1 | #include "ep_lib.hpp" |
---|
2 | #include <mpi.h> |
---|
3 | #include "ep_declaration.hpp" |
---|
4 | |
---|
5 | using namespace std; |
---|
6 | |
---|
7 | |
---|
8 | namespace ep_lib |
---|
9 | { |
---|
10 | int MPI_Intercomm_create_kernel(MPI_Comm local_comm, int local_leader, MPI_Comm peer_comm, int remote_leader, int tag, MPI_Comm *newintercomm) |
---|
11 | { |
---|
12 | |
---|
13 | int ep_rank, ep_rank_loc, mpi_rank; |
---|
14 | int ep_size, num_ep, mpi_size; |
---|
15 | |
---|
16 | ep_rank = local_comm.ep_comm_ptr->size_rank_info[0].first; |
---|
17 | ep_rank_loc = local_comm.ep_comm_ptr->size_rank_info[1].first; |
---|
18 | mpi_rank = local_comm.ep_comm_ptr->size_rank_info[2].first; |
---|
19 | ep_size = local_comm.ep_comm_ptr->size_rank_info[0].second; |
---|
20 | num_ep = local_comm.ep_comm_ptr->size_rank_info[1].second; |
---|
21 | mpi_size = local_comm.ep_comm_ptr->size_rank_info[2].second; |
---|
22 | |
---|
23 | std::vector<int> rank_info[4]; //! 0->rank_in_world of local_comm, 1->rank_in_local_parent of local_comm |
---|
24 | //! 2->rank_in_world of remote_comm, 3->rank_in_local_parent of remote_comm |
---|
25 | |
---|
26 | int rank_in_world; |
---|
27 | int rank_in_local_parent; |
---|
28 | |
---|
29 | int rank_in_peer_mpi[2]; |
---|
30 | |
---|
31 | int local_ep_size = ep_size; |
---|
32 | int remote_ep_size; |
---|
33 | |
---|
34 | |
---|
35 | ::MPI_Comm local_mpi_comm = static_cast< ::MPI_Comm>(local_comm.mpi_comm); |
---|
36 | |
---|
37 | |
---|
38 | ::MPI_Comm_rank(MPI_COMM_WORLD_STD, &rank_in_world); |
---|
39 | ::MPI_Comm_rank(static_cast< ::MPI_Comm>(local_comm.mpi_comm), &rank_in_local_parent); |
---|
40 | |
---|
41 | |
---|
42 | bool is_proc_master = false; |
---|
43 | bool is_local_leader = false; |
---|
44 | bool is_final_master = false; |
---|
45 | |
---|
46 | |
---|
47 | if(ep_rank == local_leader) { is_proc_master = true; is_local_leader = true; is_final_master = true;} |
---|
48 | if(ep_rank_loc == 0 && mpi_rank != local_comm.rank_map->at(local_leader).second) is_proc_master = true; |
---|
49 | |
---|
50 | |
---|
51 | int size_info[4]; //! used for choose size of rank_info 0-> mpi_size of local_comm, 1-> mpi_size of remote_comm |
---|
52 | |
---|
53 | int leader_info[4]; //! 0->world rank of local_leader, 1->world rank of remote leader |
---|
54 | |
---|
55 | |
---|
56 | std::vector<int> ep_info[2]; //! 0-> num_ep in local_comm, 1->num_ep in remote_comm |
---|
57 | |
---|
58 | std::vector<int> new_rank_info[4]; |
---|
59 | std::vector<int> new_ep_info[2]; |
---|
60 | |
---|
61 | std::vector<int> offset; |
---|
62 | |
---|
63 | if(is_proc_master) |
---|
64 | { |
---|
65 | |
---|
66 | size_info[0] = mpi_size; |
---|
67 | |
---|
68 | rank_info[0].resize(size_info[0]); |
---|
69 | rank_info[1].resize(size_info[0]); |
---|
70 | |
---|
71 | |
---|
72 | |
---|
73 | ep_info[0].resize(size_info[0]); |
---|
74 | |
---|
75 | vector<int> send_buf(6); |
---|
76 | vector<int> recv_buf(3*size_info[0]); |
---|
77 | |
---|
78 | send_buf[0] = rank_in_world; |
---|
79 | send_buf[1] = rank_in_local_parent; |
---|
80 | send_buf[2] = num_ep; |
---|
81 | |
---|
82 | ::MPI_Allgather(send_buf.data(), 3, MPI_INT_STD, recv_buf.data(), 3, MPI_INT_STD, local_mpi_comm); |
---|
83 | |
---|
84 | for(int i=0; i<size_info[0]; i++) |
---|
85 | { |
---|
86 | rank_info[0][i] = recv_buf[3*i]; |
---|
87 | rank_info[1][i] = recv_buf[3*i+1]; |
---|
88 | ep_info[0][i] = recv_buf[3*i+2]; |
---|
89 | } |
---|
90 | |
---|
91 | if(is_local_leader) |
---|
92 | { |
---|
93 | leader_info[0] = rank_in_world; |
---|
94 | leader_info[1] = remote_leader; |
---|
95 | |
---|
96 | ::MPI_Comm_rank(static_cast< ::MPI_Comm>(peer_comm.mpi_comm), &rank_in_peer_mpi[0]); |
---|
97 | |
---|
98 | |
---|
99 | |
---|
100 | send_buf[0] = size_info[0]; |
---|
101 | send_buf[1] = local_ep_size; |
---|
102 | send_buf[2] = rank_in_peer_mpi[0]; |
---|
103 | |
---|
104 | MPI_Request req_send, req_recv; |
---|
105 | MPI_Status sta_send, sta_recv; |
---|
106 | |
---|
107 | MPI_Isend(send_buf.data(), 3, MPI_INT_STD, remote_leader, tag, peer_comm, &req_send); |
---|
108 | MPI_Irecv(recv_buf.data(), 3, MPI_INT_STD, remote_leader, tag, peer_comm, &req_recv); |
---|
109 | |
---|
110 | |
---|
111 | MPI_Wait(&req_send, &sta_send); |
---|
112 | MPI_Wait(&req_recv, &sta_recv); |
---|
113 | |
---|
114 | size_info[1] = recv_buf[0]; |
---|
115 | remote_ep_size = recv_buf[1]; |
---|
116 | rank_in_peer_mpi[1] = recv_buf[2]; |
---|
117 | |
---|
118 | } |
---|
119 | |
---|
120 | send_buf[0] = size_info[1]; |
---|
121 | send_buf[1] = leader_info[0]; |
---|
122 | send_buf[2] = leader_info[1]; |
---|
123 | send_buf[3] = rank_in_peer_mpi[0]; |
---|
124 | send_buf[4] = rank_in_peer_mpi[1]; |
---|
125 | |
---|
126 | ::MPI_Bcast(send_buf.data(), 5, MPI_INT_STD, local_comm.rank_map->at(local_leader).second, local_mpi_comm); |
---|
127 | |
---|
128 | size_info[1] = send_buf[0]; |
---|
129 | leader_info[0] = send_buf[1]; |
---|
130 | leader_info[1] = send_buf[2]; |
---|
131 | rank_in_peer_mpi[0] = send_buf[3]; |
---|
132 | rank_in_peer_mpi[1] = send_buf[4]; |
---|
133 | |
---|
134 | |
---|
135 | rank_info[2].resize(size_info[1]); |
---|
136 | rank_info[3].resize(size_info[1]); |
---|
137 | |
---|
138 | ep_info[1].resize(size_info[1]); |
---|
139 | |
---|
140 | send_buf.resize(3*size_info[0]); |
---|
141 | recv_buf.resize(3*size_info[1]); |
---|
142 | |
---|
143 | if(is_local_leader) |
---|
144 | { |
---|
145 | MPI_Status status; |
---|
146 | |
---|
147 | std::copy ( rank_info[0].data(), rank_info[0].data() + size_info[0], send_buf.begin() ); |
---|
148 | std::copy ( rank_info[1].data(), rank_info[1].data() + size_info[0], send_buf.begin() + size_info[0] ); |
---|
149 | std::copy ( ep_info[0].data(), ep_info[0].data() + size_info[0], send_buf.begin() + 2*size_info[0] ); |
---|
150 | |
---|
151 | MPI_Send(send_buf.data(), 3*size_info[0], MPI_INT_STD, remote_leader, tag+1, peer_comm); |
---|
152 | MPI_Recv(recv_buf.data(), 3*size_info[1], MPI_INT_STD, remote_leader, tag+1, peer_comm, &status); |
---|
153 | } |
---|
154 | |
---|
155 | ::MPI_Bcast(recv_buf.data(), 3*size_info[1], MPI_INT_STD, local_comm.rank_map->at(local_leader).second, local_mpi_comm); |
---|
156 | |
---|
157 | std::copy ( recv_buf.data(), recv_buf.data() + size_info[1], rank_info[2].begin() ); |
---|
158 | std::copy ( recv_buf.data() + size_info[1], recv_buf.data() + 2*size_info[1], rank_info[3].begin() ); |
---|
159 | std::copy ( recv_buf.data() + 2*size_info[1], recv_buf.data() + 3*size_info[1], ep_info[1].begin() ); |
---|
160 | |
---|
161 | |
---|
162 | offset.resize(size_info[0]); |
---|
163 | |
---|
164 | if(leader_info[0]<leader_info[1]) // erase all ranks doubled with remote_comm, except the local leader |
---|
165 | { |
---|
166 | |
---|
167 | bool found = false; |
---|
168 | int ep_local; |
---|
169 | int ep_remote; |
---|
170 | for(int i=0; i<size_info[0]; i++) |
---|
171 | { |
---|
172 | int target = rank_info[0][i]; |
---|
173 | found = false; |
---|
174 | for(int j=0; j<size_info[1]; j++) |
---|
175 | { |
---|
176 | if(target == rank_info[2][j]) |
---|
177 | { |
---|
178 | found = true; |
---|
179 | ep_local = ep_info[0][j]; |
---|
180 | ep_remote = ep_info[1][j]; |
---|
181 | break; |
---|
182 | } |
---|
183 | } |
---|
184 | if(found) |
---|
185 | { |
---|
186 | |
---|
187 | if(target == leader_info[0]) // the leader is doubled in remote |
---|
188 | { |
---|
189 | new_rank_info[0].push_back(target); |
---|
190 | new_rank_info[1].push_back(rank_info[1][i]); |
---|
191 | |
---|
192 | new_ep_info[0].push_back(ep_local + ep_remote); |
---|
193 | offset[i] = 0; |
---|
194 | } |
---|
195 | else |
---|
196 | { |
---|
197 | offset[i] = ep_local; |
---|
198 | } |
---|
199 | } |
---|
200 | else |
---|
201 | { |
---|
202 | new_rank_info[0].push_back(target); |
---|
203 | new_rank_info[1].push_back(rank_info[1][i]); |
---|
204 | |
---|
205 | new_ep_info[0].push_back(ep_info[0][i]); |
---|
206 | |
---|
207 | offset[i] = 0; |
---|
208 | } |
---|
209 | |
---|
210 | } |
---|
211 | } |
---|
212 | |
---|
213 | else // erase rank doubled with remote leader |
---|
214 | { |
---|
215 | |
---|
216 | bool found = false; |
---|
217 | int ep_local; |
---|
218 | int ep_remote; |
---|
219 | for(int i=0; i<size_info[0]; i++) |
---|
220 | { |
---|
221 | int target = rank_info[0][i]; |
---|
222 | found = false; |
---|
223 | for(int j=0; j<size_info[1]; j++) |
---|
224 | { |
---|
225 | |
---|
226 | if(target == rank_info[2][j]) |
---|
227 | { |
---|
228 | found = true; |
---|
229 | ep_local = ep_info[0][j]; |
---|
230 | ep_remote = ep_info[1][j]; |
---|
231 | break; |
---|
232 | } |
---|
233 | } |
---|
234 | if(found) |
---|
235 | { |
---|
236 | if(target != leader_info[1]) |
---|
237 | { |
---|
238 | new_rank_info[0].push_back(target); |
---|
239 | new_rank_info[1].push_back(rank_info[1][i]); |
---|
240 | |
---|
241 | new_ep_info[0].push_back(ep_local + ep_remote); |
---|
242 | offset[i] = 0; |
---|
243 | } |
---|
244 | else // found remote leader |
---|
245 | { |
---|
246 | offset[i] = ep_remote; |
---|
247 | } |
---|
248 | } |
---|
249 | else |
---|
250 | { |
---|
251 | new_rank_info[0].push_back(target); |
---|
252 | new_rank_info[1].push_back(rank_info[1][i]); |
---|
253 | |
---|
254 | new_ep_info[0].push_back(ep_info[0][i]); |
---|
255 | offset[i] = 0; |
---|
256 | } |
---|
257 | } |
---|
258 | } |
---|
259 | |
---|
260 | if(offset[mpi_rank] == 0) |
---|
261 | { |
---|
262 | is_final_master = true; |
---|
263 | } |
---|
264 | |
---|
265 | |
---|
266 | //! size_info[4]: 2->size of new_ep_info for local, 3->size of new_ep_info for remote |
---|
267 | |
---|
268 | if(is_local_leader) |
---|
269 | { |
---|
270 | size_info[2] = new_ep_info[0].size(); |
---|
271 | MPI_Status status; |
---|
272 | MPI_Send(&size_info[2], 1, MPI_INT_STD, remote_leader, tag+2, peer_comm); |
---|
273 | MPI_Recv(&size_info[3], 1, MPI_INT_STD, remote_leader, tag+2, peer_comm, &status); |
---|
274 | } |
---|
275 | |
---|
276 | ::MPI_Bcast(&size_info[2], 2, MPI_INT_STD, local_comm.rank_map->at(local_leader).second, local_mpi_comm); |
---|
277 | |
---|
278 | new_rank_info[2].resize(size_info[3]); |
---|
279 | new_rank_info[3].resize(size_info[3]); |
---|
280 | new_ep_info[1].resize(size_info[3]); |
---|
281 | |
---|
282 | send_buf.resize(size_info[2]); |
---|
283 | recv_buf.resize(size_info[3]); |
---|
284 | |
---|
285 | if(is_local_leader) |
---|
286 | { |
---|
287 | MPI_Status status; |
---|
288 | |
---|
289 | std::copy ( new_rank_info[0].data(), new_rank_info[0].data() + size_info[2], send_buf.begin() ); |
---|
290 | std::copy ( new_rank_info[1].data(), new_rank_info[1].data() + size_info[2], send_buf.begin() + size_info[2] ); |
---|
291 | std::copy ( new_ep_info[0].data(), new_ep_info[0].data() + size_info[0], send_buf.begin() + 2*size_info[2] ); |
---|
292 | |
---|
293 | MPI_Send(send_buf.data(), 3*size_info[2], MPI_INT_STD, remote_leader, tag+3, peer_comm); |
---|
294 | MPI_Recv(recv_buf.data(), 3*size_info[3], MPI_INT_STD, remote_leader, tag+3, peer_comm, &status); |
---|
295 | } |
---|
296 | |
---|
297 | ::MPI_Bcast(recv_buf.data(), 3*size_info[3], MPI_INT_STD, local_comm.rank_map->at(local_leader).second, local_mpi_comm); |
---|
298 | |
---|
299 | std::copy ( recv_buf.data(), recv_buf.data() + size_info[3], new_rank_info[2].begin() ); |
---|
300 | std::copy ( recv_buf.data() + size_info[3], recv_buf.data() + 2*size_info[3], new_rank_info[3].begin() ); |
---|
301 | std::copy ( recv_buf.data() + 2*size_info[3], recv_buf.data() + 3*size_info[3], new_ep_info[1].begin() ); |
---|
302 | |
---|
303 | } |
---|
304 | |
---|
305 | |
---|
306 | |
---|
307 | if(is_proc_master) |
---|
308 | { |
---|
309 | //! leader_info[4]: 2-> rank of local leader in new_group generated comm; |
---|
310 | // 3-> rank of remote leader in new_group generated comm; |
---|
311 | ::MPI_Group local_group; |
---|
312 | ::MPI_Group new_group; |
---|
313 | ::MPI_Comm new_comm; |
---|
314 | ::MPI_Comm intercomm; |
---|
315 | |
---|
316 | ::MPI_Comm_group(local_mpi_comm, &local_group); |
---|
317 | |
---|
318 | ::MPI_Group_incl(local_group, size_info[2], new_rank_info[1].data(), &new_group); |
---|
319 | |
---|
320 | ::MPI_Comm_create(local_mpi_comm, new_group, &new_comm); |
---|
321 | |
---|
322 | |
---|
323 | |
---|
324 | if(is_local_leader) |
---|
325 | { |
---|
326 | ::MPI_Comm_rank(new_comm, &leader_info[2]); |
---|
327 | } |
---|
328 | |
---|
329 | ::MPI_Bcast(&leader_info[2], 1, MPI_INT_STD, local_comm.rank_map->at(local_leader).second, local_mpi_comm); |
---|
330 | |
---|
331 | if(new_comm != MPI_COMM_NULL_STD) |
---|
332 | { |
---|
333 | |
---|
334 | ::MPI_Barrier(new_comm); |
---|
335 | |
---|
336 | ::MPI_Intercomm_create(new_comm, leader_info[2], static_cast< ::MPI_Comm>(peer_comm.mpi_comm), rank_in_peer_mpi[1], tag, &intercomm); |
---|
337 | |
---|
338 | int id; |
---|
339 | |
---|
340 | ::MPI_Comm_rank(new_comm, &id); |
---|
341 | int my_num_ep = new_ep_info[0][id]; |
---|
342 | |
---|
343 | MPI_Comm *ep_intercomm; |
---|
344 | MPI_Info info; |
---|
345 | MPI_Comm_create_endpoints(new_comm, my_num_ep, info, ep_intercomm); |
---|
346 | |
---|
347 | |
---|
348 | for(int i= 0; i<my_num_ep; i++) |
---|
349 | { |
---|
350 | ep_intercomm[i].is_intercomm = true; |
---|
351 | |
---|
352 | ep_intercomm[i].ep_comm_ptr->intercomm = new ep_lib::ep_intercomm; |
---|
353 | ep_intercomm[i].ep_comm_ptr->intercomm->mpi_inter_comm = intercomm; |
---|
354 | ep_intercomm[i].ep_comm_ptr->comm_label = leader_info[0]; |
---|
355 | } |
---|
356 | |
---|
357 | |
---|
358 | #pragma omp critical (write_to_tag_list) |
---|
359 | tag_list.push_back(make_pair( make_pair(tag, min(leader_info[0], leader_info[1])) , ep_intercomm)); |
---|
360 | //printf("tag_list size = %lu\n", tag_list.size()); |
---|
361 | } |
---|
362 | |
---|
363 | |
---|
364 | } |
---|
365 | |
---|
366 | |
---|
367 | MPI_Barrier_local(local_comm); |
---|
368 | |
---|
369 | vector<int> bcast_buf(8); |
---|
370 | if(is_local_leader) |
---|
371 | { |
---|
372 | std::copy(size_info, size_info+4, bcast_buf.begin()); |
---|
373 | std::copy(leader_info, leader_info+4, bcast_buf.begin()+4); |
---|
374 | } |
---|
375 | |
---|
376 | MPI_Bcast(bcast_buf.data(), 8, MPI_INT_STD, local_leader, local_comm); |
---|
377 | |
---|
378 | if(!is_local_leader) |
---|
379 | { |
---|
380 | std::copy(bcast_buf.begin(), bcast_buf.begin()+4, size_info); |
---|
381 | std::copy(bcast_buf.begin()+4, bcast_buf.begin()+8, leader_info); |
---|
382 | } |
---|
383 | |
---|
384 | if(!is_local_leader) |
---|
385 | { |
---|
386 | new_rank_info[1].resize(size_info[2]); |
---|
387 | ep_info[1].resize(size_info[1]); |
---|
388 | offset.resize(size_info[0]); |
---|
389 | } |
---|
390 | |
---|
391 | bcast_buf.resize(size_info[2]+size_info[1]+size_info[0]+1); |
---|
392 | |
---|
393 | if(is_local_leader) |
---|
394 | { |
---|
395 | bcast_buf[0] = remote_ep_size; |
---|
396 | std::copy(new_rank_info[1].data(), new_rank_info[1].data()+size_info[2], bcast_buf.begin()+1); |
---|
397 | std::copy(ep_info[1].data(), ep_info[1].data()+size_info[1], bcast_buf.begin()+size_info[2]+1); |
---|
398 | std::copy(offset.data(), offset.data()+size_info[0], bcast_buf.begin()+size_info[2]+size_info[1]+1); |
---|
399 | } |
---|
400 | |
---|
401 | MPI_Bcast(bcast_buf.data(), size_info[2]+size_info[1]+size_info[0]+1, MPI_INT_STD, local_leader, local_comm); |
---|
402 | |
---|
403 | if(!is_local_leader) |
---|
404 | { |
---|
405 | remote_ep_size = bcast_buf[0]; |
---|
406 | std::copy(bcast_buf.data()+1, bcast_buf.data()+1+size_info[2], new_rank_info[1].begin()); |
---|
407 | std::copy(bcast_buf.data()+1+size_info[2], bcast_buf.data()+1+size_info[2]+size_info[1], ep_info[1].begin()); |
---|
408 | std::copy(bcast_buf.data()+1+size_info[2]+size_info[1], bcast_buf.data()+1+size_info[2]+size_info[1]+size_info[0], offset.begin()); |
---|
409 | } |
---|
410 | |
---|
411 | int my_position = offset[rank_in_local_parent]+ep_rank_loc; |
---|
412 | |
---|
413 | |
---|
414 | MPI_Barrier_local(local_comm); |
---|
415 | #pragma omp flush |
---|
416 | |
---|
417 | |
---|
418 | #pragma omp critical (read_from_tag_list) |
---|
419 | { |
---|
420 | bool found = false; |
---|
421 | while(!found) |
---|
422 | { |
---|
423 | for(std::list<std::pair < std::pair<int,int>, MPI_Comm* > >::iterator iter = tag_list.begin(); iter!=tag_list.end(); iter++) |
---|
424 | { |
---|
425 | if((*iter).first == make_pair(tag, min(leader_info[0], leader_info[1]))) |
---|
426 | { |
---|
427 | *newintercomm = iter->second[my_position]; |
---|
428 | found = true; |
---|
429 | break; |
---|
430 | } |
---|
431 | } |
---|
432 | } |
---|
433 | } |
---|
434 | |
---|
435 | MPI_Barrier_local(local_comm); |
---|
436 | if(is_proc_master) |
---|
437 | { |
---|
438 | for(std::list<std::pair < std::pair<int,int>, MPI_Comm* > >::iterator iter = tag_list.begin(); iter!=tag_list.end(); iter++) |
---|
439 | { |
---|
440 | if((*iter).first == make_pair(tag, min(leader_info[0], leader_info[1]))) |
---|
441 | { |
---|
442 | tag_list.erase(iter); |
---|
443 | break; |
---|
444 | } |
---|
445 | } |
---|
446 | } |
---|
447 | |
---|
448 | int intercomm_ep_rank, intercomm_ep_rank_loc, intercomm_mpi_rank; |
---|
449 | int intercomm_ep_size, intercomm_num_ep, intercomm_mpi_size; |
---|
450 | |
---|
451 | intercomm_ep_rank = newintercomm->ep_comm_ptr->size_rank_info[0].first; |
---|
452 | intercomm_ep_rank_loc = newintercomm->ep_comm_ptr->size_rank_info[1].first; |
---|
453 | intercomm_mpi_rank = newintercomm->ep_comm_ptr->size_rank_info[2].first; |
---|
454 | intercomm_ep_size = newintercomm->ep_comm_ptr->size_rank_info[0].second; |
---|
455 | intercomm_num_ep = newintercomm->ep_comm_ptr->size_rank_info[1].second; |
---|
456 | intercomm_mpi_size = newintercomm->ep_comm_ptr->size_rank_info[2].second; |
---|
457 | |
---|
458 | MPI_Bcast(&remote_ep_size, 1, MPI_INT_STD, local_leader, local_comm); |
---|
459 | |
---|
460 | int my_rank_map_elem[2]; |
---|
461 | |
---|
462 | my_rank_map_elem[0] = intercomm_ep_rank; |
---|
463 | my_rank_map_elem[1] = (*newintercomm).ep_comm_ptr->comm_label; |
---|
464 | |
---|
465 | vector<pair<int, int> > local_rank_map_array; |
---|
466 | vector<pair<int, int> > remote_rank_map_array; |
---|
467 | |
---|
468 | |
---|
469 | (*newintercomm).ep_comm_ptr->intercomm->local_rank_map = new RANK_MAP; |
---|
470 | (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->resize(local_ep_size); |
---|
471 | |
---|
472 | MPI_Allgather(my_rank_map_elem, 2, MPI_INT_STD, (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->data(), 2, MPI_INT_STD, local_comm); |
---|
473 | |
---|
474 | (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map = new RANK_MAP; |
---|
475 | (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->resize(remote_ep_size); |
---|
476 | |
---|
477 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[0] = local_comm.ep_comm_ptr->size_rank_info[0]; |
---|
478 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[1] = local_comm.ep_comm_ptr->size_rank_info[1]; |
---|
479 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[2] = local_comm.ep_comm_ptr->size_rank_info[2]; |
---|
480 | |
---|
481 | int local_intercomm_size = intercomm_ep_size; |
---|
482 | int remote_intercomm_size; |
---|
483 | |
---|
484 | int new_bcast_root_0 = 0; |
---|
485 | int new_bcast_root = 0; |
---|
486 | |
---|
487 | |
---|
488 | if(is_local_leader) |
---|
489 | { |
---|
490 | MPI_Status status; |
---|
491 | MPI_Send((*newintercomm).ep_comm_ptr->intercomm->local_rank_map->data(), 2*local_ep_size, MPI_INT_STD, remote_leader, tag+4, peer_comm); |
---|
492 | MPI_Recv((*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->data(), 2*remote_ep_size, MPI_INT_STD, remote_leader, tag+4, peer_comm, &status); |
---|
493 | |
---|
494 | MPI_Send(&local_intercomm_size, 1, MPI_INT_STD, remote_leader, tag+5, peer_comm); |
---|
495 | MPI_Recv(&remote_intercomm_size, 1, MPI_INT_STD, remote_leader, tag+5, peer_comm, &status); |
---|
496 | |
---|
497 | new_bcast_root_0 = intercomm_ep_rank; |
---|
498 | } |
---|
499 | |
---|
500 | MPI_Allreduce(&new_bcast_root_0, &new_bcast_root, 1, MPI_INT_STD, MPI_SUM_STD, *newintercomm); |
---|
501 | |
---|
502 | |
---|
503 | MPI_Bcast((*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->data(), 2*remote_ep_size, MPI_INT_STD, local_leader, local_comm); |
---|
504 | MPI_Bcast(&remote_intercomm_size, 1, MPI_INT_STD, new_bcast_root, *newintercomm); |
---|
505 | |
---|
506 | |
---|
507 | (*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map = new RANK_MAP; |
---|
508 | (*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map->resize(remote_intercomm_size); |
---|
509 | |
---|
510 | |
---|
511 | |
---|
512 | |
---|
513 | if(is_local_leader) |
---|
514 | { |
---|
515 | MPI_Status status; |
---|
516 | MPI_Send((*newintercomm).rank_map->data(), 2*local_intercomm_size, MPI_INT_STD, remote_leader, tag+6, peer_comm); |
---|
517 | MPI_Recv((*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map->data(), 2*remote_intercomm_size, MPI_INT_STD, remote_leader, tag+6, peer_comm, &status); |
---|
518 | } |
---|
519 | |
---|
520 | MPI_Bcast((*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map->data(), 2*remote_intercomm_size, MPI_INT_STD, new_bcast_root, *newintercomm); |
---|
521 | |
---|
522 | (*newintercomm).ep_comm_ptr->intercomm->local_comm = &(local_comm.ep_comm_ptr->comm_list[ep_rank_loc]); |
---|
523 | (*newintercomm).ep_comm_ptr->intercomm->intercomm_tag = tag; |
---|
524 | |
---|
525 | /* |
---|
526 | for(int i=0; i<local_ep_size; i++) |
---|
527 | if(local_comm.ep_comm_ptr->comm_label == 0) printf("ep_rank (from EP) = %d, local_rank_map[%d] = (%d,%d)\n", intercomm_ep_rank, i, |
---|
528 | (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->at(i).first, (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->at(i).second); |
---|
529 | |
---|
530 | for(int i=0; i<remote_ep_size; i++) |
---|
531 | if(local_comm.ep_comm_ptr->comm_label == 0) printf("ep_rank (from EP) = %d, remote_rank_map[%d] = (%d,%d)\n", intercomm_ep_rank, i, |
---|
532 | (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->at(i).first, (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->at(i).second); |
---|
533 | |
---|
534 | for(int i=0; i<remote_intercomm_size; i++) |
---|
535 | if(local_comm.ep_comm_ptr->comm_label == 0) printf("ep_rank (from EP) = %d, intercomm_rank_map[%d] = (%d,%d)\n", intercomm_ep_rank, i, |
---|
536 | (*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map->at(i).first, (*newintercomm).ep_comm_ptr->intercomm->intercomm_rank_map->at(i).second); |
---|
537 | */ |
---|
538 | |
---|
539 | // for(int i=0; i<(*newintercomm).rank_map->size(); i++) |
---|
540 | // if(local_comm.ep_comm_ptr->comm_label != 99) printf("ep_rank = %d, rank_map[%d] = (%d,%d)\n", intercomm_ep_rank, i, |
---|
541 | // (*newintercomm).rank_map->at(i).first, (*newintercomm).rank_map->at(i).second); |
---|
542 | |
---|
543 | // MPI_Comm *test_comm = newintercomm->ep_comm_ptr->intercomm->local_comm; |
---|
544 | // int test_rank; |
---|
545 | // MPI_Comm_rank(*test_comm, &test_rank); |
---|
546 | // printf("=================test_rank = %d\n", test_rank); |
---|
547 | |
---|
548 | |
---|
549 | |
---|
550 | return MPI_SUCCESS; |
---|
551 | |
---|
552 | } |
---|
553 | |
---|
554 | |
---|
555 | |
---|
556 | |
---|
557 | int MPI_Intercomm_create_unique_leader(MPI_Comm local_comm, int local_leader, MPI_Comm peer_comm, int remote_leader, int tag, MPI_Comm *newintercomm) |
---|
558 | { |
---|
559 | //! mpi_size of local comm = 1 |
---|
560 | //! same world rank of leaders |
---|
561 | |
---|
562 | int ep_rank, ep_rank_loc, mpi_rank; |
---|
563 | int ep_size, num_ep, mpi_size; |
---|
564 | |
---|
565 | ep_rank = local_comm.ep_comm_ptr->size_rank_info[0].first; |
---|
566 | ep_rank_loc = local_comm.ep_comm_ptr->size_rank_info[1].first; |
---|
567 | mpi_rank = local_comm.ep_comm_ptr->size_rank_info[2].first; |
---|
568 | ep_size = local_comm.ep_comm_ptr->size_rank_info[0].second; |
---|
569 | num_ep = local_comm.ep_comm_ptr->size_rank_info[1].second; |
---|
570 | mpi_size = local_comm.ep_comm_ptr->size_rank_info[2].second; |
---|
571 | |
---|
572 | |
---|
573 | |
---|
574 | std::vector<int> rank_info[4]; //! 0->rank_in_world of local_comm, 1->rank_in_local_parent of local_comm |
---|
575 | //! 2->rank_in_world of remote_comm, 3->rank_in_local_parent of remote_comm |
---|
576 | |
---|
577 | int rank_in_world; |
---|
578 | |
---|
579 | int rank_in_peer_mpi[2]; |
---|
580 | |
---|
581 | ::MPI_Comm_rank(MPI_COMM_WORLD_STD, &rank_in_world); |
---|
582 | |
---|
583 | |
---|
584 | int local_num_ep = num_ep; |
---|
585 | int remote_num_ep; |
---|
586 | int total_num_ep; |
---|
587 | |
---|
588 | int leader_rank_in_peer[2]; |
---|
589 | |
---|
590 | int my_position; |
---|
591 | int tag_label[2]; |
---|
592 | |
---|
593 | vector<int> send_buf(4); |
---|
594 | vector<int> recv_buf(4); |
---|
595 | |
---|
596 | |
---|
597 | if(ep_rank == local_leader) |
---|
598 | { |
---|
599 | MPI_Status status; |
---|
600 | |
---|
601 | |
---|
602 | |
---|
603 | MPI_Comm_rank(peer_comm, &leader_rank_in_peer[0]); |
---|
604 | |
---|
605 | send_buf[0] = local_num_ep; |
---|
606 | send_buf[1] = leader_rank_in_peer[0]; |
---|
607 | |
---|
608 | MPI_Request req_s, req_r; |
---|
609 | |
---|
610 | MPI_Isend(send_buf.data(), 2, MPI_INT_STD, remote_leader, tag, peer_comm, &req_s); |
---|
611 | MPI_Irecv(recv_buf.data(), 2, MPI_INT_STD, remote_leader, tag, peer_comm, &req_r); |
---|
612 | |
---|
613 | |
---|
614 | MPI_Wait(&req_s, &status); |
---|
615 | MPI_Wait(&req_r, &status); |
---|
616 | |
---|
617 | recv_buf[2] = leader_rank_in_peer[0]; |
---|
618 | |
---|
619 | } |
---|
620 | |
---|
621 | MPI_Bcast(recv_buf.data(), 3, MPI_INT_STD, local_leader, local_comm); |
---|
622 | |
---|
623 | remote_num_ep = recv_buf[0]; |
---|
624 | leader_rank_in_peer[1] = recv_buf[1]; |
---|
625 | leader_rank_in_peer[0] = recv_buf[2]; |
---|
626 | |
---|
627 | total_num_ep = local_num_ep + remote_num_ep; |
---|
628 | |
---|
629 | |
---|
630 | if(leader_rank_in_peer[0] < leader_rank_in_peer[1]) |
---|
631 | { |
---|
632 | my_position = ep_rank_loc; |
---|
633 | //! LEADER create EP |
---|
634 | if(ep_rank == local_leader) |
---|
635 | { |
---|
636 | ::MPI_Comm mpi_dup; |
---|
637 | |
---|
638 | ::MPI_Comm_dup(static_cast< ::MPI_Comm>(local_comm.mpi_comm), &mpi_dup); |
---|
639 | |
---|
640 | MPI_Comm *ep_intercomm; |
---|
641 | MPI_Info info; |
---|
642 | MPI_Comm_create_endpoints(mpi_dup, total_num_ep, info, ep_intercomm); |
---|
643 | |
---|
644 | |
---|
645 | for(int i=0; i<total_num_ep; i++) |
---|
646 | { |
---|
647 | ep_intercomm[i].is_intercomm = true; |
---|
648 | ep_intercomm[i].ep_comm_ptr->intercomm = new ep_lib::ep_intercomm; |
---|
649 | ep_intercomm[i].ep_comm_ptr->intercomm->mpi_inter_comm = 0; |
---|
650 | |
---|
651 | ep_intercomm[i].ep_comm_ptr->comm_label = leader_rank_in_peer[0]; |
---|
652 | } |
---|
653 | |
---|
654 | tag_label[0] = TAG++; |
---|
655 | tag_label[1] = rank_in_world; |
---|
656 | |
---|
657 | #pragma omp critical (write_to_tag_list) |
---|
658 | tag_list.push_back(make_pair( make_pair(tag_label[0], tag_label[1]) , ep_intercomm)); |
---|
659 | |
---|
660 | MPI_Request req_s; |
---|
661 | MPI_Status sta_s; |
---|
662 | MPI_Isend(tag_label, 2, MPI_INT_STD, remote_leader, tag, peer_comm, &req_s); |
---|
663 | |
---|
664 | MPI_Wait(&req_s, &sta_s); |
---|
665 | |
---|
666 | } |
---|
667 | } |
---|
668 | else |
---|
669 | { |
---|
670 | //! Wait for EP creation |
---|
671 | my_position = remote_num_ep + ep_rank_loc; |
---|
672 | if(ep_rank == local_leader) |
---|
673 | { |
---|
674 | MPI_Status status; |
---|
675 | MPI_Request req_r; |
---|
676 | MPI_Irecv(tag_label, 2, MPI_INT_STD, remote_leader, tag, peer_comm, &req_r); |
---|
677 | MPI_Wait(&req_r, &status); |
---|
678 | } |
---|
679 | } |
---|
680 | |
---|
681 | MPI_Bcast(tag_label, 2, MPI_INT_STD, local_leader, local_comm); |
---|
682 | |
---|
683 | |
---|
684 | |
---|
685 | |
---|
686 | #pragma omp critical (read_from_tag_list) |
---|
687 | { |
---|
688 | bool found = false; |
---|
689 | while(!found) |
---|
690 | { |
---|
691 | for(std::list<std::pair < std::pair<int,int>, MPI_Comm* > >::iterator iter = tag_list.begin(); iter!=tag_list.end(); iter++) |
---|
692 | { |
---|
693 | if((*iter).first == make_pair(tag_label[0], tag_label[1])) |
---|
694 | { |
---|
695 | *newintercomm = iter->second[my_position]; |
---|
696 | found = true; |
---|
697 | // tag_list.erase(iter); |
---|
698 | break; |
---|
699 | } |
---|
700 | } |
---|
701 | } |
---|
702 | } |
---|
703 | |
---|
704 | MPI_Barrier_local(local_comm); |
---|
705 | |
---|
706 | if(leader_rank_in_peer[0] < leader_rank_in_peer[1]) |
---|
707 | { |
---|
708 | for(std::list<std::pair < std::pair<int,int>, MPI_Comm* > >::iterator iter = tag_list.begin(); iter!=tag_list.end(); iter++) |
---|
709 | { |
---|
710 | if((*iter).first == make_pair(tag_label[0], tag_label[1])) |
---|
711 | { |
---|
712 | tag_list.erase(iter); |
---|
713 | break; |
---|
714 | } |
---|
715 | } |
---|
716 | } |
---|
717 | |
---|
718 | |
---|
719 | |
---|
720 | int intercomm_ep_rank, intercomm_ep_rank_loc, intercomm_mpi_rank; |
---|
721 | int intercomm_ep_size, intercomm_num_ep, intercomm_mpi_size; |
---|
722 | |
---|
723 | intercomm_ep_rank = newintercomm->ep_comm_ptr->size_rank_info[0].first; |
---|
724 | intercomm_ep_rank_loc = newintercomm->ep_comm_ptr->size_rank_info[1].first; |
---|
725 | intercomm_mpi_rank = newintercomm->ep_comm_ptr->size_rank_info[2].first; |
---|
726 | intercomm_ep_size = newintercomm->ep_comm_ptr->size_rank_info[0].second; |
---|
727 | intercomm_num_ep = newintercomm->ep_comm_ptr->size_rank_info[1].second; |
---|
728 | intercomm_mpi_size = newintercomm->ep_comm_ptr->size_rank_info[2].second; |
---|
729 | |
---|
730 | |
---|
731 | |
---|
732 | (*newintercomm).ep_comm_ptr->intercomm->local_rank_map = new RANK_MAP; |
---|
733 | (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map = new RANK_MAP; |
---|
734 | (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->resize(local_num_ep); |
---|
735 | (*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->resize(remote_num_ep); |
---|
736 | |
---|
737 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[0] = local_comm.ep_comm_ptr->size_rank_info[0]; |
---|
738 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[1] = local_comm.ep_comm_ptr->size_rank_info[1]; |
---|
739 | (*newintercomm).ep_comm_ptr->intercomm->size_rank_info[2] = local_comm.ep_comm_ptr->size_rank_info[2]; |
---|
740 | |
---|
741 | |
---|
742 | |
---|
743 | int local_rank_map_ele[2]; |
---|
744 | local_rank_map_ele[0] = intercomm_ep_rank; |
---|
745 | local_rank_map_ele[1] = (*newintercomm).ep_comm_ptr->comm_label; |
---|
746 | |
---|
747 | MPI_Allgather(local_rank_map_ele, 2, MPI_INT_STD, (*newintercomm).ep_comm_ptr->intercomm->local_rank_map->data(), 2, MPI_INT_STD, local_comm); |
---|
748 | |
---|
749 | if(ep_rank == local_leader) |
---|
750 | { |
---|
751 | MPI_Status status; |
---|
752 | MPI_Request req_s, req_r; |
---|
753 | |
---|
754 | MPI_Isend((*newintercomm).ep_comm_ptr->intercomm->local_rank_map->data(), 2*local_num_ep, MPI_INT_STD, remote_leader, tag, peer_comm, &req_s); |
---|
755 | MPI_Irecv((*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->data(), 2*remote_num_ep, MPI_INT_STD, remote_leader, tag, peer_comm, &req_r); |
---|
756 | |
---|
757 | |
---|
758 | MPI_Wait(&req_s, &status); |
---|
759 | MPI_Wait(&req_r, &status); |
---|
760 | |
---|
761 | } |
---|
762 | |
---|
763 | MPI_Bcast((*newintercomm).ep_comm_ptr->intercomm->remote_rank_map->data(), 2*remote_num_ep, MPI_INT_STD, local_leader, local_comm); |
---|
764 | (*newintercomm).ep_comm_ptr->intercomm->local_comm = &(local_comm.ep_comm_ptr->comm_list[ep_rank_loc]); |
---|
765 | (*newintercomm).ep_comm_ptr->intercomm->intercomm_tag = tag; |
---|
766 | |
---|
767 | |
---|
768 | return MPI_SUCCESS; |
---|
769 | } |
---|
770 | |
---|
771 | |
---|
772 | } |
---|