1 /* Copyright (c) 2007-2015. 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 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)");
12 msg_bar_t creation_bar = nullptr;
14 typedef struct s_smpi_mpi_win{
21 std::vector<MPI_Request> *requests;
23 MPI_Win* connected_wins;
27 int count; //for ordering the accs
31 MPI_Win smpi_mpi_win_create( void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm){
34 int comm_size = smpi_comm_size(comm);
35 int rank=smpi_comm_rank(comm);
36 XBT_DEBUG("Creating window");
38 win = xbt_new(s_smpi_mpi_win_t, 1);
41 win->disp_unit = disp_unit;
44 if(info!=MPI_INFO_NULL)
49 win->group = MPI_GROUP_NULL;
50 win->requests = new std::vector<MPI_Request>();
51 win->connected_wins = xbt_new0(MPI_Win, comm_size);
52 win->connected_wins[rank] = win;
55 win->bar = MSG_barrier_init(comm_size);
57 mpi_coll_allgather_fun(&(win->connected_wins[rank]), sizeof(MPI_Win), MPI_BYTE, win->connected_wins, sizeof(MPI_Win),
60 mpi_coll_bcast_fun(&(win->bar), sizeof(msg_bar_t), MPI_BYTE, 0, comm);
62 mpi_coll_barrier_fun(comm);
67 int smpi_mpi_win_free( MPI_Win* win){
68 //As per the standard, perform a barrier to ensure every async comm is finished
69 MSG_barrier_wait((*win)->bar);
70 delete (*win)->requests;
71 xbt_free((*win)->connected_wins);
72 if ((*win)->name != nullptr){
73 xbt_free((*win)->name);
75 if((*win)->info!=MPI_INFO_NULL){
76 MPI_Info_free(&(*win)->info);
79 mpi_coll_barrier_fun((*win)->comm);
80 int rank=smpi_comm_rank((*win)->comm);
82 MSG_barrier_destroy((*win)->bar);
88 void smpi_mpi_win_get_name(MPI_Win win, char* name, int* length){
89 if(win->name==nullptr){
94 *length = strlen(win->name);
95 strncpy(name, win->name, *length+1);
98 void smpi_mpi_win_get_group(MPI_Win win, MPI_Group* group){
99 if(win->comm != MPI_COMM_NULL){
100 *group = smpi_comm_group(win->comm);
102 *group = MPI_GROUP_NULL;
106 void smpi_mpi_win_set_name(MPI_Win win, char* name){
107 win->name = xbt_strdup(name);
110 int smpi_mpi_win_fence( int assert, MPI_Win win){
111 XBT_DEBUG("Entering fence");
114 if(assert != MPI_MODE_NOPRECEDE){
115 MSG_barrier_wait(win->bar);
117 std::vector<MPI_Request> *reqs = win->requests;
118 int size = static_cast<int>(reqs->size());
119 // start all requests that have been prepared by another process
121 for(auto req: *reqs){
122 if (req && (req->flags & PREPARED))
126 MPI_Request* treqs = &(*reqs)[0];
128 smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
132 win->assert = assert;
134 MSG_barrier_wait(win->bar);
135 XBT_DEBUG("Leaving fence ");
140 int smpi_mpi_put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
141 MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Win win)
143 if(win->opened==0)//check that post/start has been done
145 //get receiver pointer
146 MPI_Win recv_win = win->connected_wins[target_rank];
148 void* recv_addr = static_cast<void*> ( static_cast<char*>(recv_win->base) + target_disp * recv_win->disp_unit);
149 XBT_DEBUG("Entering MPI_Put to %d", target_rank);
151 if(target_rank != smpi_comm_rank(win->comm)){
152 //prepare send_request
153 MPI_Request sreq = smpi_rma_send_init(origin_addr, origin_count, origin_datatype, smpi_process_index(),
154 smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+1, win->comm, MPI_OP_NULL);
156 //prepare receiver request
157 MPI_Request rreq = smpi_rma_recv_init(recv_addr, target_count, target_datatype, smpi_process_index(),
158 smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+1, recv_win->comm, MPI_OP_NULL);
160 //push request to receiver's win
161 recv_win->requests->push_back(rreq);
164 smpi_mpi_start(sreq);
166 //push request to sender's win
167 win->requests->push_back(sreq);
169 smpi_datatype_copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype);
175 int smpi_mpi_get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
176 MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Win win)
178 if(win->opened==0)//check that post/start has been done
181 MPI_Win send_win = win->connected_wins[target_rank];
183 void* send_addr = static_cast<void*>(static_cast<char*>(send_win->base) + target_disp * send_win->disp_unit);
184 XBT_DEBUG("Entering MPI_Get from %d", target_rank);
186 if(target_rank != smpi_comm_rank(win->comm)){
187 //prepare send_request
188 MPI_Request sreq = smpi_rma_send_init(send_addr, target_count, target_datatype,
189 smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), SMPI_RMA_TAG+2, send_win->comm,
192 //prepare receiver request
193 MPI_Request rreq = smpi_rma_recv_init(origin_addr, origin_count, origin_datatype,
194 smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), SMPI_RMA_TAG+2, win->comm,
197 //start the send, with another process than us as sender.
198 smpi_mpi_start(sreq);
200 //push request to receiver's win
201 send_win->requests->push_back(sreq);
204 smpi_mpi_start(rreq);
206 //push request to sender's win
207 win->requests->push_back(rreq);
209 smpi_datatype_copy(send_addr, target_count, target_datatype, origin_addr, origin_count, origin_datatype);
216 int smpi_mpi_accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
217 MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Op op, MPI_Win win)
219 if(win->opened==0)//check that post/start has been done
221 //FIXME: local version
222 //get receiver pointer
223 MPI_Win recv_win = win->connected_wins[target_rank];
225 void* recv_addr = static_cast<void*>(static_cast<char*>(recv_win->base) + target_disp * recv_win->disp_unit);
226 XBT_DEBUG("Entering MPI_Accumulate to %d", target_rank);
227 //As the tag will be used for ordering of the operations, add count to it
228 //prepare send_request
229 MPI_Request sreq = smpi_rma_send_init(origin_addr, origin_count, origin_datatype,
230 smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+3+win->count, win->comm, op);
232 //prepare receiver request
233 MPI_Request rreq = smpi_rma_recv_init(recv_addr, target_count, target_datatype,
234 smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+3+win->count, recv_win->comm, op);
237 //push request to receiver's win
238 recv_win->requests->push_back(rreq);
240 smpi_mpi_start(sreq);
242 //push request to sender's win
243 win->requests->push_back(sreq);
248 int smpi_mpi_win_start(MPI_Group group, int assert, MPI_Win win){
249 /* From MPI forum advices
250 The call to MPI_WIN_COMPLETE does not return until the put call has completed at the origin; and the target window
251 will be accessed by the put operation only after the call to MPI_WIN_START has matched a call to MPI_WIN_POST by
252 the target process. This still leaves much choice to implementors. The call to MPI_WIN_START can block until the
253 matching call to MPI_WIN_POST occurs at all target processes. One can also have implementations where the call to
254 MPI_WIN_START is nonblocking, but the call to MPI_PUT blocks until the matching call to MPI_WIN_POST occurred; or
255 implementations where the first two calls are nonblocking, but the call to MPI_WIN_COMPLETE blocks until the call
256 to MPI_WIN_POST occurred; or even implementations where all three calls can complete before any target process
257 called MPI_WIN_POST --- the data put must be buffered, in this last case, so as to allow the put to complete at the
258 origin ahead of its completion at the target. However, once the call to MPI_WIN_POST is issued, the sequence above
259 must complete, without further dependencies. */
261 //naive, blocking implementation.
263 int size = smpi_group_size(group);
264 MPI_Request* reqs = xbt_new0(MPI_Request, size);
267 int src=smpi_group_index(group,j);
268 if(src!=smpi_process_index()){
269 reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+4, MPI_COMM_WORLD);
275 smpi_mpi_startall(size, reqs);
276 smpi_mpi_waitall(size, reqs, MPI_STATUSES_IGNORE);
278 smpi_mpi_request_free(&reqs[i]);
281 win->opened++; //we're open for business !
283 smpi_group_use(group);
287 int smpi_mpi_win_post(MPI_Group group, int assert, MPI_Win win){
288 //let's make a synchronous send here
290 int size = smpi_group_size(group);
291 MPI_Request* reqs = xbt_new0(MPI_Request, size);
294 int dst=smpi_group_index(group,j);
295 if(dst!=smpi_process_index()){
296 reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+4, MPI_COMM_WORLD);
303 smpi_mpi_startall(size, reqs);
304 smpi_mpi_waitall(size, reqs, MPI_STATUSES_IGNORE);
306 smpi_mpi_request_free(&reqs[i]);
309 win->opened++; //we're open for business !
311 smpi_group_use(group);
315 int smpi_mpi_win_complete(MPI_Win win){
317 xbt_die("Complete called on already opened MPI_Win");
319 XBT_DEBUG("Entering MPI_Win_Complete");
321 int size = smpi_group_size(win->group);
322 MPI_Request* reqs = xbt_new0(MPI_Request, size);
325 int dst=smpi_group_index(win->group,j);
326 if(dst!=smpi_process_index()){
327 reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+5, MPI_COMM_WORLD);
333 XBT_DEBUG("Win_complete - Sending sync messages to %d processes", size);
334 smpi_mpi_startall(size, reqs);
335 smpi_mpi_waitall(size, reqs, MPI_STATUSES_IGNORE);
338 smpi_mpi_request_free(&reqs[i]);
342 //now we can finish RMA calls
344 std::vector<MPI_Request> *reqqs = win->requests;
345 size = static_cast<int>(reqqs->size());
347 XBT_DEBUG("Win_complete - Finishing %d RMA calls", size);
348 // start all requests that have been prepared by another process
349 for (auto req: *reqqs){
350 if (req && (req->flags & PREPARED))
354 MPI_Request* treqs = &(*reqqs)[0];
355 smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
357 smpi_group_unuse(win->group);
358 win->opened--; //we're closed for business !
362 int smpi_mpi_win_wait(MPI_Win win){
363 //naive, blocking implementation.
364 XBT_DEBUG("Entering MPI_Win_Wait");
366 int size = smpi_group_size(win->group);
367 MPI_Request* reqs = xbt_new0(MPI_Request, size);
370 int src=smpi_group_index(win->group,j);
371 if(src!=smpi_process_index()){
372 reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+5, MPI_COMM_WORLD);
378 XBT_DEBUG("Win_wait - Receiving sync messages from %d processes", size);
379 smpi_mpi_startall(size, reqs);
380 smpi_mpi_waitall(size, reqs, MPI_STATUSES_IGNORE);
382 smpi_mpi_request_free(&reqs[i]);
386 std::vector<MPI_Request> *reqqs = win->requests;
387 size = static_cast<int>(reqqs->size());
389 XBT_DEBUG("Win_complete - Finishing %d RMA calls", size);
391 // start all requests that have been prepared by another process
392 for(auto req: *reqqs){
393 if (req && (req->flags & PREPARED))
397 MPI_Request* treqs = &(*reqqs)[0];
398 smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
400 smpi_group_unuse(win->group);
401 win->opened--; //we're opened for business !