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_synchro_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_synchro_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);
348 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
349 //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
350 mailbox = smpi_process_mailbox_small();
351 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
352 XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox);
353 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
356 mailbox = smpi_process_mailbox();
357 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
358 action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
360 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
361 XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox);
362 mailbox = smpi_process_mailbox_small();
365 XBT_DEBUG("yes there was something for us in the large mailbox");
368 mailbox = smpi_process_mailbox_small();
369 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
370 smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
373 XBT_DEBUG("No, nothing in the permanent receive mailbox");
374 mailbox = smpi_process_mailbox();
376 XBT_DEBUG("yes there was something for us in the small mailbox");
380 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
381 double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
383 simcall_process_sleep(sleeptime);
384 XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
387 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
388 request->real_size=request->size;
389 smpi_datatype_use(request->old_type);
390 smpi_comm_use(request->comm);
391 request->action = simcall_comm_irecv(mailbox, request->buf,
392 &request->real_size, &match_recv,
393 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
394 : &smpi_comm_null_copy_buffer_callback,
396 XBT_DEBUG("recv simcall posted");
402 int receiver = request->dst;
405 int rank = request->src;
406 if (TRACE_smpi_view_internals()) {
407 TRACE_smpi_send(rank, rank, receiver,request->size);
410 print_request("New send", request);
412 //if we are giving back the control to the user without waiting for completion, we have to inject timings
413 double sleeptime = 0.0;
414 if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
415 //isend and send timings may be different
416 sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
419 if(sleeptime != 0.0){
420 simcall_process_sleep(sleeptime);
421 XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
424 if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
425 mailbox = smpi_process_remote_mailbox(receiver);
426 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
427 smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
429 if (! (request->flags & SSEND)){
430 mailbox = smpi_process_remote_mailbox_small(receiver);
431 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
433 mailbox = smpi_process_remote_mailbox_small(receiver);
434 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
435 action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
437 XBT_DEBUG("No, we are first, send to large mailbox");
438 mailbox = smpi_process_remote_mailbox(receiver);
442 XBT_DEBUG("Yes there was something for us in the large mailbox");
445 mailbox = smpi_process_remote_mailbox(receiver);
446 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
449 void* buf = request->buf;
450 if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
452 request->detached = 1;
453 XBT_DEBUG("Send request %p is detached", request);
455 if(request->old_type->has_subtype == 0){
456 oldbuf = request->buf;
457 if (!smpi_process_get_replaying() && oldbuf && request->size!=0){
458 if((smpi_privatize_global_variables)
459 && ((char*)request->buf >= start_data_exe)
460 && ((char*)request->buf < start_data_exe + size_data_exe )){
461 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
462 smpi_switch_data_segment(request->src);
464 buf = xbt_malloc(request->size);
465 memcpy(buf,oldbuf,request->size);
466 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
471 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
472 request->real_size=request->size;
473 smpi_datatype_use(request->old_type);
474 smpi_comm_use(request->comm);
476 simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
477 buf, request->real_size,
479 &xbt_free_f, // how to free the userdata if a detached send fails
480 !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
481 : &smpi_comm_null_copy_buffer_callback,
483 // detach if msg size < eager/rdv switch limit
485 XBT_DEBUG("send simcall posted");
490 /* FIXME: detached sends are not traceable (request->action == NULL) */
492 simcall_set_category(request->action, TRACE_internal_smpi_get_category());
500 void smpi_mpi_startall(int count, MPI_Request * requests)
503 if(requests==NULL) return;
505 for(i = 0; i < count; i++) {
506 smpi_mpi_start(requests[i]);
510 void smpi_mpi_request_free(MPI_Request * request)
512 if((*request) != MPI_REQUEST_NULL){
513 (*request)->refcount--;
514 if((*request)->refcount<0) xbt_die("wrong refcount");
516 if((*request)->refcount==0){
517 print_request("Destroying", (*request));
519 *request = MPI_REQUEST_NULL;
521 print_request("Decrementing", (*request));
524 xbt_die("freeing an already free request");
529 MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype,
530 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
532 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
534 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag,
535 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
537 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
538 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
544 MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype,
545 int src, int dst, int tag, MPI_Comm comm, MPI_Op op)
547 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
549 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
550 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
552 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag,
553 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
560 MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
561 int dst, int tag, MPI_Comm comm)
563 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
564 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
565 comm, PERSISTENT | ISEND | SEND | PREPARED);
569 MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
570 int dst, 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 | ISEND | SEND);
575 smpi_mpi_start(request);
579 MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
580 int dst, int tag, MPI_Comm comm)
582 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
583 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
584 comm, NON_PERSISTENT | ISEND | SSEND | SEND);
585 smpi_mpi_start(request);
589 MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
590 int src, int tag, MPI_Comm comm)
592 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
593 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,
594 comm, PERSISTENT | RECV | PREPARED);
598 MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
599 int src, int tag, MPI_Comm comm)
601 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
602 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,
603 comm, NON_PERSISTENT | RECV);
604 smpi_mpi_start(request);
608 void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
609 int tag, MPI_Comm comm, MPI_Status * status)
611 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
612 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
613 smpi_mpi_wait(&request, status);
619 void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
620 int tag, MPI_Comm comm)
622 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
623 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
624 comm, NON_PERSISTENT | SEND);
626 smpi_mpi_start(request);
627 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
631 void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
632 int dst, int tag, MPI_Comm comm)
634 MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
635 request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag,
636 comm, NON_PERSISTENT | SSEND | SEND);
638 smpi_mpi_start(request);
639 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
643 void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
644 int dst, int sendtag, void *recvbuf, int recvcount,
645 MPI_Datatype recvtype, int src, int recvtag,
646 MPI_Comm comm, MPI_Status * status)
648 MPI_Request requests[2];
650 int myid=smpi_process_index();
651 if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) {
652 smpi_datatype_copy(sendbuf, sendcount, sendtype,
653 recvbuf, recvcount, recvtype);
657 smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
659 smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
660 smpi_mpi_startall(2, requests);
661 smpi_mpi_waitall(2, requests, stats);
662 smpi_mpi_request_free(&requests[0]);
663 smpi_mpi_request_free(&requests[1]);
664 if(status != MPI_STATUS_IGNORE) {
665 // Copy receive status
670 int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
672 return status->count / smpi_datatype_size(datatype);
675 static void finish_wait(MPI_Request * request, MPI_Status * status)
677 MPI_Request req = *request;
678 smpi_empty_status(status);
680 if(!(req->detached && req->flags & SEND)
681 && !(req->flags & PREPARED)){
682 if(status != MPI_STATUS_IGNORE) {
683 int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
684 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src);
685 status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
686 status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
687 // this handles the case were size in receive differs from size in send
688 // FIXME: really this should just contain the count of receive-type blocks,
690 status->count = req->real_size;
693 print_request("Finishing", req);
694 MPI_Datatype datatype = req->old_type;
696 if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
697 if (!smpi_process_get_replaying()){
698 if( smpi_privatize_global_variables
699 && ((char*)req->old_buf >= start_data_exe)
700 && ((char*)req->old_buf < start_data_exe + size_data_exe )
702 XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
703 smpi_switch_data_segment(smpi_process_index());
707 if(datatype->has_subtype == 1){
708 // This part handles the problem of non-contignous memory
709 // the unserialization at the reception
710 s_smpi_subtype_t *subtype = datatype->substruct;
711 if(req->flags & RECV)
712 subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
713 if(req->detached == 0) free(req->buf);
714 }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
715 int n =req->real_size/smpi_datatype_size(datatype);
716 smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
719 smpi_comm_unuse(req->comm);
720 smpi_datatype_unuse(datatype);
725 if (TRACE_smpi_view_internals()) {
726 if(req->flags & RECV){
727 int rank = smpi_process_index();
728 int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
729 TRACE_smpi_recv(rank, src_traced, rank);
734 if(req->detached_sender!=NULL){
735 smpi_mpi_request_free(&(req->detached_sender));
737 if(req->flags & PERSISTENT)
739 req->flags |= FINISHED;
741 smpi_mpi_request_free(request);
745 int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
748 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
751 //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
752 static int nsleeps = 1;
753 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
755 smpi_empty_status(status);
757 if (!((*request)->flags & PREPARED)) {
758 if ((*request)->action != NULL)
759 flag = simcall_comm_test((*request)->action);
761 finish_wait(request, status);
762 nsleeps=1;//reset the number of sleeps we will do next time
763 if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
764 *request = MPI_REQUEST_NULL;
772 int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
779 *index = MPI_UNDEFINED;
781 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
782 map = xbt_new(int, count);
784 for(i = 0; i < count; i++) {
785 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action &&
786 !(requests[i]->flags & PREPARED)) {
787 xbt_dynar_push(comms, &requests[i]->action);
793 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
794 static int nsleeps = 1;
795 if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
797 i = simcall_comm_testany(comms);
798 // not MPI_UNDEFINED, as this is a simix return code
801 finish_wait(&requests[*index], status);
802 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
803 requests[*index] = MPI_REQUEST_NULL;
810 //all requests are null or inactive, return true
812 smpi_empty_status(status);
815 xbt_dynar_free(&comms);
821 int smpi_mpi_testall(int count, MPI_Request requests[],
825 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
828 for(i=0; i<count; i++){
829 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
830 if (smpi_mpi_test(&requests[i], pstat)!=1){
833 requests[i]=MPI_REQUEST_NULL;
836 smpi_empty_status(pstat);
838 if(status != MPI_STATUSES_IGNORE) {
845 void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
847 //FIXME find another wait to avoid busy waiting ?
848 // the issue here is that we have to wait on a nonexistent comm
850 smpi_mpi_iprobe(source, tag, comm, &flag, status);
851 XBT_DEBUG("Busy Waiting on probing : %d", flag);
855 void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
857 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,
858 comm, PERSISTENT | RECV);
860 //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
861 //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
862 static int nsleeps = 1;
863 if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
864 // behave like a receive, but don't do it
867 print_request("New iprobe", request);
868 // We have to test both mailboxes as we don't know if we will receive one one or another
869 if (sg_cfg_get_int("smpi/async_small_thres")>0){
870 mailbox = smpi_process_mailbox_small();
871 XBT_DEBUG("trying to probe the perm recv mailbox");
872 request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
874 if (request->action==NULL){
875 mailbox = smpi_process_mailbox();
876 XBT_DEBUG("trying to probe the other mailbox");
877 request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
881 MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
883 if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
884 status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
885 status->MPI_TAG = req->tag;
886 status->MPI_ERROR = MPI_SUCCESS;
887 status->count = req->real_size;
889 nsleeps=1;//reset the number of sleeps we will do next time
895 smpi_mpi_request_free(&request);
900 void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
902 print_request("Waiting", *request);
903 if ((*request)->flags & PREPARED) {
904 smpi_empty_status(status);
908 if ((*request)->action != NULL) { // this is not a detached send
909 simcall_comm_wait((*request)->action, -1.0);
911 if((MC_is_active() || MC_record_replay_is_active()) && (*request)->action)
912 (*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
915 finish_wait(request, status);
916 if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT))
917 *request = MPI_REQUEST_NULL;
918 // FIXME for a detached send, finish_wait is not called:
921 int smpi_mpi_waitany(int count, MPI_Request requests[],
928 index = MPI_UNDEFINED;
930 // Wait for a request to complete
931 comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
932 map = xbt_new(int, count);
934 XBT_DEBUG("Wait for one of %d", count);
935 for(i = 0; i < count; i++) {
936 if (requests[i] != MPI_REQUEST_NULL
937 && !(requests[i]->flags & PREPARED)
938 && !(requests[i]->flags & FINISHED)) {
939 if (requests[i]->action != NULL) {
940 XBT_DEBUG("Waiting any %p ", requests[i]);
941 xbt_dynar_push(comms, &requests[i]->action);
945 //This is a finished detached request, let's return this one
946 size=0;//so we free the dynar but don't do the waitany call
948 finish_wait(&requests[i], status);//cleanup if refcount = 0
949 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
950 requests[i]=MPI_REQUEST_NULL;//set to null
956 i = simcall_comm_waitany(comms);
958 // not MPI_UNDEFINED, as this is a simix return code
961 finish_wait(&requests[index], status);
962 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
963 requests[index] = MPI_REQUEST_NULL;
967 xbt_dynar_free(&comms);
970 if (index==MPI_UNDEFINED)
971 smpi_empty_status(status);
976 int smpi_mpi_waitall(int count, MPI_Request requests[],
981 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
982 int retvalue = MPI_SUCCESS;
983 //tag invalid requests in the set
984 if (status != MPI_STATUSES_IGNORE) {
985 for (c = 0; c < count; c++) {
986 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ||
987 (requests[c]->flags & PREPARED)) {
988 smpi_empty_status(&status[c]);
989 } else if (requests[c]->src == MPI_PROC_NULL) {
990 smpi_empty_status(&status[c]);
991 status[c].MPI_SOURCE = MPI_PROC_NULL;
995 for(c = 0; c < count; c++) {
997 if (MC_is_active() || MC_record_replay_is_active()) {
998 smpi_mpi_wait(&requests[c], pstat);
1001 index = smpi_mpi_waitany(count, requests, pstat);
1002 if (index == MPI_UNDEFINED)
1004 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1005 requests[index]=MPI_REQUEST_NULL;
1007 if (status != MPI_STATUSES_IGNORE) {
1008 status[index] = *pstat;
1009 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1010 retvalue = MPI_ERR_IN_STATUS;
1017 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
1018 MPI_Status status[])
1020 int i, count, index;
1022 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1025 for(i = 0; i < incount; i++)
1027 index=smpi_mpi_waitany(incount, requests, pstat);
1028 if(index!=MPI_UNDEFINED){
1029 indices[count] = index;
1031 if(status != MPI_STATUSES_IGNORE) {
1032 status[index] = *pstat;
1034 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
1035 requests[index]=MPI_REQUEST_NULL;
1037 return MPI_UNDEFINED;
1043 int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
1044 MPI_Status status[])
1046 int i, count, count_dead;
1048 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1052 for(i = 0; i < incount; i++) {
1053 if((requests[i] != MPI_REQUEST_NULL)) {
1054 if(smpi_mpi_test(&requests[i], pstat)) {
1057 if(status != MPI_STATUSES_IGNORE) {
1060 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
1061 requests[i]=MPI_REQUEST_NULL;
1067 if(count_dead==incount)return MPI_UNDEFINED;
1071 void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
1074 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1075 nary_tree_bcast(buf, count, datatype, root, comm, 4);
1078 void smpi_mpi_barrier(MPI_Comm comm)
1080 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
1081 nary_tree_barrier(comm, 4);
1084 void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1085 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1086 int root, MPI_Comm comm)
1088 int system_tag = COLL_TAG_GATHER;
1089 int rank, size, src, index;
1090 MPI_Aint lb = 0, recvext = 0;
1091 MPI_Request *requests;
1093 rank = smpi_comm_rank(comm);
1094 size = smpi_comm_size(comm);
1096 // Send buffer to root
1097 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1099 // FIXME: check for errors
1100 smpi_datatype_extent(recvtype, &lb, &recvext);
1101 // Local copy from root
1102 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1103 (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
1104 // Receive buffers from senders
1105 requests = xbt_new(MPI_Request, size - 1);
1107 for(src = 0; src < size; src++) {
1109 requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
1110 recvcount, recvtype,
1111 src, system_tag, comm);
1115 // Wait for completion of irecv's.
1116 smpi_mpi_startall(size - 1, requests);
1117 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1118 for(src = 0; src < size-1; src++) {
1119 smpi_mpi_request_free(&requests[src]);
1126 void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
1127 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1131 int rank = smpi_process_index();
1134 /* arbitrarily choose root as rank 0 */
1135 size = smpi_comm_size(comm);
1137 displs = xbt_new(int, size);
1138 for (i = 0; i < size; i++) {
1140 count += recvcounts[i];
1142 tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1144 mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
1145 smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
1146 recvcounts[rank], datatype, 0, comm);
1148 smpi_free_tmp_buffer(tmpbuf);
1151 void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1152 void *recvbuf, int *recvcounts, int *displs,
1153 MPI_Datatype recvtype, int root, MPI_Comm comm)
1155 int system_tag = COLL_TAG_GATHERV;
1156 int rank, size, src, index;
1157 MPI_Aint lb = 0, recvext = 0;
1158 MPI_Request *requests;
1160 rank = smpi_comm_rank(comm);
1161 size = smpi_comm_size(comm);
1163 // Send buffer to root
1164 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
1166 // FIXME: check for errors
1167 smpi_datatype_extent(recvtype, &lb, &recvext);
1168 // Local copy from root
1169 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1170 (char *)recvbuf + displs[root] * recvext,
1171 recvcounts[root], recvtype);
1172 // Receive buffers from senders
1173 requests = xbt_new(MPI_Request, size - 1);
1175 for(src = 0; src < size; src++) {
1178 smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
1179 recvcounts[src], recvtype, src, system_tag, comm);
1183 // Wait for completion of irecv's.
1184 smpi_mpi_startall(size - 1, requests);
1185 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1186 for(src = 0; src < size-1; src++) {
1187 smpi_mpi_request_free(&requests[src]);
1193 void smpi_mpi_allgather(void *sendbuf, int sendcount,
1194 MPI_Datatype sendtype, void *recvbuf,
1195 int recvcount, MPI_Datatype recvtype,
1198 int system_tag = COLL_TAG_ALLGATHER;
1199 int rank, size, other, index;
1200 MPI_Aint lb = 0, recvext = 0;
1201 MPI_Request *requests;
1203 rank = smpi_comm_rank(comm);
1204 size = smpi_comm_size(comm);
1205 // FIXME: check for errors
1206 smpi_datatype_extent(recvtype, &lb, &recvext);
1207 // Local copy from self
1208 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1209 (char *)recvbuf + rank * recvcount * recvext, recvcount,
1211 // Send/Recv buffers to/from others;
1212 requests = xbt_new(MPI_Request, 2 * (size - 1));
1214 for(other = 0; other < size; other++) {
1217 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1220 requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
1221 recvcount, recvtype, other,
1226 // Wait for completion of all comms.
1227 smpi_mpi_startall(2 * (size - 1), requests);
1228 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1229 for(other = 0; other < 2*(size-1); other++) {
1230 smpi_mpi_request_free(&requests[other]);
1235 void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
1236 MPI_Datatype sendtype, void *recvbuf,
1237 int *recvcounts, int *displs,
1238 MPI_Datatype recvtype, MPI_Comm comm)
1240 int system_tag = COLL_TAG_ALLGATHERV;
1241 int rank, size, other, index;
1242 MPI_Aint lb = 0, recvext = 0;
1243 MPI_Request *requests;
1245 rank = smpi_comm_rank(comm);
1246 size = smpi_comm_size(comm);
1247 // FIXME: check for errors
1248 smpi_datatype_extent(recvtype, &lb, &recvext);
1249 // Local copy from self
1250 smpi_datatype_copy(sendbuf, sendcount, sendtype,
1251 (char *)recvbuf + displs[rank] * recvext,
1252 recvcounts[rank], recvtype);
1253 // Send buffers to others;
1254 requests = xbt_new(MPI_Request, 2 * (size - 1));
1256 for(other = 0; other < size; other++) {
1259 smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
1263 smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
1264 recvtype, other, system_tag, comm);
1268 // Wait for completion of all comms.
1269 smpi_mpi_startall(2 * (size - 1), requests);
1270 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
1271 for(other = 0; other < 2*(size-1); other++) {
1272 smpi_mpi_request_free(&requests[other]);
1277 void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
1278 void *recvbuf, int recvcount, MPI_Datatype recvtype,
1279 int root, MPI_Comm comm)
1281 int system_tag = COLL_TAG_SCATTER;
1282 int rank, size, dst, index;
1283 MPI_Aint lb = 0, sendext = 0;
1284 MPI_Request *requests;
1286 rank = smpi_comm_rank(comm);
1287 size = smpi_comm_size(comm);
1289 // Recv buffer from root
1290 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1293 // FIXME: check for errors
1294 smpi_datatype_extent(sendtype, &lb, &sendext);
1295 // Local copy from root
1296 if(recvbuf!=MPI_IN_PLACE){
1297 smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
1298 sendcount, sendtype, recvbuf, recvcount, recvtype);
1300 // Send buffers to receivers
1301 requests = xbt_new(MPI_Request, size - 1);
1303 for(dst = 0; dst < size; dst++) {
1305 requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
1306 sendcount, sendtype, dst,
1311 // Wait for completion of isend's.
1312 smpi_mpi_startall(size - 1, requests);
1313 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1314 for(dst = 0; dst < size-1; dst++) {
1315 smpi_mpi_request_free(&requests[dst]);
1321 void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
1322 MPI_Datatype sendtype, void *recvbuf, int recvcount,
1323 MPI_Datatype recvtype, int root, MPI_Comm comm)
1325 int system_tag = COLL_TAG_SCATTERV;
1326 int rank, size, dst, index;
1327 MPI_Aint lb = 0, sendext = 0;
1328 MPI_Request *requests;
1330 rank = smpi_comm_rank(comm);
1331 size = smpi_comm_size(comm);
1333 // Recv buffer from root
1334 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
1337 // FIXME: check for errors
1338 smpi_datatype_extent(sendtype, &lb, &sendext);
1339 // Local copy from root
1340 if(recvbuf!=MPI_IN_PLACE){
1341 smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
1342 sendtype, recvbuf, recvcount, recvtype);
1344 // Send buffers to receivers
1345 requests = xbt_new(MPI_Request, size - 1);
1347 for(dst = 0; dst < size; dst++) {
1350 smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
1351 sendtype, dst, system_tag, comm);
1355 // Wait for completion of isend's.
1356 smpi_mpi_startall(size - 1, requests);
1357 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
1358 for(dst = 0; dst < size-1; dst++) {
1359 smpi_mpi_request_free(&requests[dst]);
1365 void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
1366 MPI_Datatype datatype, MPI_Op op, int root,
1369 int system_tag = COLL_TAG_REDUCE;
1370 int rank, size, src, index;
1371 MPI_Aint lb = 0, dataext = 0;
1372 MPI_Request *requests;
1376 char* sendtmpbuf = (char*) sendbuf;
1377 if( sendbuf == MPI_IN_PLACE ) {
1378 sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype));
1379 smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
1382 rank = smpi_comm_rank(comm);
1383 size = smpi_comm_size(comm);
1384 //non commutative case, use a working algo from openmpi
1385 if(!smpi_op_is_commute(op)){
1386 smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count,
1387 datatype, op, root, comm);
1392 // Send buffer to root
1393 smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm);
1395 // FIXME: check for errors
1396 smpi_datatype_extent(datatype, &lb, &dataext);
1397 // Local copy from root
1398 if (sendtmpbuf && recvbuf)
1399 smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
1400 // Receive buffers from senders
1401 //TODO: make a MPI_barrier here ?
1402 requests = xbt_new(MPI_Request, size - 1);
1403 tmpbufs = xbt_new(void *, size - 1);
1405 for(src = 0; src < size; src++) {
1407 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1409 if (!smpi_process_get_replaying())
1410 tmpbufs[index] = xbt_malloc(count * dataext);
1412 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1414 smpi_irecv_init(tmpbufs[index], count, datatype, src,
1419 // Wait for completion of irecv's.
1420 smpi_mpi_startall(size - 1, requests);
1421 for(src = 0; src < size - 1; src++) {
1422 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1423 XBT_DEBUG("finished waiting any request with index %d", index);
1424 if(index == MPI_UNDEFINED) {
1427 smpi_mpi_request_free(&requests[index]);
1429 if(op) /* op can be MPI_OP_NULL that does nothing */
1430 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1432 for(index = 0; index < size - 1; index++) {
1433 smpi_free_tmp_buffer(tmpbufs[index]);
1438 if( sendbuf == MPI_IN_PLACE ) {
1439 smpi_free_tmp_buffer(sendtmpbuf);
1444 void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
1445 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1447 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
1448 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
1451 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
1452 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1454 int system_tag = -888;
1455 int rank, size, other, index;
1456 MPI_Aint lb = 0, dataext = 0;
1457 MPI_Request *requests;
1460 rank = smpi_comm_rank(comm);
1461 size = smpi_comm_size(comm);
1463 // FIXME: check for errors
1464 smpi_datatype_extent(datatype, &lb, &dataext);
1466 // Local copy from self
1467 smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
1469 // Send/Recv buffers to/from others;
1470 requests = xbt_new(MPI_Request, size - 1);
1471 tmpbufs = xbt_new(void *, rank);
1473 for(other = 0; other < rank; other++) {
1474 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1476 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1478 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1482 for(other = rank + 1; other < size; other++) {
1484 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1487 // Wait for completion of all comms.
1488 smpi_mpi_startall(size - 1, requests);
1490 if(smpi_op_is_commute(op)){
1491 for(other = 0; other < size - 1; other++) {
1492 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1493 if(index == MPI_UNDEFINED) {
1497 // #Request is below rank: it's a irecv
1498 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1502 //non commutative case, wait in order
1503 for(other = 0; other < size - 1; other++) {
1504 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1506 smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1510 for(index = 0; index < rank; index++) {
1511 smpi_free_tmp_buffer(tmpbufs[index]);
1513 for(index = 0; index < size-1; index++) {
1514 smpi_mpi_request_free(&requests[index]);
1520 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
1521 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
1523 int system_tag = -888;
1524 int rank, size, other, index;
1525 MPI_Aint lb = 0, dataext = 0;
1526 MPI_Request *requests;
1528 int recvbuf_is_empty=1;
1529 rank = smpi_comm_rank(comm);
1530 size = smpi_comm_size(comm);
1532 // FIXME: check for errors
1533 smpi_datatype_extent(datatype, &lb, &dataext);
1535 // Send/Recv buffers to/from others;
1536 requests = xbt_new(MPI_Request, size - 1);
1537 tmpbufs = xbt_new(void *, rank);
1539 for(other = 0; other < rank; other++) {
1540 // FIXME: possibly overkill we we have contiguous/noncontiguous data
1542 tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
1544 smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
1548 for(other = rank + 1; other < size; other++) {
1550 smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
1553 // Wait for completion of all comms.
1554 smpi_mpi_startall(size - 1, requests);
1555 if(smpi_op_is_commute(op)){
1556 for(other = 0; other < size - 1; other++) {
1557 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
1558 if(index == MPI_UNDEFINED) {
1562 if(recvbuf_is_empty){
1563 smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
1566 // #Request is below rank: it's a irecv
1567 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
1571 //non commutative case, wait in order
1572 for(other = 0; other < size - 1; other++) {
1573 smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
1575 if(recvbuf_is_empty){
1576 smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
1578 }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
1582 for(index = 0; index < rank; index++) {
1583 smpi_free_tmp_buffer(tmpbufs[index]);
1585 for(index = 0; index < size-1; index++) {
1586 smpi_mpi_request_free(&requests[index]);