1 /* Copyright (c) 2007-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include "xbt/replay.h"
12 #include "simix/smx_private.h"
13 #include "surf/surf.h"
14 #include "simgrid/sg_config.h"
15 #include "colls/colls.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
20 static int match_recv(void* a, void* b, smx_action_t ignored) {
21 MPI_Request ref = (MPI_Request)a;
22 MPI_Request req = (MPI_Request)b;
23 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
25 xbt_assert(ref, "Cannot match recv against null reference");
26 xbt_assert(req, "Cannot match recv against null request");
27 if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
28 && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){
29 //we match, we can transfer some values
30 // FIXME : move this to the copy function ?
31 if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
32 if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
33 if(ref->real_size < req->real_size) ref->truncated = 1;
35 ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
37 XBT_DEBUG("match succeeded");
42 static int match_send(void* a, void* b,smx_action_t ignored) {
43 MPI_Request ref = (MPI_Request)a;
44 MPI_Request req = (MPI_Request)b;
45 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
46 xbt_assert(ref, "Cannot match send against null reference");
47 xbt_assert(req, "Cannot match send against null request");
49 if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
50 && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag))
52 if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
53 if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
54 if(req->real_size < ref->real_size) req->truncated = 1;
56 req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
58 XBT_DEBUG("match succeeded");
64 typedef struct s_smpi_factor *smpi_factor_t;
65 typedef struct s_smpi_factor {
68 double values[4];//arbitrary set to 4
70 xbt_dynar_t smpi_os_values = NULL;
71 xbt_dynar_t smpi_or_values = NULL;
72 xbt_dynar_t smpi_ois_values = NULL;
74 double smpi_wtime_sleep = 0.0;
75 double smpi_iprobe_sleep = 1e-4;
76 double smpi_test_sleep = 1e-4;
79 // Methods used to parse and store the values for timing injections in smpi
80 // These are taken from surf/network.c and generalized to have more factors
81 // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
83 static int factor_cmp(const void *pa, const void *pb)
85 return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 :
86 (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0;
90 static xbt_dynar_t parse_factor(const char *smpi_coef_string)
93 unsigned int iter = 0;
97 xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
99 smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
100 radical_elements = xbt_str_split(smpi_coef_string, ";");
101 xbt_dynar_foreach(radical_elements, iter, value) {
102 memset(&fact, 0, sizeof(s_smpi_factor_t));
103 radical_elements2 = xbt_str_split(value, ":");
104 if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
105 xbt_die("Malformed radical for smpi factor!");
106 for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
108 fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
110 fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
115 xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
116 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
117 xbt_dynar_free(&radical_elements2);
119 xbt_dynar_free(&radical_elements);
121 xbt_dynar_sort(smpi_factor, &factor_cmp);
122 xbt_dynar_foreach(smpi_factor, iter, fact) {
123 XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
128 static double smpi_os(double size)
130 if (!smpi_os_values) {
131 smpi_os_values = parse_factor(sg_cfg_get_string("smpi/os"));
132 smpi_register_static(smpi_os_values, xbt_dynar_free_voidp);
134 unsigned int iter = 0;
135 s_smpi_factor_t fact;
137 xbt_dynar_foreach(smpi_os_values, iter, fact) {
138 if (size <= fact.factor) {
139 XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
142 current=fact.values[0]+fact.values[1]*size;
145 XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
150 static double smpi_ois(double size)
152 if (!smpi_ois_values) {
153 smpi_ois_values = parse_factor(sg_cfg_get_string("smpi/ois"));
154 smpi_register_static(smpi_ois_values, xbt_dynar_free_voidp);
156 unsigned int iter = 0;
157 s_smpi_factor_t fact;
159 xbt_dynar_foreach(smpi_ois_values, iter, fact) {
160 if (size <= fact.factor) {
161 XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
164 current=fact.values[0]+fact.values[1]*size;
167 XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
172 static double smpi_or(double size)
174 if (!smpi_or_values) {
175 smpi_or_values = parse_factor(sg_cfg_get_string("smpi/or"));
176 smpi_register_static(smpi_or_values, xbt_dynar_free_voidp);
178 unsigned int iter = 0;
179 s_smpi_factor_t fact;
181 xbt_dynar_foreach(smpi_or_values, iter, fact) {
182 if (size <= fact.factor) {
183 XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
186 current=fact.values[0]+fact.values[1]*size;
188 XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
193 double smpi_mpi_wtime(){
195 if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
197 time = SIMIX_get_clock();
198 //to avoid deadlocks if called too many times
199 if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep);
202 time = SIMIX_get_clock();
207 static MPI_Request build_request(void *buf, int count,
208 MPI_Datatype datatype, int src, int dst,
209 int tag, MPI_Comm comm, unsigned flags)
211 MPI_Request request = NULL;
213 void *old_buf = NULL;
215 request = xbt_new(s_smpi_mpi_request_t, 1);
217 s_smpi_subtype_t *subtype = datatype->substruct;
219 if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){
220 // This part handles the problem of non-contiguous memory
222 buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
223 if ((datatype->has_subtype == 1) && (flags & SEND)) {
224 subtype->serialize(old_buf, buf, count, datatype->substruct);
229 // This part handles the problem of non-contiguous memory (for the
230 // unserialisation at the reception)
231 request->old_buf = old_buf;
232 request->old_type = datatype;
234 request->size = smpi_datatype_size(datatype) * count;
238 request->comm = comm;
239 request->action = NULL;
240 request->flags = flags;
241 request->detached = 0;
242 request->detached_sender = NULL;
243 request->real_src = 0;
245 request->truncated = 0;
246 request->real_size = 0;
247 request->real_tag = 0;
248 if(flags & PERSISTENT)
249 request->refcount = 1;
251 request->refcount = 0;
252 request->op = MPI_REPLACE;
257 if (flags & SEND) smpi_datatype_unuse(datatype);
263 void smpi_empty_status(MPI_Status * status)
265 if(status != MPI_STATUS_IGNORE) {
266 status->MPI_SOURCE = MPI_ANY_SOURCE;
267 status->MPI_TAG = MPI_ANY_TAG;
268 status->MPI_ERROR = MPI_SUCCESS;
273 void smpi_action_trace_run(char *path)
277 xbt_dict_cursor_t cursor;
281 action_fp = fopen(path, "r");
282 if (action_fp == NULL)
283 xbt_die("Cannot open %s: %s", path, strerror(errno));
286 if (!xbt_dict_is_empty(action_queues)) {
288 ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
291 xbt_dict_foreach(action_queues, cursor, name, todo) {
292 XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
298 xbt_dict_free(&action_queues);
299 action_queues = xbt_dict_new_homogeneous(NULL);
302 static void smpi_mpi_request_free_voidp(void* request)
304 MPI_Request req = request;
305 smpi_mpi_request_free(&req);
308 /* MPI Low level calls */
309 MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
310 int dst, int tag, MPI_Comm comm)
312 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
313 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
314 comm, PERSISTENT | SEND | PREPARED);
318 MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
319 int dst, int tag, MPI_Comm comm)
321 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
322 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
323 comm, PERSISTENT | SSEND | SEND | PREPARED);
327 MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
328 int src, int tag, MPI_Comm comm)
330 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
331 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
332 comm, PERSISTENT | RECV | PREPARED);
336 void smpi_mpi_start(MPI_Request request)
340 xbt_assert(!request->action, "Cannot (re)start a non-finished communication");
341 request->flags &= ~PREPARED;
342 request->flags &= ~FINISHED;
345 if (request->flags & RECV) {
346 print_request("New recv", request);
347 //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match !
348 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres"))
349 mailbox = smpi_process_mailbox_small();
351 mailbox = smpi_process_mailbox();
352 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
353 request->real_size=request->size;
354 smpi_datatype_use(request->old_type);
355 smpi_comm_use(request->comm);
356 request->action = simcall_comm_irecv(mailbox, request->buf,
357 &request->real_size, &match_recv,
358 &smpi_comm_copy_buffer_callback,
361 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
362 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
364 simcall_process_sleep(sleeptime);
365 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
371 int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst);
374 int rank = request->src;
375 if (TRACE_smpi_view_internals()) {
376 TRACE_smpi_send(rank, rank, receiver,request->size);
379 /* if(receiver == MPI_UNDEFINED) {*/
380 /* XBT_WARN("Trying to send a message to a wrong rank");*/
383 print_request("New send", request);
384 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
385 mailbox = smpi_process_remote_mailbox_small(receiver);
387 XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
388 mailbox = smpi_process_remote_mailbox(receiver);
391 void* buf = request->buf;
392 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
394 request->detached = 1;
395 XBT_DEBUG("Send request %p is detached", request);
397 if(request->old_type->has_subtype == 0){
398 oldbuf = request->buf;
399 if (!_xbt_replay_is_active() && oldbuf && request->size!=0){
400 if((smpi_privatize_global_variables)
401 && ((char*)request->buf >= start_data_exe)
402 && ((char*)request->buf < start_data_exe + size_data_exe )){
403 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
404 switch_data_segment(request->src);
406 buf = xbt_malloc(request->size);
407 memcpy(buf,oldbuf,request->size);
408 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
413 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
414 request->real_size=request->size;
415 smpi_datatype_use(request->old_type);
416 smpi_comm_use(request->comm);
418 //if we are giving back the control to the user without waiting for completion, we have to inject timings
419 double sleeptime = 0.0;
420 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
421 //isend and send timings may be different
422 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
425 if(sleeptime != 0.0){
426 simcall_process_sleep(sleeptime);
427 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
430 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
431 buf, request->real_size,
433 &xbt_free_f, // how to free the userdata if a detached send fails
434 &smpi_comm_copy_buffer_callback,
436 // detach if msg size < eager/rdv switch limit
440 /* FIXME: detached sends are not traceable (request->action == NULL) */
442 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
450 void smpi_mpi_startall(int count, MPI_Request * requests)
453 if(requests==NULL) return;
455 for(i = 0; i < count; i++) {
456 smpi_mpi_start(requests[i]);
460 void smpi_mpi_request_free(MPI_Request * request)
462 if((*request) != MPI_REQUEST_NULL){
463 (*request)->refcount--;
464 if((*request)->refcount<0) xbt_die("wrong refcount");
466 if((*request)->refcount==0){
467 print_request("Destroying", (*request));
469 *request = MPI_REQUEST_NULL;
471 print_request("Decrementing", (*request));
474 xbt_die("freeing an already free request");
479 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
480 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
482 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
484 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
485 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
487 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
488 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
494 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
495 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
497 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
499 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
500 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
502 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
503 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
510 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
511 int dst, int tag, MPI_Comm comm)
513 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
514 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
515 comm, PERSISTENT | ISEND | SEND | PREPARED);
519 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
520 int dst, int tag, MPI_Comm comm)
522 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
523 request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
524 comm, NON_PERSISTENT | ISEND | SEND);
525 smpi_mpi_start(request);
529 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
530 int dst, int tag, MPI_Comm comm)
532 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
533 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
534 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
535 smpi_mpi_start(request);
539 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
540 int src, int tag, MPI_Comm comm)
542 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
543 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
544 comm, PERSISTENT | RECV | PREPARED);
548 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
549 int src, int tag, MPI_Comm comm)
551 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
552 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
553 comm, NON_PERSISTENT | RECV);
554 smpi_mpi_start(request);
558 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
559 int tag, MPI_Comm comm, MPI_Status * status)
561 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
562 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
563 smpi_mpi_wait(&request, status);
569 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
570 int tag, MPI_Comm comm)
572 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
573 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
574 comm, NON_PERSISTENT | SEND);
576 smpi_mpi_start(request);
577 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
581 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
582 int dst, int tag, MPI_Comm comm)
584 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
585 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
586 comm, NON_PERSISTENT | SSEND | SEND);
588 smpi_mpi_start(request);
589 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
593 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
594 int dst, int sendtag, void *recvbuf, int recvcount,
595 MPI_Datatype recvtype, int src, int recvtag,
596 MPI_Comm comm, MPI_Status * status)
598 MPI_Request requests[2];
600 int myid=smpi_process_index();
601 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
602 smpi_datatype_copy(sendbuf, sendcount, sendtype,
603 recvbuf, recvcount, recvtype);
607 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
609 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
610 smpi_mpi_startall(2, requests);
611 smpi_mpi_waitall(2, requests, stats);
612 smpi_mpi_request_free(&requests[0]);
613 smpi_mpi_request_free(&requests[1]);
614 if(status != MPI_STATUS_IGNORE) {
615 // Copy receive status
620 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
622 return status->count / smpi_datatype_size(datatype);
625 static void finish_wait(MPI_Request * request, MPI_Status * status)
627 MPI_Request req = *request;
628 smpi_empty_status(status);
630 if(!(req->detached && req->flags & SEND)
631 && !(req->flags & PREPARED)){
632 if(status != MPI_STATUS_IGNORE) {
633 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
634 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
635 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
636 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
637 // this handles the case were size in receive differs from size in send
638 // FIXME: really this should just contain the count of receive-type blocks,
640 status->count = req->real_size;
643 print_request("Finishing", req);
644 MPI_Datatype datatype = req->old_type;
646 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
647 if (!_xbt_replay_is_active()){
648 if( smpi_privatize_global_variables
649 && ((char*)req->old_buf >= start_data_exe)
650 && ((char*)req->old_buf < start_data_exe + size_data_exe )
652 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
653 switch_data_segment(smpi_process_index());
657 if(datatype->has_subtype == 1){
658 // This part handles the problem of non-contignous memory
659 // the unserialization at the reception
660 s_smpi_subtype_t *subtype = datatype->substruct;
661 if(req->flags & RECV)
662 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
663 if(req->detached == 0) free(req->buf);
664 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
665 int n =req->real_size/smpi_datatype_size(datatype);
666 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
669 smpi_comm_unuse(req->comm);
670 smpi_datatype_unuse(datatype);
675 if (TRACE_smpi_view_internals()) {
676 if(req->flags & RECV){
677 int rank = smpi_process_index();
678 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
679 TRACE_smpi_recv(rank, src_traced, rank);
684 if(req->detached_sender!=NULL){
685 smpi_mpi_request_free(&(req->detached_sender));
687 if(req->flags & PERSISTENT)
689 req->flags |= FINISHED;
691 smpi_mpi_request_free(request);
695 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
698 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
701 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
702 static int nsleeps = 1;
703 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
705 smpi_empty_status(status);
707 if (!((*request)->flags & PREPARED)) {
708 if ((*request)->action != NULL)
709 flag = simcall_comm_test((*request)->action);
711 finish_wait(request, status);
712 nsleeps=1;//reset the number of sleeps we will do next time
713 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
714 *request = MPI_REQUEST_NULL;
722 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
729 *index = MPI_UNDEFINED;
731 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
732 map = xbt_new(int, count);
734 for(i = 0; i < count; i++) {
735 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
736 !(requests[i]->flags & PREPARED)) {
737 xbt_dynar_push(comms, &requests[i]->action);
743 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
744 static int nsleeps = 1;
745 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
747 i = simcall_comm_testany(comms);
748 // not MPI_UNDEFINED, as this is a simix return code
751 finish_wait(&requests[*index], status);
752 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
753 requests[*index] = MPI_REQUEST_NULL;
760 //all requests are null or inactive, return true
762 smpi_empty_status(status);
765 xbt_dynar_free(&comms);
771 int smpi_mpi_testall(int count, MPI_Request requests[],
775 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
778 for(i=0; i<count; i++){
779 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
780 if (smpi_mpi_test(&requests[i], pstat)!=1){
783 requests[i]=MPI_REQUEST_NULL;
786 smpi_empty_status(pstat);
788 if(status != MPI_STATUSES_IGNORE) {
795 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
797 //FIXME find another wait to avoid busy waiting ?
798 // the issue here is that we have to wait on a nonexistent comm
800 smpi_mpi_iprobe(source, tag, comm, &flag, status);
801 XBT_DEBUG("Busy Waiting on probing : %d", flag);
805 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
807 MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
808 comm, PERSISTENT | RECV);
810 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
811 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
812 static int nsleeps = 1;
813 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
814 // behave like a receive, but don't do it
817 print_request("New iprobe", request);
818 // We have to test both mailboxes as we don't know if we will receive one one or another
819 if (sg_cfg_get_int("smpi/async_small_thres")>0){
820 mailbox = smpi_process_mailbox_small();
821 XBT_DEBUG("trying to probe the perm recv mailbox");
822 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
824 if (request->action==NULL){
825 mailbox = smpi_process_mailbox();
826 XBT_DEBUG("trying to probe the other mailbox");
827 request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
831 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
833 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
834 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
835 status->MPI_TAG = req->tag;
836 status->MPI_ERROR = MPI_SUCCESS;
837 status->count = req->real_size;
839 nsleeps=1;//reset the number of sleeps we will do next time
845 smpi_mpi_request_free(&request);
850 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
852 print_request("Waiting", *request);
853 if ((*request)->flags & PREPARED) {
854 smpi_empty_status(status);
858 if ((*request)->action != NULL) { // this is not a detached send
859 simcall_comm_wait((*request)->action, -1.0);
861 if(MC_is_active() && (*request)->action)
862 (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison
866 finish_wait(request, status);
867 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
868 *request = MPI_REQUEST_NULL;
869 // FIXME for a detached send, finish_wait is not called:
872 int smpi_mpi_waitany(int count, MPI_Request requests[],
879 index = MPI_UNDEFINED;
881 // Wait for a request to complete
882 comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
883 map = xbt_new(int, count);
885 XBT_DEBUG("Wait for one of %d", count);
886 for(i = 0; i < count; i++) {
887 if (requests[i] != MPI_REQUEST_NULL
888 && !(requests[i]->flags & PREPARED)
889 && !(requests[i]->flags & FINISHED)) {
890 if (requests[i]->action != NULL) {
891 XBT_DEBUG("Waiting any %p ", requests[i]);
892 xbt_dynar_push(comms, &requests[i]->action);
896 //This is a finished detached request, let's return this one
897 size=0;//so we free the dynar but don't do the waitany call
899 finish_wait(&requests[i], status);//cleanup if refcount = 0
900 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
901 requests[i]=MPI_REQUEST_NULL;//set to null
907 i = simcall_comm_waitany(comms);
909 // not MPI_UNDEFINED, as this is a simix return code
912 finish_wait(&requests[index], status);
913 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
914 requests[index] = MPI_REQUEST_NULL;
918 xbt_dynar_free(&comms);
921 if (index==MPI_UNDEFINED)
922 smpi_empty_status(status);
927 int smpi_mpi_waitall(int count, MPI_Request requests[],
932 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
933 int retvalue = MPI_SUCCESS;
934 //tag invalid requests in the set
935 if (status != MPI_STATUSES_IGNORE) {
936 for (c = 0; c < count; c++) {
937 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
938 (requests[c]->flags & PREPARED)) {
939 smpi_empty_status(&status[c]);
940 } else if (requests[c]->src == MPI_PROC_NULL) {
941 smpi_empty_status(&status[c]);
942 status[c].MPI_SOURCE = MPI_PROC_NULL;
946 for(c = 0; c < count; c++) {
948 if (MC_is_active()) {
949 smpi_mpi_wait(&requests[c], pstat);
952 index = smpi_mpi_waitany(count, requests, pstat);
953 if (index == MPI_UNDEFINED)
955 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
956 requests[index]=MPI_REQUEST_NULL;
958 if (status != MPI_STATUSES_IGNORE) {
959 status[index] = *pstat;
960 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
961 retvalue = MPI_ERR_IN_STATUS;
968 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
973 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
976 for(i = 0; i < incount; i++)
978 index=smpi_mpi_waitany(incount, requests, pstat);
979 if(index!=MPI_UNDEFINED){
980 indices[count] = index;
982 if(status != MPI_STATUSES_IGNORE) {
983 status[index] = *pstat;
985 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
986 requests[index]=MPI_REQUEST_NULL;
988 return MPI_UNDEFINED;
994 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
997 int i, count, count_dead;
999 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1003 for(i = 0; i < incount; i++) {
1004 if((requests[i] != MPI_REQUEST_NULL)) {
1005 if(smpi_mpi_test(&requests[i], pstat)) {
1008 if(status != MPI_STATUSES_IGNORE) {
1011 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1012 requests[i]=MPI_REQUEST_NULL;
1018 if(count_dead==incount)return MPI_UNDEFINED;
1022 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1025 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1026 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1029 void smpi_mpi_barrier(MPI_Comm comm)
1031 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1032 nary_tree_barrier(comm, 4);
1035 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1036 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1037 int root, MPI_Comm comm)
1039 int system_tag = COLL_TAG_GATHER;
1040 int rank, size, src, index;
1041 MPI_Aint lb = 0, recvext = 0;
1042 MPI_Request *requests;
1044 rank = smpi_comm_rank(comm);
1045 size = smpi_comm_size(comm);
1047 // Send buffer to root
1048 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1050 // FIXME: check for errors
1051 smpi_datatype_extent(recvtype, &lb, &recvext);
1052 // Local copy from root
1053 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1054 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1055 // Receive buffers from senders
1056 requests = xbt_new(MPI_Request, size - 1);
1058 for(src = 0; src < size; src++) {
1060 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1061 recvcount, recvtype,
1062 src, system_tag, comm);
1066 // Wait for completion of irecv's.
1067 smpi_mpi_startall(size - 1, requests);
1068 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1069 for(src = 0; src < size-1; src++) {
1070 smpi_mpi_request_free(&requests[src]);
1077 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1078 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1082 int rank = smpi_process_index();
1085 /* arbitrarily choose root as rank 0 */
1086 size = smpi_comm_size(comm);
1088 displs = xbt_new(int, size);
1089 for (i = 0; i < size; i++) {
1091 count += recvcounts[i];
1093 tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1094 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1095 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1096 recvcounts[rank], datatype, 0, comm);
1101 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1102 void *recvbuf, int *recvcounts, int *displs,
1103 MPI_Datatype recvtype, int root, MPI_Comm comm)
1105 int system_tag = COLL_TAG_GATHERV;
1106 int rank, size, src, index;
1107 MPI_Aint lb = 0, recvext = 0;
1108 MPI_Request *requests;
1110 rank = smpi_comm_rank(comm);
1111 size = smpi_comm_size(comm);
1113 // Send buffer to root
1114 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1116 // FIXME: check for errors
1117 smpi_datatype_extent(recvtype, &lb, &recvext);
1118 // Local copy from root
1119 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1120 (char *)recvbuf + displs[root] * recvext,
1121 recvcounts[root], recvtype);
1122 // Receive buffers from senders
1123 requests = xbt_new(MPI_Request, size - 1);
1125 for(src = 0; src < size; src++) {
1128 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1129 recvcounts[src], recvtype, src, system_tag, comm);
1133 // Wait for completion of irecv's.
1134 smpi_mpi_startall(size - 1, requests);
1135 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1136 for(src = 0; src < size-1; src++) {
1137 smpi_mpi_request_free(&requests[src]);
1143 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1144 MPI_Datatype sendtype, void *recvbuf,
1145 int recvcount, MPI_Datatype recvtype,
1148 int system_tag = COLL_TAG_ALLGATHER;
1149 int rank, size, other, index;
1150 MPI_Aint lb = 0, recvext = 0;
1151 MPI_Request *requests;
1153 rank = smpi_comm_rank(comm);
1154 size = smpi_comm_size(comm);
1155 // FIXME: check for errors
1156 smpi_datatype_extent(recvtype, &lb, &recvext);
1157 // Local copy from self
1158 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1159 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1161 // Send/Recv buffers to/from others;
1162 requests = xbt_new(MPI_Request, 2 * (size - 1));
1164 for(other = 0; other < size; other++) {
1167 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1170 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1171 recvcount, recvtype, other,
1176 // Wait for completion of all comms.
1177 smpi_mpi_startall(2 * (size - 1), requests);
1178 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1179 for(other = 0; other < 2*(size-1); other++) {
1180 smpi_mpi_request_free(&requests[other]);
1185 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1186 MPI_Datatype sendtype, void *recvbuf,
1187 int *recvcounts, int *displs,
1188 MPI_Datatype recvtype, MPI_Comm comm)
1190 int system_tag = COLL_TAG_ALLGATHERV;
1191 int rank, size, other, index;
1192 MPI_Aint lb = 0, recvext = 0;
1193 MPI_Request *requests;
1195 rank = smpi_comm_rank(comm);
1196 size = smpi_comm_size(comm);
1197 // FIXME: check for errors
1198 smpi_datatype_extent(recvtype, &lb, &recvext);
1199 // Local copy from self
1200 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1201 (char *)recvbuf + displs[rank] * recvext,
1202 recvcounts[rank], recvtype);
1203 // Send buffers to others;
1204 requests = xbt_new(MPI_Request, 2 * (size - 1));
1206 for(other = 0; other < size; other++) {
1209 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1213 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1214 recvtype, other, system_tag, comm);
1218 // Wait for completion of all comms.
1219 smpi_mpi_startall(2 * (size - 1), requests);
1220 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1221 for(other = 0; other < 2*(size-1); other++) {
1222 smpi_mpi_request_free(&requests[other]);
1227 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1228 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1229 int root, MPI_Comm comm)
1231 int system_tag = COLL_TAG_SCATTER;
1232 int rank, size, dst, index;
1233 MPI_Aint lb = 0, sendext = 0;
1234 MPI_Request *requests;
1236 rank = smpi_comm_rank(comm);
1237 size = smpi_comm_size(comm);
1239 // Recv buffer from root
1240 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1243 // FIXME: check for errors
1244 smpi_datatype_extent(sendtype, &lb, &sendext);
1245 // Local copy from root
1246 if(recvbuf!=MPI_IN_PLACE){
1247 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1248 sendcount, sendtype, recvbuf, recvcount, recvtype);
1250 // Send buffers to receivers
1251 requests = xbt_new(MPI_Request, size - 1);
1253 for(dst = 0; dst < size; dst++) {
1255 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1256 sendcount, sendtype, dst,
1261 // Wait for completion of isend's.
1262 smpi_mpi_startall(size - 1, requests);
1263 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1264 for(dst = 0; dst < size-1; dst++) {
1265 smpi_mpi_request_free(&requests[dst]);
1271 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1272 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1273 MPI_Datatype recvtype, int root, MPI_Comm comm)
1275 int system_tag = COLL_TAG_SCATTERV;
1276 int rank, size, dst, index;
1277 MPI_Aint lb = 0, sendext = 0;
1278 MPI_Request *requests;
1280 rank = smpi_comm_rank(comm);
1281 size = smpi_comm_size(comm);
1283 // Recv buffer from root
1284 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1287 // FIXME: check for errors
1288 smpi_datatype_extent(sendtype, &lb, &sendext);
1289 // Local copy from root
1290 if(recvbuf!=MPI_IN_PLACE){
1291 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1292 sendtype, recvbuf, recvcount, recvtype);
1294 // Send buffers to receivers
1295 requests = xbt_new(MPI_Request, size - 1);
1297 for(dst = 0; dst < size; dst++) {
1300 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1301 sendtype, dst, system_tag, comm);
1305 // Wait for completion of isend's.
1306 smpi_mpi_startall(size - 1, requests);
1307 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1308 for(dst = 0; dst < size-1; dst++) {
1309 smpi_mpi_request_free(&requests[dst]);
1315 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1316 MPI_Datatype datatype, MPI_Op op, int root,
1319 int system_tag = COLL_TAG_REDUCE;
1320 int rank, size, src, index;
1321 MPI_Aint lb = 0, dataext = 0;
1322 MPI_Request *requests;
1326 char* sendtmpbuf = (char*) sendbuf;
1327 if( sendbuf == MPI_IN_PLACE ) {
1328 sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
1329 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1332 rank = smpi_comm_rank(comm);
1333 size = smpi_comm_size(comm);
1334 //non commutative case, use a working algo from openmpi
1335 if(!smpi_op_is_commute(op)){
1336 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1337 datatype, op, root, comm);
1342 // Send buffer to root
1343 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1345 // FIXME: check for errors
1346 smpi_datatype_extent(datatype, &lb, &dataext);
1347 // Local copy from root
1348 if (sendtmpbuf && recvbuf)
1349 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1350 // Receive buffers from senders
1351 //TODO: make a MPI_barrier here ?
1352 requests = xbt_new(MPI_Request, size - 1);
1353 tmpbufs = xbt_new(void *, size - 1);
1355 for(src = 0; src < size; src++) {
1357 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1359 tmpbufs[index] = xbt_malloc(count * dataext);
1361 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1366 // Wait for completion of irecv's.
1367 smpi_mpi_startall(size - 1, requests);
1368 for(src = 0; src < size - 1; src++) {
1369 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1370 XBT_DEBUG("finished waiting any request with index %d", index);
1371 if(index == MPI_UNDEFINED) {
1374 smpi_mpi_request_free(&requests[index]);
1376 if(op) /* op can be MPI_OP_NULL that does nothing */
1377 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1379 for(index = 0; index < size - 1; index++) {
1380 xbt_free(tmpbufs[index]);
1385 if( sendbuf == MPI_IN_PLACE ) {
1386 xbt_free(sendtmpbuf);
1391 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1392 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1394 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1395 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1398 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1399 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1401 int system_tag = -888;
1402 int rank, size, other, index;
1403 MPI_Aint lb = 0, dataext = 0;
1404 MPI_Request *requests;
1407 rank = smpi_comm_rank(comm);
1408 size = smpi_comm_size(comm);
1410 // FIXME: check for errors
1411 smpi_datatype_extent(datatype, &lb, &dataext);
1413 // Local copy from self
1414 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1416 // Send/Recv buffers to/from others;
1417 requests = xbt_new(MPI_Request, size - 1);
1418 tmpbufs = xbt_new(void *, rank);
1420 for(other = 0; other < rank; other++) {
1421 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1423 tmpbufs[index] = xbt_malloc(count * dataext);
1425 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1429 for(other = rank + 1; other < size; other++) {
1431 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1434 // Wait for completion of all comms.
1435 smpi_mpi_startall(size - 1, requests);
1437 if(smpi_op_is_commute(op)){
1438 for(other = 0; other < size - 1; other++) {
1439 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1440 if(index == MPI_UNDEFINED) {
1444 // #Request is below rank: it's a irecv
1445 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1449 //non commutative case, wait in order
1450 for(other = 0; other < size - 1; other++) {
1451 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1453 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1457 for(index = 0; index < rank; index++) {
1458 xbt_free(tmpbufs[index]);
1460 for(index = 0; index < size-1; index++) {
1461 smpi_mpi_request_free(&requests[index]);
1467 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1468 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1470 int system_tag = -888;
1471 int rank, size, other, index;
1472 MPI_Aint lb = 0, dataext = 0;
1473 MPI_Request *requests;
1475 int recvbuf_is_empty=1;
1476 rank = smpi_comm_rank(comm);
1477 size = smpi_comm_size(comm);
1479 // FIXME: check for errors
1480 smpi_datatype_extent(datatype, &lb, &dataext);
1482 // Send/Recv buffers to/from others;
1483 requests = xbt_new(MPI_Request, size - 1);
1484 tmpbufs = xbt_new(void *, rank);
1486 for(other = 0; other < rank; other++) {
1487 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1489 tmpbufs[index] = xbt_malloc(count * dataext);
1491 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1495 for(other = rank + 1; other < size; other++) {
1497 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1500 // Wait for completion of all comms.
1501 smpi_mpi_startall(size - 1, requests);
1502 if(smpi_op_is_commute(op)){
1503 for(other = 0; other < size - 1; other++) {
1504 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1505 if(index == MPI_UNDEFINED) {
1509 if(recvbuf_is_empty){
1510 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1513 // #Request is below rank: it's a irecv
1514 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1518 //non commutative case, wait in order
1519 for(other = 0; other < size - 1; other++) {
1520 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1522 if(recvbuf_is_empty){
1523 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1525 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1529 for(index = 0; index < rank; index++) {
1530 xbt_free(tmpbufs[index]);
1532 for(index = 0; index < size-1; index++) {
1533 smpi_mpi_request_free(&requests[index]);