[688] | 1 | #include "mpi_routing.hpp" |
---|
[694] | 2 | #include "mpi.hpp" |
---|
[688] | 3 | #include "node.hpp" |
---|
| 4 | #include "elt.hpp" |
---|
| 5 | #include "timerRemap.hpp" |
---|
| 6 | #include <iostream> |
---|
| 7 | |
---|
| 8 | namespace sphereRemap { |
---|
| 9 | |
---|
| 10 | const int verbose = 0; |
---|
| 11 | |
---|
[1638] | 12 | CMPIRouting::CMPIRouting(ep_lib::MPI_Comm comm) : communicator(comm) |
---|
[688] | 13 | { |
---|
[1638] | 14 | ep_lib::MPI_Comm_rank(comm, &mpiRank); |
---|
| 15 | ep_lib::MPI_Comm_size(comm, &mpiSize); |
---|
[688] | 16 | } |
---|
| 17 | |
---|
| 18 | /* sparse alltoallv when it is known that only a subset of ranks will communicate, |
---|
| 19 | but message lengths are *known* to receiver */ |
---|
| 20 | template <typename T> |
---|
[1638] | 21 | void alltoalls_known(const vector<vector<T> >& send, vector<vector<T> >& recv, const vector<int>& ranks, ep_lib::MPI_Comm communicator) |
---|
[688] | 22 | { |
---|
[1638] | 23 | vector<ep_lib::MPI_Request> request(ranks.size() * 2); |
---|
| 24 | vector<ep_lib::MPI_Status> status(ranks.size() * 2); |
---|
[688] | 25 | |
---|
| 26 | // communicate data |
---|
| 27 | int nbRequest = 0; |
---|
| 28 | for (int i = 0; i < ranks.size(); i++) |
---|
| 29 | if (recv[i].size()) |
---|
[1638] | 30 | ep_lib::MPI_Irecv(&recv[i][0], recv[i].size()*sizeof(T), EP_CHAR, ranks[i], 0, communicator, &request[nbRequest++]); |
---|
[688] | 31 | for (int i = 0; i < ranks.size(); i++) |
---|
| 32 | if (send[i].size()) |
---|
[1638] | 33 | ep_lib::MPI_Isend((void *) &send[i][0], send[i].size()*sizeof(T), EP_CHAR, ranks[i], 0, communicator, &request[nbRequest++]); |
---|
| 34 | ep_lib::MPI_Waitall(nbRequest, &request[0], &status[0]); |
---|
[688] | 35 | } |
---|
| 36 | |
---|
| 37 | /* sparse alltoallv when it is known that only a subset of ranks will communicate, |
---|
| 38 | but message lengths are *unknown* to receiver */ |
---|
| 39 | template <typename T> |
---|
[1638] | 40 | void alltoalls_unknown(const vector<vector<T> >& send, vector<vector<T> >& recv, const vector<int>& ranks, ep_lib::MPI_Comm communicator) |
---|
[688] | 41 | { |
---|
[1638] | 42 | vector<ep_lib::MPI_Request> request(ranks.size() * 2); |
---|
| 43 | vector<ep_lib::MPI_Status> status(ranks.size() * 2); |
---|
[688] | 44 | |
---|
| 45 | // communicate sizes |
---|
| 46 | int nbRequest = 0; |
---|
| 47 | vector<int> sendSizes(ranks.size()); |
---|
| 48 | vector<int> recvSizes(ranks.size()); |
---|
| 49 | for (int i = 0; i < ranks.size(); i++) |
---|
| 50 | sendSizes[i] = send[i].size(); |
---|
| 51 | for (int i = 0; i < ranks.size(); i++) |
---|
[1638] | 52 | ep_lib::MPI_Irecv(&recvSizes[i], 1, EP_INT, ranks[i], 0, communicator, &request[nbRequest++]); |
---|
[688] | 53 | for (int i = 0; i < ranks.size(); i++) |
---|
[1638] | 54 | ep_lib::MPI_Isend(&sendSizes[i], 1, EP_INT, ranks[i], 0, communicator, &request[nbRequest++]); |
---|
| 55 | ep_lib::MPI_Waitall(nbRequest, &request[0], &status[0]); |
---|
[688] | 56 | |
---|
| 57 | // allocate |
---|
| 58 | for (int i = 0; i < ranks.size(); i++) |
---|
| 59 | if (recvSizes[i]) recv[i].resize(recvSizes[i]); |
---|
| 60 | // communicate data |
---|
| 61 | alltoalls_known(send, recv, ranks, communicator); |
---|
| 62 | } |
---|
| 63 | |
---|
| 64 | void setElementsSendCnt(int route, vector<int>& sendCnt) |
---|
| 65 | { |
---|
| 66 | sendCnt[route]++; |
---|
| 67 | } |
---|
| 68 | |
---|
| 69 | void setElementsSendCnt(const vector<int>& route, vector<int>& sendCnt) |
---|
| 70 | { |
---|
| 71 | for (int i = 0; i < route.size(); i++) |
---|
| 72 | sendCnt[route[i]]++; |
---|
| 73 | } |
---|
| 74 | |
---|
| 75 | void setTargetElementIndex(int route, vector<int>& indices, const int *rankToIndex) |
---|
| 76 | { |
---|
| 77 | int index = rankToIndex[route]; |
---|
| 78 | indices.push_back(index); |
---|
| 79 | } |
---|
| 80 | |
---|
| 81 | void setTargetElementIndex(const vector<int>& route, vector<int>& indices, const int *rankToIndex) |
---|
| 82 | { |
---|
| 83 | for (int i = 0; i < route.size(); i++) |
---|
| 84 | { |
---|
| 85 | int index = rankToIndex[route[i]]; |
---|
| 86 | indices.push_back(index); |
---|
| 87 | } |
---|
| 88 | } |
---|
| 89 | |
---|
| 90 | template<typename T> |
---|
| 91 | void CMPIRouting::init(const vector<T>& route, CMPICascade *cascade) |
---|
| 92 | { |
---|
| 93 | vector<int> nbElementSend(mpiSize); |
---|
| 94 | int *toSend = new int[mpiSize]; |
---|
| 95 | int *recvCount = new int[mpiSize]; |
---|
| 96 | int *targetRankToIndex; |
---|
| 97 | |
---|
| 98 | for (int i = 0; i < route.size(); i++) |
---|
| 99 | setElementsSendCnt(route[i], nbElementSend); |
---|
| 100 | |
---|
| 101 | nbTarget = 0; |
---|
| 102 | vector<int> destRanks; |
---|
| 103 | for (int i = 0; i < mpiSize; i++) |
---|
| 104 | { |
---|
| 105 | if (nbElementSend[i] == 0) |
---|
| 106 | toSend[i] = 0; |
---|
| 107 | else |
---|
| 108 | { |
---|
| 109 | destRanks.push_back(i); |
---|
| 110 | toSend[i] = 1; |
---|
| 111 | nbTarget++; |
---|
| 112 | } |
---|
| 113 | recvCount[i] = 1; |
---|
| 114 | } |
---|
| 115 | //int recvCntDeb = (stree) ? stree->scatter_reduce(destRanks) : -1; |
---|
| 116 | |
---|
| 117 | MPI_Barrier(communicator); |
---|
| 118 | CTimer::get("CMPIRouting::init(reduce_scatter)").reset(); |
---|
| 119 | CTimer::get("CMPIRouting::init(reduce_scatter)").resume(); |
---|
[1638] | 120 | ep_lib::MPI_Reduce_scatter(toSend, &nbSource, recvCount, EP_INT, EP_SUM, communicator); |
---|
[688] | 121 | CTimer::get("CMPIRouting::init(reduce_scatter)").suspend(); |
---|
| 122 | CTimer::get("CMPIRouting::init(reduce_scatter)").print(); |
---|
| 123 | |
---|
[1638] | 124 | ep_lib::MPI_Alloc_mem(nbTarget *sizeof(int), EP_INFO_NULL, &targetRank); |
---|
| 125 | ep_lib::MPI_Alloc_mem(nbSource *sizeof(int), EP_INFO_NULL, &sourceRank); |
---|
[688] | 126 | |
---|
| 127 | targetRankToIndex = new int[mpiSize]; |
---|
| 128 | int index = 0; |
---|
| 129 | for (int i = 0; i < mpiSize; i++) |
---|
| 130 | { |
---|
| 131 | if (toSend[i] == 1) |
---|
| 132 | { |
---|
| 133 | targetRank[index] = i; |
---|
| 134 | targetRankToIndex[i] = index; |
---|
| 135 | index++; |
---|
| 136 | } |
---|
| 137 | } |
---|
| 138 | |
---|
[1638] | 139 | ep_lib::MPI_Barrier(communicator); |
---|
[688] | 140 | CTimer::get("CMPIRouting::init(get_source)").reset(); |
---|
| 141 | CTimer::get("CMPIRouting::init(get_source)").resume(); |
---|
| 142 | |
---|
[1638] | 143 | ep_lib::MPI_Request *request = new ep_lib::MPI_Request[nbSource + nbTarget]; |
---|
| 144 | ep_lib::MPI_Status *status = new ep_lib::MPI_Status[nbSource + nbTarget]; |
---|
[688] | 145 | |
---|
| 146 | int indexRequest = 0; |
---|
| 147 | if (verbose) cout << "CMPIRouting::init nbSource : " << nbSource << " nbTarget : " << nbTarget << endl; |
---|
| 148 | // assert(recvCntDeb == -1 || recvCntDeb == nbSource); |
---|
| 149 | //cout << mpiRank << "DEB : " << recvCntDeb << " nbSource " << nbSource << " nbTarget : " << nbTarget << endl; |
---|
| 150 | for (int i = 0; i < nbSource; i++) |
---|
| 151 | { |
---|
[1638] | 152 | #ifdef _usingMPI |
---|
| 153 | ep_lib::MPI_Irecv(&sourceRank[i], 1, EP_INT, MPI_ANY_SOURCE, 0, communicator, &request[indexRequest]); |
---|
| 154 | #elif _usingEP |
---|
| 155 | ep_lib::MPI_Irecv(&sourceRank[i], 1, EP_INT, -2, 0, communicator, &request[indexRequest]); |
---|
| 156 | #endif |
---|
| 157 | indexRequest++; |
---|
[688] | 158 | } |
---|
| 159 | MPI_Barrier(communicator); |
---|
| 160 | for (int i = 0; i < nbTarget; i++) |
---|
| 161 | { |
---|
[1638] | 162 | ep_lib::MPI_Isend(&mpiRank, 1, EP_INT, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 163 | indexRequest++; |
---|
| 164 | } |
---|
| 165 | MPI_Waitall(indexRequest, request, status); |
---|
| 166 | MPI_Barrier(communicator); //needed |
---|
| 167 | CTimer::get("CMPIRouting::init(get_source)").suspend(); |
---|
| 168 | CTimer::get("CMPIRouting::init(get_source)").print(); |
---|
| 169 | |
---|
| 170 | CTimer::get("CMPIRouting::init(get_source)").reset(); |
---|
| 171 | CTimer::get("CMPIRouting::init(get_source)").resume(); |
---|
| 172 | |
---|
| 173 | indexRequest = 0; |
---|
| 174 | for (int i = 0; i < nbSource; i++) |
---|
| 175 | { |
---|
[1638] | 176 | #ifdef _usingMPI |
---|
| 177 | ep_lib::MPI_Irecv(&sourceRank[i], 1, EP_INT, MPI_ANY_SOURCE, 0, communicator, &request[indexRequest]); |
---|
| 178 | #elif _usingEP |
---|
| 179 | ep_lib::MPI_Irecv(&sourceRank[i], 1, EP_INT, -2, 0, communicator, &request[indexRequest]); |
---|
| 180 | #endif |
---|
| 181 | indexRequest++; |
---|
[688] | 182 | } |
---|
| 183 | |
---|
| 184 | for (int i = 0; i < nbTarget; i++) |
---|
| 185 | { |
---|
[1638] | 186 | ep_lib::MPI_Isend(&mpiRank, 1, EP_INT, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 187 | indexRequest++; |
---|
| 188 | } |
---|
| 189 | MPI_Waitall(indexRequest, request, status); |
---|
| 190 | MPI_Barrier(communicator); |
---|
| 191 | CTimer::get("CMPIRouting::init(get_source)").suspend(); |
---|
| 192 | CTimer::get("CMPIRouting::init(get_source)").print(); |
---|
| 193 | |
---|
| 194 | CTimer::get("CMPIRouting::init(send_element)").reset(); |
---|
| 195 | CTimer::get("CMPIRouting::init(send_element)").resume(); |
---|
| 196 | |
---|
| 197 | nbTargetElement.resize(nbTarget); |
---|
| 198 | nbSourceElement.resize(nbSource); |
---|
| 199 | |
---|
| 200 | for (int i = 0; i < route.size(); i++) |
---|
| 201 | setTargetElementIndex(route[i], targetElementIndex, targetRankToIndex); |
---|
| 202 | |
---|
| 203 | for (int i = 0; i < targetElementIndex.size(); i++) |
---|
| 204 | nbTargetElement[targetElementIndex[i]]++; |
---|
| 205 | |
---|
| 206 | indexRequest = 0; |
---|
| 207 | totalSourceElement = 0; |
---|
| 208 | totalTargetElement = 0; |
---|
| 209 | for (int i = 0; i < nbSource; i++) |
---|
| 210 | { |
---|
[1638] | 211 | ep_lib::MPI_Irecv(&nbSourceElement[i], 1, EP_INT, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 212 | indexRequest++; |
---|
| 213 | } |
---|
| 214 | |
---|
| 215 | for (int i = 0; i < nbTarget; i++) |
---|
| 216 | { |
---|
| 217 | totalTargetElement += nbTargetElement[i]; |
---|
[1638] | 218 | ep_lib::MPI_Isend(&nbTargetElement[i], 1, EP_INT, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 219 | indexRequest++; |
---|
| 220 | } |
---|
| 221 | |
---|
| 222 | MPI_Waitall(indexRequest, request, status); |
---|
| 223 | |
---|
| 224 | CTimer::get("CMPIRouting::init(send_element)").suspend(); |
---|
| 225 | CTimer::get("CMPIRouting::init(send_element)").print(); |
---|
| 226 | |
---|
| 227 | totalSourceElement = 0; |
---|
| 228 | for (int i = 0; i < nbSource; i++) |
---|
| 229 | totalSourceElement += nbSourceElement[i]; |
---|
| 230 | |
---|
| 231 | sourceElementIndex.resize(totalSourceElement); |
---|
| 232 | |
---|
| 233 | totalSourceElement = 0; |
---|
| 234 | for (int i = 0; i < nbSource; i++) |
---|
| 235 | { |
---|
| 236 | for (int j = 0; j < nbSourceElement[i]; j++) |
---|
| 237 | { |
---|
| 238 | sourceElementIndex[totalSourceElement] = i; |
---|
| 239 | totalSourceElement++; |
---|
| 240 | } |
---|
| 241 | } |
---|
| 242 | |
---|
| 243 | delete[] toSend; |
---|
| 244 | delete[] recvCount; |
---|
| 245 | delete[] request; |
---|
| 246 | delete[] status; |
---|
| 247 | } |
---|
| 248 | |
---|
| 249 | |
---|
| 250 | int CMPIRouting::getTotalSourceElement(void) |
---|
| 251 | { |
---|
| 252 | return totalSourceElement; |
---|
| 253 | } |
---|
| 254 | |
---|
| 255 | template<typename T> |
---|
| 256 | void CMPIRouting::transferToTarget(T* targetElements, T* sourceElements) |
---|
| 257 | { |
---|
| 258 | char** targetBuffer = new char*[nbTarget]; |
---|
| 259 | int* indexTargetBuffer = new int[nbTarget]; |
---|
| 260 | |
---|
| 261 | for(int i=0;i<nbTarget; i++) |
---|
| 262 | { |
---|
| 263 | targetBuffer[i] = new char[sizeof(T)*nbTargetElement[i]]; |
---|
| 264 | indexTargetBuffer[i]= 0; |
---|
| 265 | } |
---|
| 266 | |
---|
| 267 | char** sourceBuffer = new char*[nbSource]; |
---|
| 268 | int* indexSourceBuffer = new int[nbSource]; |
---|
| 269 | |
---|
| 270 | for(int i=0;i<nbSource; i++) |
---|
| 271 | { |
---|
| 272 | sourceBuffer[i] = new char[sizeof(T)*nbSourceElement[i]]; |
---|
| 273 | indexSourceBuffer[i]= 0; |
---|
| 274 | } |
---|
| 275 | |
---|
| 276 | // pack the data |
---|
| 277 | int index; |
---|
| 278 | for(int i=0;i<totalTargetElement;i++) |
---|
| 279 | { |
---|
| 280 | index=targetElementIndex[i]; |
---|
| 281 | *((T*) &(targetBuffer[index][indexTargetBuffer[index]])) = targetElements[i]; |
---|
| 282 | indexTargetBuffer[index]+=sizeof(T); |
---|
| 283 | } |
---|
| 284 | |
---|
| 285 | |
---|
[1638] | 286 | ep_lib::MPI_Request* request=new ep_lib::MPI_Request[nbSource+nbTarget]; |
---|
| 287 | ep_lib::MPI_Status* status=new ep_lib::MPI_Status[nbSource+nbTarget]; |
---|
[688] | 288 | int indexRequest=0; |
---|
| 289 | |
---|
[1638] | 290 | ep_lib::MPI_Barrier(communicator); |
---|
[688] | 291 | CTimer::get("CMPIRouting::transferToTarget").reset(); |
---|
| 292 | CTimer::get("CMPIRouting::transferToTarget").resume(); |
---|
| 293 | |
---|
| 294 | for(int i=0; i<nbSource; i++) |
---|
| 295 | { |
---|
[1638] | 296 | ep_lib::MPI_Irecv(sourceBuffer[i],nbSourceElement[i]*sizeof(T),EP_CHAR, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 297 | indexRequest++; |
---|
| 298 | } |
---|
| 299 | |
---|
| 300 | for(int i=0;i<nbTarget; i++) |
---|
| 301 | { |
---|
[1638] | 302 | ep_lib::MPI_Isend(targetBuffer[i],nbTargetElement[i]*sizeof(T), EP_CHAR, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 303 | indexRequest++; |
---|
| 304 | } |
---|
| 305 | |
---|
[1638] | 306 | ep_lib::MPI_Waitall(indexRequest,request,status); |
---|
[688] | 307 | |
---|
| 308 | CTimer::get("CMPIRouting::transferToTarget").suspend(); |
---|
| 309 | CTimer::get("CMPIRouting::transferToTarget").print(); |
---|
[1638] | 310 | ep_lib::MPI_Barrier(communicator); |
---|
[688] | 311 | |
---|
| 312 | // unpack the data |
---|
| 313 | for(int i=0;i<totalSourceElement;i++) |
---|
| 314 | { |
---|
| 315 | index=sourceElementIndex[i]; |
---|
| 316 | sourceElements[i] = *((T*) &(sourceBuffer[index][indexSourceBuffer[index]])); |
---|
| 317 | indexSourceBuffer[index]+=sizeof(T); |
---|
| 318 | } |
---|
| 319 | |
---|
| 320 | for(int i=0;i<nbTarget; i++) delete[] targetBuffer[i]; |
---|
| 321 | for(int i=0;i<nbSource; i++) delete[] sourceBuffer[i]; |
---|
| 322 | delete[] targetBuffer; |
---|
| 323 | delete[] indexTargetBuffer; |
---|
| 324 | delete[] sourceBuffer; |
---|
| 325 | delete[] indexSourceBuffer; |
---|
| 326 | delete[] request; |
---|
| 327 | delete[] status; |
---|
| 328 | } |
---|
| 329 | |
---|
| 330 | |
---|
| 331 | template<typename T, typename t_pack, typename t_unpack> |
---|
| 332 | void CMPIRouting::transferToTarget(T* targetElements, T* sourceElements, t_pack pack, t_unpack unpack) |
---|
| 333 | { |
---|
| 334 | char** targetBuffer = new char*[nbTarget]; |
---|
| 335 | int* indexTargetBuffer = new int[nbTarget]; |
---|
| 336 | int* targetMessageSize = new int[nbTarget]; |
---|
| 337 | int* sourceMessageSize = new int[nbSource]; |
---|
| 338 | int index; |
---|
| 339 | |
---|
| 340 | // compute target buffer size |
---|
| 341 | for (int i = 0; i < nbTarget; i++) |
---|
| 342 | targetMessageSize[i] = 0; |
---|
| 343 | |
---|
| 344 | for (int i = 0; i < totalTargetElement; i++) |
---|
| 345 | { |
---|
| 346 | index = targetElementIndex[i]; |
---|
| 347 | pack(targetElements[i], NULL, targetMessageSize[index]); |
---|
| 348 | } |
---|
| 349 | |
---|
[1638] | 350 | ep_lib::MPI_Request *request = new ep_lib::MPI_Request[nbSource + nbTarget]; |
---|
| 351 | ep_lib::MPI_Status *status = new ep_lib::MPI_Status[nbSource + nbTarget]; |
---|
[688] | 352 | int indexRequest = 0; |
---|
| 353 | |
---|
[1638] | 354 | ep_lib::MPI_Barrier(communicator); |
---|
[688] | 355 | CTimer::get("CMPIRouting::transferToTarget(messageSize)").reset(); |
---|
| 356 | CTimer::get("CMPIRouting::transferToTarget(messageSize)").resume(); |
---|
| 357 | |
---|
| 358 | for(int i=0; i<nbSource; i++) |
---|
| 359 | { |
---|
[1638] | 360 | ep_lib::MPI_Irecv(&sourceMessageSize[i],1,EP_INT, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 361 | indexRequest++; |
---|
| 362 | } |
---|
| 363 | |
---|
| 364 | for(int i=0; i<nbTarget; i++) |
---|
| 365 | { |
---|
[1638] | 366 | ep_lib::MPI_Isend(&targetMessageSize[i],1, EP_INT, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 367 | indexRequest++; |
---|
| 368 | } |
---|
| 369 | |
---|
[1638] | 370 | ep_lib::MPI_Waitall(indexRequest,request,status); |
---|
[688] | 371 | |
---|
[1638] | 372 | ep_lib::MPI_Barrier(communicator); |
---|
[688] | 373 | CTimer::get("CMPIRouting::transferToTarget(messageSize)").suspend(); |
---|
| 374 | CTimer::get("CMPIRouting::transferToTarget(messageSize)").print(); |
---|
| 375 | |
---|
| 376 | for(int i=0; i<nbTarget; i++) |
---|
| 377 | { |
---|
| 378 | targetBuffer[i] = new char[targetMessageSize[i]]; |
---|
| 379 | indexTargetBuffer[i] = 0; |
---|
| 380 | } |
---|
| 381 | |
---|
| 382 | char** sourceBuffer = new char*[nbSource]; |
---|
| 383 | int* indexSourceBuffer = new int[nbSource]; |
---|
| 384 | |
---|
| 385 | for(int i=0;i<nbSource; i++) |
---|
| 386 | { |
---|
| 387 | sourceBuffer[i] = new char[sourceMessageSize[i]]; |
---|
| 388 | indexSourceBuffer[i]= 0; |
---|
| 389 | } |
---|
| 390 | |
---|
| 391 | // pack the data |
---|
| 392 | for(int i=0; i<totalTargetElement; i++) |
---|
| 393 | { |
---|
| 394 | index=targetElementIndex[i]; |
---|
| 395 | pack(targetElements[i], targetBuffer[index], indexTargetBuffer[index] ); |
---|
| 396 | } |
---|
| 397 | |
---|
| 398 | indexRequest=0; |
---|
| 399 | |
---|
| 400 | MPI_Barrier(communicator); |
---|
| 401 | CTimer::get("CMPIRouting::transferToTarget(data)").reset(); |
---|
| 402 | CTimer::get("CMPIRouting::transferToTarget(data)").resume(); |
---|
| 403 | for(int i=0; i<nbSource; i++) |
---|
| 404 | { |
---|
[1638] | 405 | ep_lib::MPI_Irecv(sourceBuffer[i],sourceMessageSize[i],EP_CHAR, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 406 | indexRequest++; |
---|
| 407 | } |
---|
| 408 | |
---|
| 409 | for(int i=0;i<nbTarget; i++) |
---|
| 410 | { |
---|
[1638] | 411 | ep_lib::MPI_Isend(targetBuffer[i],targetMessageSize[i], EP_CHAR, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 412 | indexRequest++; |
---|
| 413 | } |
---|
| 414 | |
---|
| 415 | MPI_Waitall(indexRequest,request,status); |
---|
| 416 | |
---|
| 417 | MPI_Barrier(communicator); |
---|
| 418 | CTimer::get("CMPIRouting::transferToTarget(data)").suspend(); |
---|
| 419 | CTimer::get("CMPIRouting::transferToTarget(data)").print(); |
---|
| 420 | |
---|
| 421 | // unpack the data |
---|
| 422 | for(int i=0; i<totalSourceElement; i++) |
---|
| 423 | { |
---|
| 424 | index=sourceElementIndex[i]; |
---|
| 425 | unpack(sourceElements[i], sourceBuffer[index], indexSourceBuffer[index]); |
---|
| 426 | } |
---|
| 427 | |
---|
| 428 | for (int i=0; i<nbTarget; i++) delete[] targetBuffer[i]; |
---|
| 429 | for (int i=0; i<nbSource; i++) delete[] sourceBuffer[i]; |
---|
| 430 | delete[] targetBuffer; |
---|
| 431 | delete[] indexTargetBuffer; |
---|
| 432 | delete[] targetMessageSize; |
---|
| 433 | delete[] sourceBuffer; |
---|
| 434 | delete[] indexSourceBuffer; |
---|
| 435 | delete[] sourceMessageSize; |
---|
| 436 | delete[] request; |
---|
| 437 | delete[] status; |
---|
| 438 | } |
---|
| 439 | |
---|
| 440 | template<typename T> |
---|
| 441 | void CMPIRouting::transferFromSource(T* targetElements, T* sourceElements) |
---|
| 442 | { |
---|
| 443 | char** targetBuffer = new char*[nbTarget]; |
---|
| 444 | int* indexTargetBuffer = new int[nbTarget]; |
---|
| 445 | |
---|
| 446 | for (int i = 0; i < nbTarget; i++) |
---|
| 447 | { |
---|
| 448 | targetBuffer[i] = new char[sizeof(T)*nbTargetElement[i]]; |
---|
| 449 | indexTargetBuffer[i] = 0; |
---|
| 450 | } |
---|
| 451 | |
---|
| 452 | char** sourceBuffer = new char*[nbSource]; |
---|
| 453 | int* indexSourceBuffer = new int[nbSource]; |
---|
| 454 | |
---|
| 455 | for (int i = 0; i < nbSource; i++) |
---|
| 456 | { |
---|
| 457 | sourceBuffer[i] = new char[sizeof(T)*nbSourceElement[i]]; |
---|
| 458 | indexSourceBuffer[i] = 0; |
---|
| 459 | } |
---|
| 460 | |
---|
| 461 | // pack the data |
---|
| 462 | int index; |
---|
| 463 | for (int i = 0; i < totalSourceElement; i++) |
---|
| 464 | { |
---|
| 465 | index = sourceElementIndex[i]; |
---|
| 466 | *((T*) &(sourceBuffer[index][indexSourceBuffer[index]])) = sourceElements[i]; |
---|
| 467 | indexSourceBuffer[index] += sizeof(T); |
---|
| 468 | } |
---|
| 469 | |
---|
[1638] | 470 | ep_lib::MPI_Request* request=new ep_lib::MPI_Request[nbSource+nbTarget]; |
---|
| 471 | ep_lib::MPI_Status* status=new ep_lib::MPI_Status[nbSource+nbTarget]; |
---|
[688] | 472 | int indexRequest=0; |
---|
| 473 | |
---|
| 474 | for(int i=0; i<nbSource; i++) |
---|
| 475 | { |
---|
[1638] | 476 | ep_lib::MPI_Isend(sourceBuffer[i],nbSourceElement[i]*sizeof(T),EP_CHAR, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 477 | indexRequest++; |
---|
| 478 | } |
---|
| 479 | |
---|
| 480 | for(int i=0;i<nbTarget; i++) |
---|
| 481 | { |
---|
[1638] | 482 | ep_lib::MPI_Irecv(targetBuffer[i],nbTargetElement[i]*sizeof(T), EP_CHAR, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 483 | indexRequest++; |
---|
| 484 | } |
---|
| 485 | |
---|
[1638] | 486 | ep_lib::MPI_Waitall(indexRequest,request,status); |
---|
[688] | 487 | |
---|
| 488 | // unpack the data |
---|
| 489 | for(int i=0;i<totalTargetElement;i++) |
---|
| 490 | { |
---|
| 491 | index=targetElementIndex[i]; |
---|
| 492 | targetElements[i] = *((T*) &(targetBuffer[index][indexTargetBuffer[index]])); |
---|
| 493 | indexTargetBuffer[index]+=sizeof(T); |
---|
| 494 | } |
---|
| 495 | |
---|
| 496 | for(int i=0;i<nbTarget; i++) delete[] targetBuffer[i]; |
---|
| 497 | for(int i=0;i<nbSource; i++) delete[] sourceBuffer[i]; |
---|
| 498 | delete[] targetBuffer; |
---|
| 499 | delete[] indexTargetBuffer; |
---|
| 500 | delete[] sourceBuffer; |
---|
| 501 | delete[] indexSourceBuffer; |
---|
| 502 | delete[] request; |
---|
| 503 | delete[] status; |
---|
| 504 | } |
---|
| 505 | |
---|
| 506 | |
---|
| 507 | /* number of source and target elements is known from previous call to init() */ |
---|
| 508 | template<typename T, typename t_pack, typename t_unpack> |
---|
| 509 | void CMPIRouting::transferFromSource(T *targetElements, T *sourceElements, t_pack pack, t_unpack unpack) |
---|
| 510 | { |
---|
| 511 | char **targetBuffer = new char*[nbTarget]; |
---|
| 512 | int *indexTargetBuffer = new int[nbTarget]; |
---|
| 513 | int *targetMessageSize = new int[nbTarget]; |
---|
| 514 | int *sourceMessageSize = new int[nbSource]; |
---|
| 515 | int index; |
---|
| 516 | |
---|
| 517 | // compute target buffer size |
---|
| 518 | for (int i = 0; i < nbSource; i++) sourceMessageSize[i] = 0; |
---|
| 519 | |
---|
| 520 | for (int i = 0; i < totalSourceElement; i++) |
---|
| 521 | { |
---|
| 522 | index = sourceElementIndex[i]; |
---|
| 523 | pack(sourceElements[i], NULL, sourceMessageSize[index]); |
---|
| 524 | } |
---|
| 525 | |
---|
[1638] | 526 | ep_lib::MPI_Request *request = new ep_lib::MPI_Request[nbSource + nbTarget]; |
---|
| 527 | ep_lib::MPI_Status *status = new ep_lib::MPI_Status[nbSource + nbTarget]; |
---|
[688] | 528 | int indexRequest = 0; |
---|
| 529 | for (int i = 0; i < nbSource; i++) |
---|
| 530 | { |
---|
[1638] | 531 | ep_lib::MPI_Isend(&sourceMessageSize[i], 1, EP_INT, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 532 | indexRequest++; |
---|
| 533 | } |
---|
| 534 | for (int i = 0; i < nbTarget; i++) |
---|
| 535 | { |
---|
[1638] | 536 | ep_lib::MPI_Irecv(&targetMessageSize[i], 1, EP_INT, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 537 | indexRequest++; |
---|
| 538 | } |
---|
[1638] | 539 | ep_lib::MPI_Waitall(indexRequest, request, status); |
---|
[688] | 540 | |
---|
| 541 | for (int i = 0; i < nbTarget; i++) |
---|
| 542 | { |
---|
| 543 | targetBuffer[i] = new char[targetMessageSize[i]]; |
---|
| 544 | indexTargetBuffer[i] = 0; |
---|
| 545 | } |
---|
| 546 | |
---|
| 547 | char **sourceBuffer = new char*[nbSource]; |
---|
| 548 | int *indexSourceBuffer = new int[nbSource]; |
---|
| 549 | |
---|
| 550 | for (int i = 0; i < nbSource; i++) |
---|
| 551 | { |
---|
| 552 | sourceBuffer[i] = new char[sourceMessageSize[i]]; |
---|
| 553 | indexSourceBuffer[i] = 0; |
---|
| 554 | } |
---|
| 555 | |
---|
| 556 | |
---|
| 557 | // pack the data |
---|
| 558 | for (int i = 0; i < totalSourceElement; i++) |
---|
| 559 | { |
---|
| 560 | index = sourceElementIndex[i]; |
---|
| 561 | pack(sourceElements[i], sourceBuffer[index], indexSourceBuffer[index] ); |
---|
| 562 | } |
---|
| 563 | |
---|
| 564 | indexRequest = 0; |
---|
| 565 | for (int i = 0; i < nbSource; i++) |
---|
| 566 | { |
---|
[1638] | 567 | ep_lib::MPI_Isend(sourceBuffer[i], sourceMessageSize[i], EP_CHAR, sourceRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 568 | indexRequest++; |
---|
| 569 | } |
---|
| 570 | for (int i = 0; i < nbTarget; i++) |
---|
| 571 | { |
---|
[1638] | 572 | ep_lib::MPI_Irecv(targetBuffer[i], targetMessageSize[i], EP_CHAR, targetRank[i], 0, communicator, &request[indexRequest]); |
---|
[688] | 573 | indexRequest++; |
---|
| 574 | } |
---|
[1638] | 575 | ep_lib::MPI_Waitall(indexRequest, request, status); |
---|
[688] | 576 | |
---|
| 577 | // unpack the data |
---|
| 578 | for (int i = 0; i < totalTargetElement; i++) |
---|
| 579 | { |
---|
| 580 | index = targetElementIndex[i]; |
---|
| 581 | unpack(targetElements[i], targetBuffer[index], indexTargetBuffer[index]); |
---|
| 582 | } |
---|
| 583 | |
---|
| 584 | for (int i = 0; i < nbTarget; i++) delete[] targetBuffer[i]; |
---|
| 585 | for (int i = 0; i < nbSource; i++) delete[] sourceBuffer[i]; |
---|
| 586 | delete[] targetBuffer; |
---|
| 587 | delete[] indexTargetBuffer; |
---|
| 588 | delete[] targetMessageSize; |
---|
| 589 | delete[] sourceBuffer; |
---|
| 590 | delete[] indexSourceBuffer; |
---|
| 591 | delete[] sourceMessageSize; |
---|
| 592 | delete[] request; |
---|
| 593 | delete[] status; |
---|
| 594 | } |
---|
| 595 | |
---|
| 596 | CMPIRouting::~CMPIRouting() |
---|
| 597 | { |
---|
| 598 | }; |
---|
| 599 | |
---|
| 600 | template void CMPIRouting::init(const std::vector<int>& route, CMPICascade *cascade); |
---|
| 601 | template void CMPIRouting::init(const std::vector<vector<int> >& route, CMPICascade *cascade); |
---|
| 602 | |
---|
| 603 | template void CMPIRouting::transferToTarget(int *targetElements, int *sourceElements); |
---|
| 604 | template void CMPIRouting::transferToTarget(int *targetElements, int *sourceElements, void (*pack)(int&, char*, int&), void (* unpack)(int&, char *, int&)); |
---|
| 605 | template void CMPIRouting::transferFromSource(int *targetElements, int *sourceElements); |
---|
| 606 | template void CMPIRouting::transferFromSource(int *targetElements, int *sourceElements, void (*pack)(int&, char*, int&), void (* unpack)(int&, char *, int&)); |
---|
| 607 | |
---|
| 608 | template void CMPIRouting::transferToTarget(Node* targetElements, Node* sourceElements, void (*pack)(Node&, char*, int&), void (* unpack)(Node&, char *, int&)); |
---|
| 609 | template void CMPIRouting::transferToTarget(Elt **targetElements, Elt **sourceElements, void (*pack)(Elt *, char*, int&), void (* unpack)(Elt *, char *, int&)); |
---|
| 610 | template void CMPIRouting::transferFromSource(std::vector<int> *targetElements, std::vector<int> *sourceElements, void (*pack)(const std::vector<int>&, char*, int&), void (* unpack)(std::vector<int>&, const char *, int&)); |
---|
| 611 | |
---|
| 612 | struct NES { int cnt; int risc; int rank; }; |
---|
| 613 | |
---|
| 614 | template void alltoalls_unknown(const std::vector<std::vector<NES> >& send, std::vector<std::vector<NES> >& recv, |
---|
[1638] | 615 | const std::vector<int>& ranks, ep_lib::MPI_Comm communicator); |
---|
[688] | 616 | |
---|
| 617 | template void alltoalls_known(const std::vector<std::vector<int> >& send, std::vector<std::vector<int> >& recv, |
---|
[1638] | 618 | const std::vector<int>& ranks, ep_lib::MPI_Comm communicator); |
---|
[688] | 619 | |
---|
| 620 | } |
---|