt_new(int, count);
751 XBT_DEBUG("Wait for one of %d", count);
752 for(i = 0; i < count; i++) {
753 if(requests[i] != MPI_REQUEST_NULL) {
754 if (requests[i]->action != NULL) {
755 XBT_DEBUG("Waiting any %p ", requests[i]);
756 xbt_dynar_push(comms, &requests[i]->action);
760 //This is a finished detached request, let's return this one
761 size=0;//so we free the dynar but don't do the waitany call
763 finish_wait(&requests[i], status);//cleanup if refcount = 0
764 requests[i]=MPI_REQUEST_NULL;//set to null
770 i = simcall_comm_waitany(comms);
772 // not MPI_UNDEFINED, as this is a simix return code
775 finish_wait(&requests[index], status);
779 xbt_dynar_free(&comms);
782 if (index==MPI_UNDEFINED)
783 smpi_empty_status(status);
788 int smpi_mpi_waitall(int count, MPI_Request requests[],
793 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
794 int retvalue = MPI_SUCCESS;
795 //tag invalid requests in the set
796 if (status != MPI_STATUSES_IGNORE) {
797 for (c = 0; c < count; c++) {
798 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
799 smpi_empty_status(&status[c]);
800 } else if (requests[c]->src == MPI_PROC_NULL) {
801 smpi_empty_status(&status[c]);
802 status[c].MPI_SOURCE = MPI_PROC_NULL;
806 for(c = 0; c < count; c++) {
807 if (MC_is_active()) {
808 smpi_mpi_wait(&requests[c], pstat);
811 index = smpi_mpi_waitany(count, requests, pstat);
812 if (index == MPI_UNDEFINED)
814 requests[index]=MPI_REQUEST_NULL;
816 if (status != MPI_STATUSES_IGNORE) {
817 status[index] = *pstat;
818 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
819 retvalue = MPI_ERR_IN_STATUS;
826 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
831 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
834 for(i = 0; i < incount; i++)
836 index=smpi_mpi_waitany(incount, requests, pstat);
837 if(index!=MPI_UNDEFINED){
838 indices[count] = index;
840 if(status != MPI_STATUSES_IGNORE) {
841 status[index] = *pstat;
843 requests[index]=MPI_REQUEST_NULL;
845 return MPI_UNDEFINED;
851 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
854 int i, count, count_dead;
856 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
860 for(i = 0; i < incount; i++) {
861 if((requests[i] != MPI_REQUEST_NULL)) {
862 if(smpi_mpi_test(&requests[i], pstat)) {
865 if(status != MPI_STATUSES_IGNORE) {
868 requests[i]=MPI_REQUEST_NULL;
875 if(count_dead==incount)return MPI_UNDEFINED;
879 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
882 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
883 nary_tree_bcast(buf, count, datatype, root, comm, 4);
886 void smpi_mpi_barrier(MPI_Comm comm)
888 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
889 nary_tree_barrier(comm, 4);
892 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
893 void *recvbuf, int recvcount, MPI_Datatype recvtype,
894 int root, MPI_Comm comm)
896 int system_tag = 666;
897 int rank, size, src, index;
898 MPI_Aint lb = 0, recvext = 0;
899 MPI_Request *requests;
901 rank = smpi_comm_rank(comm);
902 size = smpi_comm_size(comm);
904 // Send buffer to root
905 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
907 // FIXME: check for errors
908 smpi_datatype_extent(recvtype, &lb, &recvext);
909 // Local copy from root
910 smpi_datatype_copy(sendbuf, sendcount, sendtype,
911 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
912 // Receive buffers from senders
913 requests = xbt_new(MPI_Request, size - 1);
915 for(src = 0; src < size; src++) {
917 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
919 src, system_tag, comm);
923 // Wait for completion of irecv's.
924 smpi_mpi_startall(size - 1, requests);
925 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
930 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
931 void *recvbuf, int *recvcounts, int *displs,
932 MPI_Datatype recvtype, int root, MPI_Comm comm)
934 int system_tag = 666;
935 int rank, size, src, index;
936 MPI_Aint lb = 0, recvext = 0;
937 MPI_Request *requests;
939 rank = smpi_comm_rank(comm);
940 size = smpi_comm_size(comm);
942 // Send buffer to root
943 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
945 // FIXME: check for errors
946 smpi_datatype_extent(recvtype, &lb, &recvext);
947 // Local copy from root
948 smpi_datatype_copy(sendbuf, sendcount, sendtype,
949 (char *)recvbuf + displs[root] * recvext,
950 recvcounts[root], recvtype);
951 // Receive buffers from senders
952 requests = xbt_new(MPI_Request, size - 1);
954 for(src = 0; src < size; src++) {
957 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
958 recvcounts[src], recvtype, src, system_tag, comm);
962 // Wait for completion of irecv's.
963 smpi_mpi_startall(size - 1, requests);
964 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
969 void smpi_mpi_allgather(void *sendbuf, int sendcount,
970 MPI_Datatype sendtype, void *recvbuf,
971 int recvcount, MPI_Datatype recvtype,
974 int system_tag = 666;
975 int rank, size, other, index;
976 MPI_Aint lb = 0, recvext = 0;
977 MPI_Request *requests;
979 rank = smpi_comm_rank(comm);
980 size = smpi_comm_size(comm);
981 // FIXME: check for errors
982 smpi_datatype_extent(recvtype, &lb, &recvext);
983 // Local copy from self
984 smpi_datatype_copy(sendbuf, sendcount, sendtype,
985 (char *)recvbuf + rank * recvcount * recvext, recvcount,
987 // Send/Recv buffers to/from others;
988 requests = xbt_new(MPI_Request, 2 * (size - 1));
990 for(other = 0; other < size; other++) {
993 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
996 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
997 recvcount, recvtype, other,
1002 // Wait for completion of all comms.
1003 smpi_mpi_startall(2 * (size - 1), requests);
1004 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1008 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1009 MPI_Datatype sendtype, void *recvbuf,
1010 int *recvcounts, int *displs,
1011 MPI_Datatype recvtype, MPI_Comm comm)
1013 int system_tag = 666;
1014 int rank, size, other, index;
1015 MPI_Aint lb = 0, recvext = 0;
1016 MPI_Request *requests;
1018 rank = smpi_comm_rank(comm);
1019 size = smpi_comm_size(comm);
1020 // FIXME: check for errors
1021 smpi_datatype_extent(recvtype, &lb, &recvext);
1022 // Local copy from self
1023 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1024 (char *)recvbuf + displs[rank] * recvext,
1025 recvcounts[rank], recvtype);
1026 // Send buffers to others;
1027 requests = xbt_new(MPI_Request, 2 * (size - 1));
1029 for(other = 0; other < size; other++) {
1032 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1036 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1037 recvtype, other, system_tag, comm);
1041 // Wait for completion of all comms.
1042 smpi_mpi_startall(2 * (size - 1), requests);
1043 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1047 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1048 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1049 int root, MPI_Comm comm)
1051 int system_tag = 666;
1052 int rank, size, dst, index;
1053 MPI_Aint lb = 0, sendext = 0;
1054 MPI_Request *requests;
1056 rank = smpi_comm_rank(comm);
1057 size = smpi_comm_size(comm);
1059 // Recv buffer from root
1060 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1063 // FIXME: check for errors
1064 smpi_datatype_extent(sendtype, &lb, &sendext);
1065 // Local copy from root
1066 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1067 sendcount, sendtype, recvbuf, recvcount, recvtype);
1068 // Send buffers to receivers
1069 requests = xbt_new(MPI_Request, size - 1);
1071 for(dst = 0; dst < size; dst++) {
1073 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1074 sendcount, sendtype, dst,
1079 // Wait for completion of isend's.
1080 smpi_mpi_startall(size - 1, requests);
1081 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1086 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1087 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1088 MPI_Datatype recvtype, int root, MPI_Comm comm)
1090 int system_tag = 666;
1091 int rank, size, dst, index;
1092 MPI_Aint lb = 0, sendext = 0;
1093 MPI_Request *requests;
1095 rank = smpi_comm_rank(comm);
1096 size = smpi_comm_size(comm);
1098 // Recv buffer from root
1099 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1102 // FIXME: check for errors
1103 smpi_datatype_extent(sendtype, &lb, &sendext);
1104 // Local copy from root
1105 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1106 sendtype, recvbuf, recvcount, recvtype);
1107 // Send buffers to receivers
1108 requests = xbt_new(MPI_Request, size - 1);
1110 for(dst = 0; dst < size; dst++) {
1113 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1114 sendtype, dst, system_tag, comm);
1118 // Wait for completion of isend's.
1119 smpi_mpi_startall(size - 1, requests);
1120 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1125 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1126 MPI_Datatype datatype, MPI_Op op, int root,
1129 int system_tag = 666;
1130 int rank, size, src, index;
1131 MPI_Aint lb = 0, dataext = 0;
1132 MPI_Request *requests;
1135 rank = smpi_comm_rank(comm);
1136 size = smpi_comm_size(comm);
1138 // Send buffer to root
1139 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
1141 // FIXME: check for errors
1142 smpi_datatype_extent(datatype, &lb, &dataext);
1143 // Local copy from root
1144 if (sendbuf && recvbuf)
1145 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1146 // Receive buffers from senders
1147 //TODO: make a MPI_barrier here ?
1148 requests = xbt_new(MPI_Request, size - 1);
1149 tmpbufs = xbt_new(void *, size - 1);
1151 for(src = 0; src < size; src++) {
1153 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1155 tmpbufs[index] = xbt_malloc(count * dataext);
1157 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1162 // Wait for completion of irecv's.
1163 smpi_mpi_startall(size - 1, requests);
1164 for(src = 0; src < size - 1; src++) {
1165 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1166 XBT_DEBUG("finished waiting any request with index %d", index);
1167 if(index == MPI_UNDEFINED) {
1170 if(op) /* op can be MPI_OP_NULL that does nothing */
1171 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1173 for(index = 0; index < size - 1; index++) {
1174 xbt_free(tmpbufs[index]);
1181 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1182 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1184 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1185 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1188 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1189 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1191 int system_tag = 666;
1192 int rank, size, other, index;
1193 MPI_Aint lb = 0, dataext = 0;
1194 MPI_Request *requests;
1197 rank = smpi_comm_rank(comm);
1198 size = smpi_comm_size(comm);
1200 // FIXME: check for errors
1201 smpi_datatype_extent(datatype, &lb, &dataext);
1203 // Local copy from self
1204 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1206 // Send/Recv buffers to/from others;
1207 requests = xbt_new(MPI_Request, size - 1);
1208 tmpbufs = xbt_new(void *, rank);
1210 for(other = 0; other < rank; other++) {
1211 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1213 tmpbufs[index] = xbt_malloc(count * dataext);
1215 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1219 for(other = rank + 1; other < size; other++) {
1221 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1224 // Wait for completion of all comms.
1225 smpi_mpi_startall(size - 1, requests);
1226 for(other = 0; other < size - 1; other++) {
1227 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1228 if(index == MPI_UNDEFINED) {
1232 // #Request is below rank: it's a irecv
1233 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1236 for(index = 0; index < rank; index++) {
1237 xbt_free(tmpbufs[index]);