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)
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));
339 int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
340 /* if(receiver == MPI_UNDEFINED) {*/
341 /* XBT_WARN("Trying to send a message to a wrong rank");*/
344 print_request("New send", request);
345 if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
346 mailbox = smpi_process_remote_mailbox_small(receiver);
348 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
349 mailbox = smpi_process_remote_mailbox(receiver);
351 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
353 request->detached = 1;
355 if(request->old_type->has_subtype == 0){
356 oldbuf = request->buf;
358 request->buf = xbt_malloc(request->size);
359 memcpy(request->buf,oldbuf,request->size);
362 XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
364 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
365 request->real_size=request->size;
366 smpi_datatype_use(request->old_type);
368 //if we are giving back the control to the user without waiting for completion, we have to inject timings
369 double sleeptime =0.0;
370 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
371 //isend and send timings may be different
372 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
376 simcall_process_sleep(sleeptime);
377 XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
381 simcall_comm_isend(mailbox, request->size, -1.0,
382 request->buf, request->real_size,
384 &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
386 // detach if msg size < eager/rdv switch limit
390 /* FIXME: detached sends are not traceable (request->action == NULL) */
392 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
399 void smpi_mpi_startall(int count, MPI_Request * requests)
403 for(i = 0; i < count; i++) {
404 smpi_mpi_start(requests[i]);
408 void smpi_mpi_request_free(MPI_Request * request)
411 if((*request) != MPI_REQUEST_NULL){
412 (*request)->refcount--;
413 if((*request)->refcount<0) xbt_die("wrong refcount");
415 if((*request)->refcount==0){
417 *request = MPI_REQUEST_NULL;
420 xbt_die("freeing an already free request");
424 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
425 int dst, int tag, MPI_Comm comm)
427 MPI_Request request =
428 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
429 comm, NON_PERSISTENT | SEND);
434 MPI_Request smpi_mpi_isend(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 | ISEND | SEND);
441 smpi_mpi_start(request);
445 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
446 int dst, int tag, MPI_Comm comm)
448 MPI_Request request =
449 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
450 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
451 smpi_mpi_start(request);
457 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
458 int src, int tag, MPI_Comm comm)
460 MPI_Request request =
461 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
462 comm, NON_PERSISTENT | RECV);
466 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
467 int src, int tag, MPI_Comm comm)
469 MPI_Request request =
470 build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
471 comm, NON_PERSISTENT | RECV);
473 smpi_mpi_start(request);
477 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
478 int tag, MPI_Comm comm, MPI_Status * status)
481 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
482 smpi_mpi_wait(&request, status);
487 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
488 int tag, MPI_Comm comm)
490 MPI_Request request =
491 build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
492 comm, NON_PERSISTENT | SEND);
494 smpi_mpi_start(request);
495 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
499 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
500 int dst, int tag, MPI_Comm comm)
502 MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
503 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
506 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
507 int dst, int sendtag, void *recvbuf, int recvcount,
508 MPI_Datatype recvtype, int src, int recvtag,
509 MPI_Comm comm, MPI_Status * status)
511 MPI_Request requests[2];
515 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
517 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
518 smpi_mpi_startall(2, requests);
519 smpi_mpi_waitall(2, requests, stats);
520 if(status != MPI_STATUS_IGNORE) {
521 // Copy receive status
526 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
528 return status->count / smpi_datatype_size(datatype);
531 static void finish_wait(MPI_Request * request, MPI_Status * status)
533 MPI_Request req = *request;
534 if(!(req->detached && req->flags & SEND)){
535 if(status != MPI_STATUS_IGNORE) {
536 status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
537 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
539 status->MPI_ERROR = MPI_ERR_TRUNCATE;
540 else status->MPI_ERROR = MPI_SUCCESS ;
541 // this handles the case were size in receive differs from size in send
542 // FIXME: really this should just contain the count of receive-type blocks,
544 status->count = req->real_size;
547 print_request("Finishing", req);
548 MPI_Datatype datatype = req->old_type;
550 if(datatype->has_subtype == 1){
551 // This part handles the problem of non-contignous memory
552 // the unserialization at the reception
553 s_smpi_subtype_t *subtype = datatype->substruct;
554 if(req->flags & RECV) {
555 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
557 if(req->detached == 0) free(req->buf);
559 smpi_datatype_unuse(datatype);
562 if(req->detached_sender!=NULL){
563 smpi_mpi_request_free(&(req->detached_sender));
566 if(req->flags & NON_PERSISTENT) {
567 smpi_mpi_request_free(request);
573 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
576 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
577 if ((*request)->action == NULL)
580 flag = simcall_comm_test((*request)->action);
582 (*request)->refcount++;
583 finish_wait(request, status);
585 smpi_empty_status(status);
590 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
597 *index = MPI_UNDEFINED;
600 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
601 map = xbt_new(int, count);
603 for(i = 0; i < count; i++) {
604 if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
605 xbt_dynar_push(comms, &requests[i]->action);
611 i = simcall_comm_testany(comms);
612 // not MPI_UNDEFINED, as this is a simix return code
615 finish_wait(&requests[*index], status);
619 //all requests are null or inactive, return true
621 smpi_empty_status(status);
624 xbt_dynar_free(&comms);
631 int smpi_mpi_testall(int count, MPI_Request requests[],
635 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
638 for(i=0; i<count; i++){
639 if(requests[i]!= MPI_REQUEST_NULL){
640 if (smpi_mpi_test(&requests[i], pstat)!=1){
644 smpi_empty_status(pstat);
646 if(status != MPI_STATUSES_IGNORE) {
653 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
655 //FIXME find another wait to avoid busy waiting ?
656 // the issue here is that we have to wait on a nonexistent comm
658 smpi_mpi_iprobe(source, tag, comm, &flag, status);
659 XBT_DEBUG("Busy Waiting on probing : %d", flag);
661 simcall_process_sleep(0.0001);
666 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
667 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
668 comm, NON_PERSISTENT | RECV);
670 // behave like a receive, but don't do it
673 print_request("New iprobe", request);
674 // We have to test both mailboxes as we don't know if we will receive one one or another
675 if (sg_cfg_get_int("smpi/async_small_thres")>0){
676 mailbox = smpi_process_mailbox_small();
677 XBT_DEBUG("trying to probe the perm recv mailbox");
678 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
680 if (request->action==NULL){
681 mailbox = smpi_process_mailbox();
682 XBT_DEBUG("trying to probe the other mailbox");
683 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
687 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
689 if(status != MPI_STATUS_IGNORE) {
690 status->MPI_SOURCE = req->src;
691 status->MPI_TAG = req->tag;
692 status->MPI_ERROR = MPI_SUCCESS;
693 status->count = req->real_size;
697 smpi_mpi_request_free(&request);
702 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
704 print_request("Waiting", *request);
705 if ((*request)->action != NULL) { // this is not a detached send
706 simcall_comm_wait((*request)->action, -1.0);
708 finish_wait(request, status);
710 // FIXME for a detached send, finish_wait is not called:
713 int smpi_mpi_waitany(int count, MPI_Request requests[],
720 index = MPI_UNDEFINED;
722 // Wait for a request to complete
723 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
724 map = xbt_new(int, count);
726 XBT_DEBUG("Wait for one of %d", count);
727 for(i = 0; i < count; i++) {
728 if(requests[i] != MPI_REQUEST_NULL) {
729 if (requests[i]->action != NULL) {
730 XBT_DEBUG("Waiting any %p ", requests[i]);
731 xbt_dynar_push(comms, &requests[i]->action);
735 //This is a finished detached request, let's return this one
736 size=0;//so we free the dynar but don't do the waitany call
738 finish_wait(&requests[i], status);//cleanup if refcount = 0
739 requests[i]=MPI_REQUEST_NULL;//set to null
745 i = simcall_comm_waitany(comms);
747 // not MPI_UNDEFINED, as this is a simix return code
750 finish_wait(&requests[index], status);
754 xbt_dynar_free(&comms);
757 if (index==MPI_UNDEFINED)
758 smpi_empty_status(status);
763 int smpi_mpi_waitall(int count, MPI_Request requests[],
768 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
769 int retvalue=MPI_SUCCESS;
770 //tag invalid requests in the set
771 for(c = 0; c < count; c++) {
772 if(requests[c]==MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ){
773 if(status != MPI_STATUSES_IGNORE)
774 smpi_empty_status(&status[c]);
775 }else if(requests[c]->src == MPI_PROC_NULL ){
776 if(status != MPI_STATUSES_IGNORE) {
777 smpi_empty_status(&status[c]);
778 status[c].MPI_SOURCE=MPI_PROC_NULL;
782 for(c = 0; c < count; c++) {
784 smpi_mpi_wait(&requests[c], pstat);
787 index = smpi_mpi_waitany(count, requests, pstat);
788 if(index == MPI_UNDEFINED) {
791 if (status != MPI_STATUSES_IGNORE) {
792 status[index] = *pstat;
793 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
794 retvalue = MPI_ERR_IN_STATUS;
802 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
807 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
810 for(i = 0; i < incount; i++)
812 index=smpi_mpi_waitany(incount, requests, pstat);
813 if(index!=MPI_UNDEFINED){
814 indices[count] = index;
816 if(status != MPI_STATUSES_IGNORE) {
817 status[index] = *pstat;
820 return MPI_UNDEFINED;
826 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
829 int i, count, count_dead;
831 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
835 for(i = 0; i < incount; i++) {
836 if((requests[i] != MPI_REQUEST_NULL)) {
837 if(smpi_mpi_test(&requests[i], pstat)) {
840 if(status != MPI_STATUSES_IGNORE) {
848 if(count_dead==incount)return MPI_UNDEFINED;
852 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
855 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
856 nary_tree_bcast(buf, count, datatype, root, comm, 4);
859 void smpi_mpi_barrier(MPI_Comm comm)
861 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
862 nary_tree_barrier(comm, 4);
865 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
866 void *recvbuf, int recvcount, MPI_Datatype recvtype,
867 int root, MPI_Comm comm)
869 int system_tag = 666;
870 int rank, size, src, index;
871 MPI_Aint lb = 0, recvext = 0;
872 MPI_Request *requests;
874 rank = smpi_comm_rank(comm);
875 size = smpi_comm_size(comm);
877 // Send buffer to root
878 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
880 // FIXME: check for errors
881 smpi_datatype_extent(recvtype, &lb, &recvext);
882 // Local copy from root
883 smpi_datatype_copy(sendbuf, sendcount, sendtype,
884 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
885 // Receive buffers from senders
886 requests = xbt_new(MPI_Request, size - 1);
888 for(src = 0; src < size; src++) {
890 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
892 src, system_tag, comm);
896 // Wait for completion of irecv's.
897 smpi_mpi_startall(size - 1, requests);
898 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
903 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
904 void *recvbuf, int *recvcounts, int *displs,
905 MPI_Datatype recvtype, int root, MPI_Comm comm)
907 int system_tag = 666;
908 int rank, size, src, index;
909 MPI_Aint lb = 0, recvext = 0;
910 MPI_Request *requests;
912 rank = smpi_comm_rank(comm);
913 size = smpi_comm_size(comm);
915 // Send buffer to root
916 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
918 // FIXME: check for errors
919 smpi_datatype_extent(recvtype, &lb, &recvext);
920 // Local copy from root
921 smpi_datatype_copy(sendbuf, sendcount, sendtype,
922 (char *)recvbuf + displs[root] * recvext,
923 recvcounts[root], recvtype);
924 // Receive buffers from senders
925 requests = xbt_new(MPI_Request, size - 1);
927 for(src = 0; src < size; src++) {
930 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
931 recvcounts[src], recvtype, src, system_tag, comm);
935 // Wait for completion of irecv's.
936 smpi_mpi_startall(size - 1, requests);
937 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
942 void smpi_mpi_allgather(void *sendbuf, int sendcount,
943 MPI_Datatype sendtype, void *recvbuf,
944 int recvcount, MPI_Datatype recvtype,
947 int system_tag = 666;
948 int rank, size, other, index;
949 MPI_Aint lb = 0, recvext = 0;
950 MPI_Request *requests;
952 rank = smpi_comm_rank(comm);
953 size = smpi_comm_size(comm);
954 // FIXME: check for errors
955 smpi_datatype_extent(recvtype, &lb, &recvext);
956 // Local copy from self
957 smpi_datatype_copy(sendbuf, sendcount, sendtype,
958 (char *)recvbuf + rank * recvcount * recvext, recvcount,
960 // Send/Recv buffers to/from others;
961 requests = xbt_new(MPI_Request, 2 * (size - 1));
963 for(other = 0; other < size; other++) {
966 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
969 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
970 recvcount, recvtype, other,
975 // Wait for completion of all comms.
976 smpi_mpi_startall(2 * (size - 1), requests);
977 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
981 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
982 MPI_Datatype sendtype, void *recvbuf,
983 int *recvcounts, int *displs,
984 MPI_Datatype recvtype, MPI_Comm comm)
986 int system_tag = 666;
987 int rank, size, other, index;
988 MPI_Aint lb = 0, recvext = 0;
989 MPI_Request *requests;
991 rank = smpi_comm_rank(comm);
992 size = smpi_comm_size(comm);
993 // FIXME: check for errors
994 smpi_datatype_extent(recvtype, &lb, &recvext);
995 // Local copy from self
996 smpi_datatype_copy(sendbuf, sendcount, sendtype,
997 (char *)recvbuf + displs[rank] * recvext,
998 recvcounts[rank], recvtype);
999 // Send buffers to others;
1000 requests = xbt_new(MPI_Request, 2 * (size - 1));
1002 for(other = 0; other < size; other++) {
1005 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1009 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1010 recvtype, other, system_tag, comm);
1014 // Wait for completion of all comms.
1015 smpi_mpi_startall(2 * (size - 1), requests);
1016 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1020 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1021 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1022 int root, MPI_Comm comm)
1024 int system_tag = 666;
1025 int rank, size, dst, index;
1026 MPI_Aint lb = 0, sendext = 0;
1027 MPI_Request *requests;
1029 rank = smpi_comm_rank(comm);
1030 size = smpi_comm_size(comm);
1032 // Recv buffer from root
1033 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1036 // FIXME: check for errors
1037 smpi_datatype_extent(sendtype, &lb, &sendext);
1038 // Local copy from root
1039 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1040 sendcount, sendtype, recvbuf, recvcount, recvtype);
1041 // Send buffers to receivers
1042 requests = xbt_new(MPI_Request, size - 1);
1044 for(dst = 0; dst < size; dst++) {
1046 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1047 sendcount, sendtype, dst,
1052 // Wait for completion of isend's.
1053 smpi_mpi_startall(size - 1, requests);
1054 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1059 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1060 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1061 MPI_Datatype recvtype, int root, MPI_Comm comm)
1063 int system_tag = 666;
1064 int rank, size, dst, index;
1065 MPI_Aint lb = 0, sendext = 0;
1066 MPI_Request *requests;
1068 rank = smpi_comm_rank(comm);
1069 size = smpi_comm_size(comm);
1071 // Recv buffer from root
1072 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1075 // FIXME: check for errors
1076 smpi_datatype_extent(sendtype, &lb, &sendext);
1077 // Local copy from root
1078 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1079 sendtype, recvbuf, recvcount, recvtype);
1080 // Send buffers to receivers
1081 requests = xbt_new(MPI_Request, size - 1);
1083 for(dst = 0; dst < size; dst++) {
1086 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1087 sendtype, dst, system_tag, comm);
1091 // Wait for completion of isend's.
1092 smpi_mpi_startall(size - 1, requests);
1093 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1098 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1099 MPI_Datatype datatype, MPI_Op op, int root,
1102 int system_tag = 666;
1103 int rank, size, src, index;
1104 MPI_Aint lb = 0, dataext = 0;
1105 MPI_Request *requests;
1108 rank = smpi_comm_rank(comm);
1109 size = smpi_comm_size(comm);
1111 // Send buffer to root
1112 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
1114 // FIXME: check for errors
1115 smpi_datatype_extent(datatype, &lb, &dataext);
1116 // Local copy from root
1117 if (sendbuf && recvbuf)
1118 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1119 // Receive buffers from senders
1120 //TODO: make a MPI_barrier here ?
1121 requests = xbt_new(MPI_Request, size - 1);
1122 tmpbufs = xbt_new(void *, size - 1);
1124 for(src = 0; src < size; src++) {
1126 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1128 tmpbufs[index] = xbt_malloc(count * dataext);
1130 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1135 // Wait for completion of irecv's.
1136 smpi_mpi_startall(size - 1, requests);
1137 for(src = 0; src < size - 1; src++) {
1138 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1139 XBT_DEBUG("finished waiting any request with index %d", index);
1140 if(index == MPI_UNDEFINED) {
1143 if(op) /* op can be MPI_OP_NULL that does nothing */
1144 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1146 for(index = 0; index < size - 1; index++) {
1147 xbt_free(tmpbufs[index]);
1154 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1155 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1157 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1158 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1161 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1162 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1164 int system_tag = 666;
1165 int rank, size, other, index;
1166 MPI_Aint lb = 0, dataext = 0;
1167 MPI_Request *requests;
1170 rank = smpi_comm_rank(comm);
1171 size = smpi_comm_size(comm);
1173 // FIXME: check for errors
1174 smpi_datatype_extent(datatype, &lb, &dataext);
1176 // Local copy from self
1177 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1179 // Send/Recv buffers to/from others;
1180 requests = xbt_new(MPI_Request, size - 1);
1181 tmpbufs = xbt_new(void *, rank);
1183 for(other = 0; other < rank; other++) {
1184 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1186 tmpbufs[index] = xbt_malloc(count * dataext);
1188 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1192 for(other = rank + 1; other < size; other++) {
1194 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1197 // Wait for completion of all comms.
1198 smpi_mpi_startall(size - 1, requests);
1199 for(other = 0; other < size - 1; other++) {
1200 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1201 if(index == MPI_UNDEFINED) {
1205 // #Request is below rank: it's a irecv
1206 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1209 for(index = 0; index < rank; index++) {
1210 xbt_free(tmpbufs[index]);