[694] | 1 | #include "mpi.hpp" |
---|
[688] | 2 | #include <map> |
---|
| 3 | #include "cputime.hpp" /* time */ |
---|
| 4 | #include "meshutil.hpp" |
---|
| 5 | #include "polyg.hpp" |
---|
| 6 | #include "circle.hpp" |
---|
| 7 | #include "intersect.hpp" |
---|
| 8 | #include "intersection_ym.hpp" |
---|
| 9 | #include "errhandle.hpp" |
---|
| 10 | #include "mpi_routing.hpp" |
---|
| 11 | #include "gridRemap.hpp" |
---|
| 12 | |
---|
| 13 | #include "mapper.hpp" |
---|
[1328] | 14 | using namespace ep_lib; |
---|
[688] | 15 | |
---|
| 16 | namespace sphereRemap { |
---|
| 17 | |
---|
[1335] | 18 | extern CRemapGrid srcGrid; |
---|
| 19 | #pragma omp threadprivate(srcGrid) |
---|
| 20 | |
---|
| 21 | extern CRemapGrid tgtGrid; |
---|
| 22 | #pragma omp threadprivate(tgtGrid) |
---|
| 23 | |
---|
| 24 | |
---|
[688] | 25 | /* A subdivition of an array into N sub-arays |
---|
| 26 | can be represented by the length of the N arrays |
---|
| 27 | or by the offsets when each of the N arrays starts. |
---|
| 28 | This function convertes from the former to the latter. */ |
---|
| 29 | void cptOffsetsFromLengths(const int *lengths, int *offsets, int sz) |
---|
| 30 | { |
---|
[1460] | 31 | offsets[0] = 0; |
---|
| 32 | for (int i = 1; i < sz; i++) |
---|
| 33 | offsets[i] = offsets[i-1] + lengths[i-1]; |
---|
[688] | 34 | } |
---|
| 35 | |
---|
| 36 | |
---|
| 37 | void Mapper::setSourceMesh(const double* boundsLon, const double* boundsLat, int nVertex, int nbCells, const double* pole, const long int* globalId) |
---|
| 38 | { |
---|
| 39 | srcGrid.pole = Coord(pole[0], pole[1], pole[2]); |
---|
| 40 | |
---|
[1460] | 41 | int mpiRank, mpiSize; |
---|
| 42 | MPI_Comm_rank(communicator, &mpiRank); |
---|
| 43 | MPI_Comm_size(communicator, &mpiSize); |
---|
[688] | 44 | |
---|
[1460] | 45 | sourceElements.reserve(nbCells); |
---|
| 46 | sourceMesh.reserve(nbCells); |
---|
[688] | 47 | sourceGlobalId.resize(nbCells) ; |
---|
| 48 | |
---|
| 49 | if (globalId==NULL) |
---|
| 50 | { |
---|
| 51 | long int offset ; |
---|
| 52 | long int nb=nbCells ; |
---|
| 53 | MPI_Scan(&nb,&offset,1,MPI_LONG,MPI_SUM,communicator) ; |
---|
| 54 | offset=offset-nb ; |
---|
| 55 | for(int i=0;i<nbCells;i++) sourceGlobalId[i]=offset+i ; |
---|
| 56 | } |
---|
| 57 | else sourceGlobalId.assign(globalId,globalId+nbCells); |
---|
| 58 | |
---|
[1460] | 59 | for (int i = 0; i < nbCells; i++) |
---|
| 60 | { |
---|
| 61 | int offs = i*nVertex; |
---|
| 62 | Elt elt(boundsLon + offs, boundsLat + offs, nVertex); |
---|
| 63 | elt.src_id.rank = mpiRank; |
---|
| 64 | elt.src_id.ind = i; |
---|
[688] | 65 | elt.src_id.globalId = sourceGlobalId[i]; |
---|
[1460] | 66 | sourceElements.push_back(elt); |
---|
| 67 | sourceMesh.push_back(Node(elt.x, cptRadius(elt), &sourceElements.back())); |
---|
| 68 | cptEltGeom(sourceElements[i], Coord(pole[0], pole[1], pole[2])); |
---|
| 69 | } |
---|
[688] | 70 | |
---|
| 71 | } |
---|
| 72 | |
---|
| 73 | void Mapper::setTargetMesh(const double* boundsLon, const double* boundsLat, int nVertex, int nbCells, const double* pole, const long int* globalId) |
---|
| 74 | { |
---|
[1328] | 75 | tgtGrid.pole = Coord(pole[0], pole[1], pole[2]); |
---|
[688] | 76 | |
---|
[1460] | 77 | int mpiRank, mpiSize; |
---|
| 78 | MPI_Comm_rank(communicator, &mpiRank); |
---|
| 79 | MPI_Comm_size(communicator, &mpiSize); |
---|
[688] | 80 | |
---|
[1460] | 81 | targetElements.reserve(nbCells); |
---|
| 82 | targetMesh.reserve(nbCells); |
---|
[688] | 83 | |
---|
[1328] | 84 | targetGlobalId.resize(nbCells) ; |
---|
| 85 | if (globalId==NULL) |
---|
| 86 | { |
---|
| 87 | long int offset ; |
---|
| 88 | long int nb=nbCells ; |
---|
| 89 | MPI_Scan(&nb,&offset,1,MPI_LONG,MPI_SUM,communicator) ; |
---|
| 90 | offset=offset-nb ; |
---|
| 91 | for(int i=0;i<nbCells;i++) targetGlobalId[i]=offset+i ; |
---|
| 92 | } |
---|
| 93 | else targetGlobalId.assign(globalId,globalId+nbCells); |
---|
[688] | 94 | |
---|
[1460] | 95 | for (int i = 0; i < nbCells; i++) |
---|
| 96 | { |
---|
| 97 | int offs = i*nVertex; |
---|
| 98 | Elt elt(boundsLon + offs, boundsLat + offs, nVertex); |
---|
| 99 | targetElements.push_back(elt); |
---|
| 100 | targetMesh.push_back(Node(elt.x, cptRadius(elt), &sourceElements.back())); |
---|
| 101 | cptEltGeom(targetElements[i], Coord(pole[0], pole[1], pole[2])); |
---|
| 102 | } |
---|
[688] | 103 | |
---|
| 104 | |
---|
| 105 | } |
---|
| 106 | |
---|
| 107 | void Mapper::setSourceValue(const double* val) |
---|
| 108 | { |
---|
[1328] | 109 | int size=sourceElements.size() ; |
---|
| 110 | for(int i=0;i<size;++i) sourceElements[i].val=val[i] ; |
---|
[688] | 111 | } |
---|
| 112 | |
---|
| 113 | void Mapper::getTargetValue(double* val) |
---|
| 114 | { |
---|
[1328] | 115 | int size=targetElements.size() ; |
---|
| 116 | for(int i=0;i<size;++i) val[i]=targetElements[i].val ; |
---|
[688] | 117 | } |
---|
| 118 | |
---|
[1114] | 119 | vector<double> Mapper::computeWeights(int interpOrder, bool renormalize, bool quantity) |
---|
[688] | 120 | { |
---|
[1460] | 121 | vector<double> timings; |
---|
| 122 | int mpiSize, mpiRank; |
---|
| 123 | MPI_Comm_size(communicator, &mpiSize); |
---|
| 124 | MPI_Comm_rank(communicator, &mpiRank); |
---|
[688] | 125 | |
---|
[1328] | 126 | this->buildSSTree(sourceMesh, targetMesh); |
---|
[688] | 127 | |
---|
[1460] | 128 | if (mpiRank == 0 && verbose) cout << "Computing intersections ..." << endl; |
---|
| 129 | double tic = cputime(); |
---|
| 130 | computeIntersection(&targetElements[0], targetElements.size()); |
---|
| 131 | timings.push_back(cputime() - tic); |
---|
[688] | 132 | |
---|
[1460] | 133 | tic = cputime(); |
---|
| 134 | if (interpOrder == 2) { |
---|
| 135 | if (mpiRank == 0 && verbose) cout << "Computing grads ..." << endl; |
---|
| 136 | buildMeshTopology(); |
---|
| 137 | computeGrads(); |
---|
| 138 | } |
---|
| 139 | timings.push_back(cputime() - tic); |
---|
[688] | 140 | |
---|
[1460] | 141 | /* Prepare computation of weights */ |
---|
| 142 | /* compute number of intersections which for the first order case |
---|
| 143 | corresponds to the number of edges in the remap matrix */ |
---|
| 144 | int nIntersections = 0; |
---|
| 145 | for (int j = 0; j < targetElements.size(); j++) |
---|
| 146 | { |
---|
| 147 | Elt &elt = targetElements[j]; |
---|
| 148 | for (list<Polyg*>::iterator it = elt.is.begin(); it != elt.is.end(); it++) |
---|
| 149 | nIntersections++; |
---|
| 150 | } |
---|
| 151 | /* overallocate for NMAX neighbours for each elements */ |
---|
| 152 | remapMatrix = new double[nIntersections*NMAX]; |
---|
| 153 | srcAddress = new int[nIntersections*NMAX]; |
---|
| 154 | srcRank = new int[nIntersections*NMAX]; |
---|
| 155 | dstAddress = new int[nIntersections*NMAX]; |
---|
[1328] | 156 | sourceWeightId =new long[nIntersections*NMAX]; |
---|
| 157 | targetWeightId =new long[nIntersections*NMAX]; |
---|
[688] | 158 | |
---|
| 159 | |
---|
[1460] | 160 | if (mpiRank == 0 && verbose) cout << "Remapping..." << endl; |
---|
| 161 | tic = cputime(); |
---|
| 162 | nWeights = remap(&targetElements[0], targetElements.size(), interpOrder, renormalize, quantity); |
---|
| 163 | timings.push_back(cputime() - tic); |
---|
[688] | 164 | |
---|
[1328] | 165 | for (int i = 0; i < targetElements.size(); i++) targetElements[i].delete_intersections(); |
---|
| 166 | |
---|
[1460] | 167 | return timings; |
---|
[688] | 168 | } |
---|
| 169 | |
---|
| 170 | /** |
---|
[1328] | 171 | @param elements are cells of the target grid that are distributed over CPUs |
---|
| 172 | indepentently of the distribution of the SS-tree. |
---|
| 173 | @param nbElements is the size of the elements array. |
---|
| 174 | @param order is the order of interpolaton (must be 1 or 2). |
---|
| 175 | */ |
---|
[1114] | 176 | int Mapper::remap(Elt *elements, int nbElements, int order, bool renormalize, bool quantity) |
---|
[688] | 177 | { |
---|
[1460] | 178 | int mpiSize, mpiRank; |
---|
| 179 | MPI_Comm_size(communicator, &mpiSize); |
---|
| 180 | MPI_Comm_rank(communicator, &mpiRank); |
---|
[688] | 181 | |
---|
[1460] | 182 | /* create list of intersections (super mesh elements) for each rank */ |
---|
| 183 | multimap<int, Polyg *> *elementList = new multimap<int, Polyg *>[mpiSize]; |
---|
| 184 | for (int j = 0; j < nbElements; j++) |
---|
| 185 | { |
---|
| 186 | Elt& e = elements[j]; |
---|
| 187 | for (list<Polyg *>::iterator it = e.is.begin(); it != e.is.end(); it++) |
---|
| 188 | elementList[(*it)->id.rank].insert(pair<int, Polyg *>((*it)->id.ind, *it)); |
---|
| 189 | } |
---|
[688] | 190 | |
---|
[1460] | 191 | int *nbSendElement = new int[mpiSize]; |
---|
| 192 | int **sendElement = new int*[mpiSize]; /* indices of elements required from other rank */ |
---|
| 193 | double **recvValue = new double*[mpiSize]; |
---|
| 194 | double **recvArea = new double*[mpiSize]; |
---|
| 195 | Coord **recvGrad = new Coord*[mpiSize]; |
---|
| 196 | GloId **recvNeighIds = new GloId*[mpiSize]; /* ids of the of the source neighbours which also contribute through gradient */ |
---|
| 197 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 198 | { |
---|
| 199 | /* get size for allocation */ |
---|
| 200 | int last = -1; /* compares unequal to any index */ |
---|
| 201 | int index = -1; /* increased to starting index 0 in first iteration */ |
---|
| 202 | for (multimap<int, Polyg *>::iterator it = elementList[rank].begin(); it != elementList[rank].end(); ++it) |
---|
| 203 | { |
---|
| 204 | if (last != it->first) |
---|
| 205 | index++; |
---|
| 206 | (it->second)->id.ind = index; |
---|
| 207 | last = it->first; |
---|
| 208 | } |
---|
| 209 | nbSendElement[rank] = index + 1; |
---|
[688] | 210 | |
---|
[1460] | 211 | /* if size is non-zero allocate and collect indices of elements on other ranks that we intersect */ |
---|
| 212 | if (nbSendElement[rank] > 0) |
---|
| 213 | { |
---|
| 214 | sendElement[rank] = new int[nbSendElement[rank]]; |
---|
| 215 | recvValue[rank] = new double[nbSendElement[rank]]; |
---|
| 216 | recvArea[rank] = new double[nbSendElement[rank]]; |
---|
| 217 | if (order == 2) |
---|
| 218 | { |
---|
| 219 | recvNeighIds[rank] = new GloId[nbSendElement[rank]*(NMAX+1)]; |
---|
| 220 | recvGrad[rank] = new Coord[nbSendElement[rank]*(NMAX+1)]; |
---|
| 221 | } |
---|
| 222 | else |
---|
| 223 | recvNeighIds[rank] = new GloId[nbSendElement[rank]]; |
---|
[688] | 224 | |
---|
[1460] | 225 | last = -1; |
---|
| 226 | index = -1; |
---|
| 227 | for (multimap<int, Polyg *>::iterator it = elementList[rank].begin(); it != elementList[rank].end(); ++it) |
---|
| 228 | { |
---|
| 229 | if (last != it->first) |
---|
| 230 | index++; |
---|
| 231 | sendElement[rank][index] = it->first; |
---|
| 232 | last = it->first; |
---|
| 233 | } |
---|
| 234 | } |
---|
| 235 | } |
---|
[688] | 236 | |
---|
[1460] | 237 | /* communicate sizes of source elements to be sent (index lists and later values and gradients) */ |
---|
| 238 | int *nbRecvElement = new int[mpiSize]; |
---|
| 239 | MPI_Alltoall(nbSendElement, 1, MPI_INT, nbRecvElement, 1, MPI_INT, communicator); |
---|
[688] | 240 | |
---|
[1460] | 241 | /* communicate indices of source elements on other ranks whoes value and gradient we need (since intersection) */ |
---|
| 242 | int nbSendRequest = 0; |
---|
| 243 | int nbRecvRequest = 0; |
---|
| 244 | int **recvElement = new int*[mpiSize]; |
---|
| 245 | double **sendValue = new double*[mpiSize]; |
---|
| 246 | double **sendArea = new double*[mpiSize]; |
---|
| 247 | Coord **sendGrad = new Coord*[mpiSize]; |
---|
| 248 | GloId **sendNeighIds = new GloId*[mpiSize]; |
---|
| 249 | MPI_Request *sendRequest = new MPI_Request[4*mpiSize]; |
---|
| 250 | MPI_Request *recvRequest = new MPI_Request[4*mpiSize]; |
---|
| 251 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 252 | { |
---|
| 253 | if (nbSendElement[rank] > 0) |
---|
| 254 | { |
---|
| 255 | MPI_Issend(sendElement[rank], nbSendElement[rank], MPI_INT, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 256 | nbSendRequest++; |
---|
| 257 | } |
---|
[688] | 258 | |
---|
[1460] | 259 | if (nbRecvElement[rank] > 0) |
---|
| 260 | { |
---|
| 261 | recvElement[rank] = new int[nbRecvElement[rank]]; |
---|
| 262 | sendValue[rank] = new double[nbRecvElement[rank]]; |
---|
| 263 | sendArea[rank] = new double[nbRecvElement[rank]]; |
---|
| 264 | if (order == 2) |
---|
| 265 | { |
---|
| 266 | sendNeighIds[rank] = new GloId[nbRecvElement[rank]*(NMAX+1)]; |
---|
| 267 | sendGrad[rank] = new Coord[nbRecvElement[rank]*(NMAX+1)]; |
---|
| 268 | } |
---|
| 269 | else |
---|
| 270 | { |
---|
| 271 | sendNeighIds[rank] = new GloId[nbRecvElement[rank]]; |
---|
| 272 | } |
---|
| 273 | MPI_Irecv(recvElement[rank], nbRecvElement[rank], MPI_INT, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 274 | nbRecvRequest++; |
---|
| 275 | } |
---|
| 276 | } |
---|
| 277 | MPI_Status *status = new MPI_Status[4*mpiSize]; |
---|
| 278 | |
---|
| 279 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
| 280 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
[688] | 281 | |
---|
[1460] | 282 | /* for all indices that have been received from requesting ranks: pack values and gradients, then send */ |
---|
| 283 | nbSendRequest = 0; |
---|
| 284 | nbRecvRequest = 0; |
---|
| 285 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 286 | { |
---|
| 287 | if (nbRecvElement[rank] > 0) |
---|
| 288 | { |
---|
| 289 | int jj = 0; // jj == j if no weight writing |
---|
| 290 | for (int j = 0; j < nbRecvElement[rank]; j++) |
---|
| 291 | { |
---|
| 292 | sendValue[rank][j] = sstree.localElements[recvElement[rank][j]].val; |
---|
| 293 | sendArea[rank][j] = sstree.localElements[recvElement[rank][j]].area; |
---|
| 294 | if (order == 2) |
---|
| 295 | { |
---|
| 296 | sendGrad[rank][jj] = sstree.localElements[recvElement[rank][j]].grad; |
---|
| 297 | // cout<<"grad "<<jj<<" "<<recvElement[rank][j]<<" "<<sendGrad[rank][jj]<<" "<<sstree.localElements[recvElement[rank][j]].grad<<endl ; |
---|
| 298 | sendNeighIds[rank][jj] = sstree.localElements[recvElement[rank][j]].src_id; |
---|
| 299 | jj++; |
---|
| 300 | for (int i = 0; i < NMAX; i++) |
---|
| 301 | { |
---|
| 302 | sendGrad[rank][jj] = sstree.localElements[recvElement[rank][j]].gradNeigh[i]; |
---|
| 303 | // cout<<"grad "<<jj<<" "<<sendGrad[rank][jj]<<" "<<sstree.localElements[recvElement[rank][j]].grad<<endl ; |
---|
[1328] | 304 | sendNeighIds[rank][jj] = sstree.localElements[recvElement[rank][j]].neighId[i]; |
---|
[1460] | 305 | jj++; |
---|
| 306 | } |
---|
| 307 | } |
---|
| 308 | else |
---|
| 309 | sendNeighIds[rank][j] = sstree.localElements[recvElement[rank][j]].src_id; |
---|
| 310 | } |
---|
| 311 | MPI_Issend(sendValue[rank], nbRecvElement[rank], MPI_DOUBLE, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 312 | nbSendRequest++; |
---|
[1463] | 313 | MPI_Issend(sendArea[rank], nbRecvElement[rank], MPI_DOUBLE, rank, 1, communicator, &sendRequest[nbSendRequest]); |
---|
[1460] | 314 | nbSendRequest++; |
---|
| 315 | if (order == 2) |
---|
| 316 | { |
---|
| 317 | MPI_Issend(sendGrad[rank], 3*nbRecvElement[rank]*(NMAX+1), |
---|
[1463] | 318 | MPI_DOUBLE, rank, 2, communicator, &sendRequest[nbSendRequest]); |
---|
[1460] | 319 | nbSendRequest++; |
---|
[1463] | 320 | MPI_Issend(sendNeighIds[rank], 4*nbRecvElement[rank]*(NMAX+1), MPI_INT, rank, 3, communicator, &sendRequest[nbSendRequest]); |
---|
[1328] | 321 | //ym --> attention taille GloId |
---|
[1460] | 322 | nbSendRequest++; |
---|
| 323 | } |
---|
| 324 | else |
---|
| 325 | { |
---|
[1463] | 326 | MPI_Issend(sendNeighIds[rank], 4*nbRecvElement[rank], MPI_INT, rank, 4, communicator, &sendRequest[nbSendRequest]); |
---|
[1328] | 327 | //ym --> attention taille GloId |
---|
[1460] | 328 | nbSendRequest++; |
---|
| 329 | } |
---|
| 330 | } |
---|
| 331 | if (nbSendElement[rank] > 0) |
---|
| 332 | { |
---|
| 333 | MPI_Irecv(recvValue[rank], nbSendElement[rank], MPI_DOUBLE, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 334 | nbRecvRequest++; |
---|
[1463] | 335 | MPI_Irecv(recvArea[rank], nbSendElement[rank], MPI_DOUBLE, rank, 1, communicator, &recvRequest[nbRecvRequest]); |
---|
[1460] | 336 | nbRecvRequest++; |
---|
| 337 | if (order == 2) |
---|
| 338 | { |
---|
| 339 | MPI_Irecv(recvGrad[rank], 3*nbSendElement[rank]*(NMAX+1), |
---|
[1463] | 340 | MPI_DOUBLE, rank, 2, communicator, &recvRequest[nbRecvRequest]); |
---|
[1460] | 341 | nbRecvRequest++; |
---|
[1463] | 342 | MPI_Irecv(recvNeighIds[rank], 4*nbSendElement[rank]*(NMAX+1), MPI_INT, rank, 3, communicator, &recvRequest[nbRecvRequest]); |
---|
[1328] | 343 | //ym --> attention taille GloId |
---|
[1460] | 344 | nbRecvRequest++; |
---|
| 345 | } |
---|
| 346 | else |
---|
| 347 | { |
---|
[1463] | 348 | MPI_Irecv(recvNeighIds[rank], 4*nbSendElement[rank], MPI_INT, rank, 4, communicator, &recvRequest[nbRecvRequest]); |
---|
[1328] | 349 | //ym --> attention taille GloId |
---|
[1460] | 350 | nbRecvRequest++; |
---|
| 351 | } |
---|
| 352 | } |
---|
| 353 | } |
---|
[1328] | 354 | |
---|
[1460] | 355 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
| 356 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
[1328] | 357 | |
---|
[688] | 358 | |
---|
[1460] | 359 | /* now that all values and gradients are available use them to computed interpolated values on target |
---|
| 360 | and also to compute weights */ |
---|
| 361 | int i = 0; |
---|
| 362 | for (int j = 0; j < nbElements; j++) |
---|
| 363 | { |
---|
| 364 | Elt& e = elements[j]; |
---|
[688] | 365 | |
---|
[1460] | 366 | /* since for the 2nd order case source grid elements can contribute to a destination grid element over several "paths" |
---|
| 367 | (step1: gradient is computed using neighbours on same grid, step2: intersection uses several elements on other grid) |
---|
| 368 | accumulate them so that there is only one final weight between two elements */ |
---|
| 369 | map<GloId,double> wgt_map; |
---|
[688] | 370 | |
---|
[1460] | 371 | /* for destination element `e` loop over all intersetions/the corresponding source elements */ |
---|
| 372 | for (list<Polyg *>::iterator it = e.is.begin(); it != e.is.end(); it++) |
---|
| 373 | { |
---|
| 374 | /* it is the intersection element, so it->x and it->area are barycentre and area of intersection element (super mesh) |
---|
| 375 | but it->id is id of the source element that it intersects */ |
---|
| 376 | int n1 = (*it)->id.ind; |
---|
| 377 | int rank = (*it)->id.rank; |
---|
| 378 | double fk = recvValue[rank][n1]; |
---|
| 379 | double srcArea = recvArea[rank][n1]; |
---|
| 380 | double w = (*it)->area; |
---|
[1328] | 381 | if (quantity) w/=srcArea ; |
---|
[688] | 382 | |
---|
[1460] | 383 | /* first order: src value times weight (weight = supermesh area), later divide by target area */ |
---|
| 384 | int kk = (order == 2) ? n1 * (NMAX + 1) : n1; |
---|
| 385 | GloId neighID = recvNeighIds[rank][kk]; |
---|
| 386 | wgt_map[neighID] += w; |
---|
[688] | 387 | |
---|
[1460] | 388 | if (order == 2) |
---|
| 389 | { |
---|
| 390 | for (int k = 0; k < NMAX+1; k++) |
---|
| 391 | { |
---|
| 392 | int kk = n1 * (NMAX + 1) + k; |
---|
| 393 | GloId neighID = recvNeighIds[rank][kk]; |
---|
| 394 | if (neighID.ind != -1) wgt_map[neighID] += w * scalarprod(recvGrad[rank][kk], (*it)->x); |
---|
| 395 | } |
---|
[844] | 396 | |
---|
[1460] | 397 | } |
---|
| 398 | } |
---|
[844] | 399 | |
---|
[1328] | 400 | double renorm=0; |
---|
| 401 | if (renormalize) |
---|
| 402 | for (map<GloId,double>::iterator it = wgt_map.begin(); it != wgt_map.end(); it++) renorm+=it->second / e.area; |
---|
| 403 | else renorm=1. ; |
---|
[688] | 404 | |
---|
[1328] | 405 | for (map<GloId,double>::iterator it = wgt_map.begin(); it != wgt_map.end(); it++) |
---|
[1460] | 406 | { |
---|
[1328] | 407 | if (quantity) this->remapMatrix[i] = (it->second ) / renorm; |
---|
[1460] | 408 | else this->remapMatrix[i] = (it->second / e.area) / renorm; |
---|
| 409 | this->srcAddress[i] = it->first.ind; |
---|
| 410 | this->srcRank[i] = it->first.rank; |
---|
| 411 | this->dstAddress[i] = j; |
---|
[1328] | 412 | this->sourceWeightId[i]= it->first.globalId ; |
---|
| 413 | this->targetWeightId[i]= targetGlobalId[j] ; |
---|
[1460] | 414 | i++; |
---|
| 415 | } |
---|
| 416 | } |
---|
[1463] | 417 | |
---|
| 418 | MPI_Barrier(communicator); |
---|
[1328] | 419 | |
---|
[1460] | 420 | /* free all memory allocated in this function */ |
---|
| 421 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 422 | { |
---|
| 423 | if (nbSendElement[rank] > 0) |
---|
| 424 | { |
---|
| 425 | delete[] sendElement[rank]; |
---|
| 426 | delete[] recvValue[rank]; |
---|
| 427 | delete[] recvArea[rank]; |
---|
| 428 | if (order == 2) |
---|
| 429 | { |
---|
| 430 | delete[] recvGrad[rank]; |
---|
| 431 | } |
---|
| 432 | delete[] recvNeighIds[rank]; |
---|
| 433 | } |
---|
| 434 | if (nbRecvElement[rank] > 0) |
---|
| 435 | { |
---|
| 436 | delete[] recvElement[rank]; |
---|
| 437 | delete[] sendValue[rank]; |
---|
| 438 | delete[] sendArea[rank]; |
---|
| 439 | if (order == 2) |
---|
| 440 | delete[] sendGrad[rank]; |
---|
| 441 | delete[] sendNeighIds[rank]; |
---|
| 442 | } |
---|
| 443 | } |
---|
| 444 | delete[] status; |
---|
| 445 | delete[] sendRequest; |
---|
| 446 | delete[] recvRequest; |
---|
| 447 | delete[] elementList; |
---|
| 448 | delete[] nbSendElement; |
---|
| 449 | delete[] nbRecvElement; |
---|
| 450 | delete[] sendElement; |
---|
| 451 | delete[] recvElement; |
---|
| 452 | delete[] sendValue; |
---|
| 453 | delete[] recvValue; |
---|
| 454 | delete[] sendGrad; |
---|
| 455 | delete[] recvGrad; |
---|
| 456 | delete[] sendNeighIds; |
---|
| 457 | delete[] recvNeighIds; |
---|
| 458 | return i; |
---|
[688] | 459 | } |
---|
| 460 | |
---|
| 461 | void Mapper::computeGrads() |
---|
| 462 | { |
---|
[1328] | 463 | /* array of pointers to collect local elements and elements received from other cpu */ |
---|
| 464 | vector<Elt*> globalElements(sstree.nbLocalElements + nbNeighbourElements); |
---|
| 465 | int index = 0; |
---|
| 466 | for (int i = 0; i < sstree.nbLocalElements; i++, index++) |
---|
| 467 | globalElements[index] = &(sstree.localElements[i]); |
---|
| 468 | for (int i = 0; i < nbNeighbourElements; i++, index++) |
---|
| 469 | globalElements[index] = &neighbourElements[i]; |
---|
[688] | 470 | |
---|
[1328] | 471 | update_baryc(sstree.localElements, sstree.nbLocalElements); |
---|
| 472 | computeGradients(&globalElements[0], sstree.nbLocalElements); |
---|
[688] | 473 | } |
---|
| 474 | |
---|
| 475 | /** for each element of the source grid, finds all the neighbouring elements that share an edge |
---|
[1328] | 476 | (filling array neighbourElements). This is used later to compute gradients */ |
---|
[688] | 477 | void Mapper::buildMeshTopology() |
---|
| 478 | { |
---|
[1328] | 479 | int mpiSize, mpiRank; |
---|
| 480 | MPI_Comm_size(communicator, &mpiSize); |
---|
| 481 | MPI_Comm_rank(communicator, &mpiRank); |
---|
[688] | 482 | |
---|
[1328] | 483 | vector<Node> *routingList = new vector<Node>[mpiSize]; |
---|
| 484 | vector<vector<int> > routes(sstree.localTree.leafs.size()); |
---|
[688] | 485 | |
---|
[1328] | 486 | sstree.routeIntersections(routes, sstree.localTree.leafs); |
---|
[688] | 487 | |
---|
[1328] | 488 | for (int i = 0; i < routes.size(); ++i) |
---|
| 489 | for (int k = 0; k < routes[i].size(); ++k) |
---|
| 490 | routingList[routes[i][k]].push_back(sstree.localTree.leafs[i]); |
---|
| 491 | routingList[mpiRank].clear(); |
---|
[688] | 492 | |
---|
| 493 | |
---|
[1328] | 494 | CMPIRouting mpiRoute(communicator); |
---|
| 495 | mpiRoute.init(routes); |
---|
| 496 | int nRecv = mpiRoute.getTotalSourceElement(); |
---|
[1460] | 497 | // cout << mpiRank << " NRECV " << nRecv << "(" << routes.size() << ")"<< endl; |
---|
[688] | 498 | |
---|
[1328] | 499 | int *nbSendNode = new int[mpiSize]; |
---|
| 500 | int *nbRecvNode = new int[mpiSize]; |
---|
| 501 | int *sendMessageSize = new int[mpiSize]; |
---|
| 502 | int *recvMessageSize = new int[mpiSize]; |
---|
[688] | 503 | |
---|
[1328] | 504 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 505 | { |
---|
| 506 | nbSendNode[rank] = routingList[rank].size(); |
---|
| 507 | sendMessageSize[rank] = 0; |
---|
| 508 | for (size_t j = 0; j < routingList[rank].size(); j++) |
---|
| 509 | { |
---|
| 510 | Elt *elt = (Elt *) (routingList[rank][j].data); |
---|
| 511 | sendMessageSize[rank] += packedPolygonSize(*elt); |
---|
| 512 | } |
---|
| 513 | } |
---|
[688] | 514 | |
---|
[1328] | 515 | MPI_Alltoall(nbSendNode, 1, MPI_INT, nbRecvNode, 1, MPI_INT, communicator); |
---|
| 516 | MPI_Alltoall(sendMessageSize, 1, MPI_INT, recvMessageSize, 1, MPI_INT, communicator); |
---|
[688] | 517 | |
---|
[1328] | 518 | char **sendBuffer = new char*[mpiSize]; |
---|
| 519 | char **recvBuffer = new char*[mpiSize]; |
---|
| 520 | int *pos = new int[mpiSize]; |
---|
[688] | 521 | |
---|
[1328] | 522 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 523 | { |
---|
| 524 | if (nbSendNode[rank] > 0) sendBuffer[rank] = new char[sendMessageSize[rank]]; |
---|
| 525 | if (nbRecvNode[rank] > 0) recvBuffer[rank] = new char[recvMessageSize[rank]]; |
---|
| 526 | } |
---|
[688] | 527 | |
---|
[1328] | 528 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 529 | { |
---|
| 530 | pos[rank] = 0; |
---|
| 531 | for (size_t j = 0; j < routingList[rank].size(); j++) |
---|
| 532 | { |
---|
| 533 | Elt *elt = (Elt *) (routingList[rank][j].data); |
---|
| 534 | packPolygon(*elt, sendBuffer[rank], pos[rank]); |
---|
| 535 | } |
---|
| 536 | } |
---|
| 537 | delete [] routingList; |
---|
[688] | 538 | |
---|
| 539 | |
---|
[1328] | 540 | int nbSendRequest = 0; |
---|
| 541 | int nbRecvRequest = 0; |
---|
| 542 | MPI_Request *sendRequest = new MPI_Request[mpiSize]; |
---|
| 543 | MPI_Request *recvRequest = new MPI_Request[mpiSize]; |
---|
| 544 | MPI_Status *status = new MPI_Status[mpiSize]; |
---|
[688] | 545 | |
---|
[1328] | 546 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 547 | { |
---|
| 548 | if (nbSendNode[rank] > 0) |
---|
| 549 | { |
---|
| 550 | MPI_Issend(sendBuffer[rank], sendMessageSize[rank], MPI_CHAR, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 551 | nbSendRequest++; |
---|
| 552 | } |
---|
| 553 | if (nbRecvNode[rank] > 0) |
---|
| 554 | { |
---|
| 555 | MPI_Irecv(recvBuffer[rank], recvMessageSize[rank], MPI_CHAR, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 556 | nbRecvRequest++; |
---|
| 557 | } |
---|
| 558 | } |
---|
[688] | 559 | |
---|
[1460] | 560 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
| 561 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
[688] | 562 | |
---|
[1328] | 563 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 564 | if (nbSendNode[rank] > 0) delete [] sendBuffer[rank]; |
---|
| 565 | delete [] sendBuffer; |
---|
[688] | 566 | |
---|
[1328] | 567 | char **sendBuffer2 = new char*[mpiSize]; |
---|
| 568 | char **recvBuffer2 = new char*[mpiSize]; |
---|
[688] | 569 | |
---|
[1328] | 570 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 571 | { |
---|
| 572 | nbSendNode[rank] = 0; |
---|
| 573 | sendMessageSize[rank] = 0; |
---|
[688] | 574 | |
---|
[1328] | 575 | if (nbRecvNode[rank] > 0) |
---|
| 576 | { |
---|
| 577 | set<NodePtr> neighbourList; |
---|
| 578 | pos[rank] = 0; |
---|
| 579 | for (int j = 0; j < nbRecvNode[rank]; j++) |
---|
| 580 | { |
---|
| 581 | Elt elt; |
---|
| 582 | unpackPolygon(elt, recvBuffer[rank], pos[rank]); |
---|
| 583 | Node node(elt.x, cptRadius(elt), &elt); |
---|
| 584 | findNeighbour(sstree.localTree.root, &node, neighbourList); |
---|
| 585 | } |
---|
| 586 | nbSendNode[rank] = neighbourList.size(); |
---|
| 587 | for (set<NodePtr>::iterator it = neighbourList.begin(); it != neighbourList.end(); it++) |
---|
| 588 | { |
---|
| 589 | Elt *elt = (Elt *) ((*it)->data); |
---|
| 590 | sendMessageSize[rank] += packedPolygonSize(*elt); |
---|
| 591 | } |
---|
[688] | 592 | |
---|
[1328] | 593 | sendBuffer2[rank] = new char[sendMessageSize[rank]]; |
---|
| 594 | pos[rank] = 0; |
---|
[688] | 595 | |
---|
[1328] | 596 | for (set<NodePtr>::iterator it = neighbourList.begin(); it != neighbourList.end(); it++) |
---|
| 597 | { |
---|
| 598 | Elt *elt = (Elt *) ((*it)->data); |
---|
| 599 | packPolygon(*elt, sendBuffer2[rank], pos[rank]); |
---|
| 600 | } |
---|
| 601 | } |
---|
| 602 | } |
---|
| 603 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 604 | if (nbRecvNode[rank] > 0) delete [] recvBuffer[rank]; |
---|
| 605 | delete [] recvBuffer; |
---|
[688] | 606 | |
---|
| 607 | |
---|
[1328] | 608 | MPI_Barrier(communicator); |
---|
| 609 | MPI_Alltoall(nbSendNode, 1, MPI_INT, nbRecvNode, 1, MPI_INT, communicator); |
---|
| 610 | MPI_Alltoall(sendMessageSize, 1, MPI_INT, recvMessageSize, 1, MPI_INT, communicator); |
---|
[688] | 611 | |
---|
[1328] | 612 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 613 | if (nbRecvNode[rank] > 0) recvBuffer2[rank] = new char[recvMessageSize[rank]]; |
---|
[688] | 614 | |
---|
[1328] | 615 | nbSendRequest = 0; |
---|
| 616 | nbRecvRequest = 0; |
---|
[688] | 617 | |
---|
[1328] | 618 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 619 | { |
---|
| 620 | if (nbSendNode[rank] > 0) |
---|
| 621 | { |
---|
| 622 | MPI_Issend(sendBuffer2[rank], sendMessageSize[rank], MPI_CHAR, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 623 | nbSendRequest++; |
---|
| 624 | } |
---|
| 625 | if (nbRecvNode[rank] > 0) |
---|
| 626 | { |
---|
| 627 | MPI_Irecv(recvBuffer2[rank], recvMessageSize[rank], MPI_CHAR, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 628 | nbRecvRequest++; |
---|
| 629 | } |
---|
| 630 | } |
---|
[1205] | 631 | |
---|
[1328] | 632 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
| 633 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
[688] | 634 | |
---|
[1328] | 635 | int nbNeighbourNodes = 0; |
---|
| 636 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 637 | nbNeighbourNodes += nbRecvNode[rank]; |
---|
[688] | 638 | |
---|
[1328] | 639 | neighbourElements = new Elt[nbNeighbourNodes]; |
---|
| 640 | nbNeighbourElements = nbNeighbourNodes; |
---|
[688] | 641 | |
---|
[1328] | 642 | int index = 0; |
---|
| 643 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 644 | { |
---|
| 645 | pos[rank] = 0; |
---|
| 646 | for (int j = 0; j < nbRecvNode[rank]; j++) |
---|
| 647 | { |
---|
| 648 | unpackPolygon(neighbourElements[index], recvBuffer2[rank], pos[rank]); |
---|
| 649 | neighbourElements[index].id.ind = sstree.localTree.leafs.size() + index; |
---|
| 650 | index++; |
---|
| 651 | } |
---|
| 652 | } |
---|
| 653 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 654 | { |
---|
| 655 | if (nbRecvNode[rank] > 0) delete [] recvBuffer2[rank]; |
---|
| 656 | if (nbSendNode[rank] > 0) delete [] sendBuffer2[rank]; |
---|
| 657 | } |
---|
| 658 | delete [] recvBuffer2; |
---|
| 659 | delete [] sendBuffer2; |
---|
| 660 | delete [] sendMessageSize; |
---|
| 661 | delete [] recvMessageSize; |
---|
| 662 | delete [] nbSendNode; |
---|
| 663 | delete [] nbRecvNode; |
---|
| 664 | delete [] sendRequest; |
---|
| 665 | delete [] recvRequest; |
---|
| 666 | delete [] status; |
---|
| 667 | delete [] pos; |
---|
[688] | 668 | |
---|
[1328] | 669 | /* re-compute on received elements to avoid having to send this information */ |
---|
| 670 | neighbourNodes.resize(nbNeighbourNodes); |
---|
| 671 | setCirclesAndLinks(neighbourElements, neighbourNodes); |
---|
| 672 | cptAllEltsGeom(neighbourElements, nbNeighbourNodes, srcGrid.pole); |
---|
[688] | 673 | |
---|
[1328] | 674 | /* the local SS tree must include nodes from other cpus if they are potential |
---|
| 675 | intersector of a local node */ |
---|
| 676 | sstree.localTree.insertNodes(neighbourNodes); |
---|
[688] | 677 | |
---|
[1328] | 678 | /* for every local element, |
---|
| 679 | use the SS-tree to find all elements (including neighbourElements) |
---|
| 680 | who are potential neighbours because their circles intersect, |
---|
| 681 | then check all canditates for common edges to build up connectivity information |
---|
| 682 | */ |
---|
| 683 | for (int j = 0; j < sstree.localTree.leafs.size(); j++) |
---|
| 684 | { |
---|
| 685 | Node& node = sstree.localTree.leafs[j]; |
---|
[688] | 686 | |
---|
[1328] | 687 | /* find all leafs whoes circles that intersect node's circle and save into node->intersectors */ |
---|
| 688 | node.search(sstree.localTree.root); |
---|
[688] | 689 | |
---|
[1328] | 690 | Elt *elt = (Elt *)(node.data); |
---|
[688] | 691 | |
---|
[1328] | 692 | for (int i = 0; i < elt->n; i++) elt->neighbour[i] = NOT_FOUND; |
---|
[688] | 693 | |
---|
[1328] | 694 | /* for element `elt` loop through all nodes in the SS-tree |
---|
| 695 | whoes circles intersect with the circle around `elt` (the SS intersectors) |
---|
| 696 | and check if they are neighbours in the sense that the two elements share an edge. |
---|
| 697 | If they do, save this information for elt */ |
---|
| 698 | for (list<NodePtr>::iterator it = (node.intersectors).begin(); it != (node.intersectors).end(); ++it) |
---|
| 699 | { |
---|
| 700 | Elt *elt2 = (Elt *)((*it)->data); |
---|
| 701 | set_neighbour(*elt, *elt2); |
---|
| 702 | } |
---|
| 703 | |
---|
[1460] | 704 | /* |
---|
| 705 | for (int i = 0; i < elt->n; i++) |
---|
| 706 | { |
---|
| 707 | if (elt->neighbour[i] == NOT_FOUND) |
---|
| 708 | error_exit("neighbour not found"); |
---|
| 709 | } |
---|
| 710 | */ |
---|
[1328] | 711 | } |
---|
[688] | 712 | } |
---|
| 713 | |
---|
| 714 | /** @param elements are the target grid elements */ |
---|
| 715 | void Mapper::computeIntersection(Elt *elements, int nbElements) |
---|
| 716 | { |
---|
[1328] | 717 | int mpiSize, mpiRank; |
---|
| 718 | MPI_Comm_size(communicator, &mpiSize); |
---|
| 719 | MPI_Comm_rank(communicator, &mpiRank); |
---|
[688] | 720 | |
---|
[1328] | 721 | MPI_Barrier(communicator); |
---|
[688] | 722 | |
---|
[1328] | 723 | vector<Node> *routingList = new vector<Node>[mpiSize]; |
---|
[688] | 724 | |
---|
[1328] | 725 | vector<Node> routeNodes; routeNodes.reserve(nbElements); |
---|
| 726 | for (int j = 0; j < nbElements; j++) |
---|
| 727 | { |
---|
| 728 | elements[j].id.ind = j; |
---|
| 729 | elements[j].id.rank = mpiRank; |
---|
| 730 | routeNodes.push_back(Node(elements[j].x, cptRadius(elements[j]), &elements[j])); |
---|
| 731 | } |
---|
[688] | 732 | |
---|
[1328] | 733 | vector<vector<int> > routes(routeNodes.size()); |
---|
| 734 | sstree.routeIntersections(routes, routeNodes); |
---|
| 735 | for (int i = 0; i < routes.size(); ++i) |
---|
| 736 | for (int k = 0; k < routes[i].size(); ++k) |
---|
| 737 | routingList[routes[i][k]].push_back(routeNodes[i]); |
---|
[688] | 738 | |
---|
[1328] | 739 | if (verbose >= 2) |
---|
| 740 | { |
---|
| 741 | cout << " --> rank " << mpiRank << " nbElements " << nbElements << " : "; |
---|
| 742 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 743 | cout << routingList[rank].size() << " "; |
---|
| 744 | cout << endl; |
---|
| 745 | } |
---|
| 746 | MPI_Barrier(communicator); |
---|
[688] | 747 | |
---|
[1328] | 748 | int *nbSendNode = new int[mpiSize]; |
---|
| 749 | int *nbRecvNode = new int[mpiSize]; |
---|
| 750 | int *sentMessageSize = new int[mpiSize]; |
---|
| 751 | int *recvMessageSize = new int[mpiSize]; |
---|
[688] | 752 | |
---|
[1328] | 753 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 754 | { |
---|
| 755 | nbSendNode[rank] = routingList[rank].size(); |
---|
| 756 | sentMessageSize[rank] = 0; |
---|
| 757 | for (size_t j = 0; j < routingList[rank].size(); j++) |
---|
| 758 | { |
---|
| 759 | Elt *elt = (Elt *) (routingList[rank][j].data); |
---|
| 760 | sentMessageSize[rank] += packedPolygonSize(*elt); |
---|
| 761 | } |
---|
| 762 | } |
---|
[688] | 763 | |
---|
[1328] | 764 | MPI_Alltoall(nbSendNode, 1, MPI_INT, nbRecvNode, 1, MPI_INT, communicator); |
---|
| 765 | MPI_Alltoall(sentMessageSize, 1, MPI_INT, recvMessageSize, 1, MPI_INT, communicator); |
---|
[688] | 766 | |
---|
[1328] | 767 | int total = 0; |
---|
[688] | 768 | |
---|
[1328] | 769 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 770 | { |
---|
| 771 | total = total + nbRecvNode[rank]; |
---|
| 772 | } |
---|
[688] | 773 | |
---|
[1328] | 774 | if (verbose >= 2) cout << "---> rank " << mpiRank << " : compute intersection : total received nodes " << total << endl; |
---|
| 775 | char **sendBuffer = new char*[mpiSize]; |
---|
| 776 | char **recvBuffer = new char*[mpiSize]; |
---|
| 777 | int *pos = new int[mpiSize]; |
---|
[688] | 778 | |
---|
[1328] | 779 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 780 | { |
---|
| 781 | if (nbSendNode[rank] > 0) sendBuffer[rank] = new char[sentMessageSize[rank]]; |
---|
| 782 | if (nbRecvNode[rank] > 0) recvBuffer[rank] = new char[recvMessageSize[rank]]; |
---|
| 783 | } |
---|
[688] | 784 | |
---|
[1328] | 785 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 786 | { |
---|
| 787 | pos[rank] = 0; |
---|
| 788 | for (size_t j = 0; j < routingList[rank].size(); j++) |
---|
| 789 | { |
---|
| 790 | Elt* elt = (Elt *) (routingList[rank][j].data); |
---|
| 791 | packPolygon(*elt, sendBuffer[rank], pos[rank]); |
---|
| 792 | } |
---|
| 793 | } |
---|
| 794 | delete [] routingList; |
---|
[688] | 795 | |
---|
[1328] | 796 | int nbSendRequest = 0; |
---|
| 797 | int nbRecvRequest = 0; |
---|
| 798 | MPI_Request *sendRequest = new MPI_Request[mpiSize]; |
---|
| 799 | MPI_Request *recvRequest = new MPI_Request[mpiSize]; |
---|
| 800 | MPI_Status *status = new MPI_Status[mpiSize]; |
---|
[688] | 801 | |
---|
[1328] | 802 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 803 | { |
---|
| 804 | if (nbSendNode[rank] > 0) |
---|
| 805 | { |
---|
| 806 | MPI_Issend(sendBuffer[rank], sentMessageSize[rank], MPI_CHAR, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 807 | nbSendRequest++; |
---|
| 808 | } |
---|
| 809 | if (nbRecvNode[rank] > 0) |
---|
| 810 | { |
---|
| 811 | MPI_Irecv(recvBuffer[rank], recvMessageSize[rank], MPI_CHAR, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 812 | nbRecvRequest++; |
---|
| 813 | } |
---|
| 814 | } |
---|
[1205] | 815 | |
---|
[1328] | 816 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
| 817 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
| 818 | char **sendBuffer2 = new char*[mpiSize]; |
---|
| 819 | char **recvBuffer2 = new char*[mpiSize]; |
---|
[688] | 820 | |
---|
[1328] | 821 | double tic = cputime(); |
---|
| 822 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 823 | { |
---|
| 824 | sentMessageSize[rank] = 0; |
---|
[688] | 825 | |
---|
[1328] | 826 | if (nbRecvNode[rank] > 0) |
---|
| 827 | { |
---|
| 828 | Elt *recvElt = new Elt[nbRecvNode[rank]]; |
---|
| 829 | pos[rank] = 0; |
---|
| 830 | for (int j = 0; j < nbRecvNode[rank]; j++) |
---|
| 831 | { |
---|
| 832 | unpackPolygon(recvElt[j], recvBuffer[rank], pos[rank]); |
---|
| 833 | cptEltGeom(recvElt[j], tgtGrid.pole); |
---|
| 834 | Node recvNode(recvElt[j].x, cptRadius(recvElt[j]), &recvElt[j]); |
---|
| 835 | recvNode.search(sstree.localTree.root); |
---|
| 836 | /* for a node holding an element of the target, loop throught candidates for intersecting source */ |
---|
| 837 | for (list<NodePtr>::iterator it = (recvNode.intersectors).begin(); it != (recvNode.intersectors).end(); ++it) |
---|
| 838 | { |
---|
| 839 | Elt *elt2 = (Elt *) ((*it)->data); |
---|
| 840 | /* recvElt is target, elt2 is source */ |
---|
| 841 | // intersect(&recvElt[j], elt2); |
---|
| 842 | intersect_ym(&recvElt[j], elt2); |
---|
| 843 | } |
---|
[1460] | 844 | |
---|
[1328] | 845 | if (recvElt[j].is.size() > 0) sentMessageSize[rank] += packIntersectionSize(recvElt[j]); |
---|
[688] | 846 | |
---|
[1328] | 847 | // here recvNode goes out of scope |
---|
| 848 | } |
---|
[688] | 849 | |
---|
[1328] | 850 | if (sentMessageSize[rank] > 0) |
---|
| 851 | { |
---|
| 852 | sentMessageSize[rank] += sizeof(int); |
---|
| 853 | sendBuffer2[rank] = new char[sentMessageSize[rank]]; |
---|
| 854 | *((int *) sendBuffer2[rank]) = 0; |
---|
| 855 | pos[rank] = sizeof(int); |
---|
| 856 | for (int j = 0; j < nbRecvNode[rank]; j++) |
---|
| 857 | { |
---|
| 858 | packIntersection(recvElt[j], sendBuffer2[rank], pos[rank]); |
---|
| 859 | //FIXME should be deleted: recvElt[j].delete_intersections(); // intersection areas have been packed to buffer and won't be used any more |
---|
| 860 | } |
---|
| 861 | } |
---|
| 862 | delete [] recvElt; |
---|
[1460] | 863 | |
---|
[1328] | 864 | } |
---|
| 865 | } |
---|
| 866 | delete [] pos; |
---|
[688] | 867 | |
---|
[1328] | 868 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 869 | { |
---|
| 870 | if (nbSendNode[rank] > 0) delete [] sendBuffer[rank]; |
---|
| 871 | if (nbRecvNode[rank] > 0) delete [] recvBuffer[rank]; |
---|
| 872 | nbSendNode[rank] = 0; |
---|
| 873 | } |
---|
[688] | 874 | |
---|
[1328] | 875 | if (verbose >= 2) cout << "Rank " << mpiRank << " Compute (internal) intersection " << cputime() - tic << " s" << endl; |
---|
| 876 | MPI_Alltoall(sentMessageSize, 1, MPI_INT, recvMessageSize, 1, MPI_INT, communicator); |
---|
[688] | 877 | |
---|
[1328] | 878 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 879 | if (recvMessageSize[rank] > 0) |
---|
| 880 | recvBuffer2[rank] = new char[recvMessageSize[rank]]; |
---|
[688] | 881 | |
---|
[1328] | 882 | nbSendRequest = 0; |
---|
| 883 | nbRecvRequest = 0; |
---|
[688] | 884 | |
---|
[1328] | 885 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 886 | { |
---|
| 887 | if (sentMessageSize[rank] > 0) |
---|
| 888 | { |
---|
| 889 | MPI_Issend(sendBuffer2[rank], sentMessageSize[rank], MPI_CHAR, rank, 0, communicator, &sendRequest[nbSendRequest]); |
---|
| 890 | nbSendRequest++; |
---|
| 891 | } |
---|
| 892 | if (recvMessageSize[rank] > 0) |
---|
| 893 | { |
---|
| 894 | MPI_Irecv(recvBuffer2[rank], recvMessageSize[rank], MPI_CHAR, rank, 0, communicator, &recvRequest[nbRecvRequest]); |
---|
| 895 | nbRecvRequest++; |
---|
| 896 | } |
---|
| 897 | } |
---|
[688] | 898 | |
---|
[1460] | 899 | MPI_Waitall(nbRecvRequest, recvRequest, status); |
---|
| 900 | MPI_Waitall(nbSendRequest, sendRequest, status); |
---|
[688] | 901 | |
---|
[1328] | 902 | delete [] sendRequest; |
---|
| 903 | delete [] recvRequest; |
---|
| 904 | delete [] status; |
---|
| 905 | for (int rank = 0; rank < mpiSize; rank++) |
---|
| 906 | { |
---|
| 907 | if (nbRecvNode[rank] > 0) |
---|
| 908 | { |
---|
| 909 | if (sentMessageSize[rank] > 0) |
---|
| 910 | delete [] sendBuffer2[rank]; |
---|
| 911 | } |
---|
[688] | 912 | |
---|
[1328] | 913 | if (recvMessageSize[rank] > 0) |
---|
| 914 | { |
---|
| 915 | unpackIntersection(elements, recvBuffer2[rank]); |
---|
| 916 | delete [] recvBuffer2[rank]; |
---|
| 917 | } |
---|
| 918 | } |
---|
| 919 | delete [] sendBuffer2; |
---|
| 920 | delete [] recvBuffer2; |
---|
| 921 | delete [] sendBuffer; |
---|
| 922 | delete [] recvBuffer; |
---|
| 923 | |
---|
| 924 | delete [] nbSendNode; |
---|
| 925 | delete [] nbRecvNode; |
---|
| 926 | delete [] sentMessageSize; |
---|
| 927 | delete [] recvMessageSize; |
---|
[688] | 928 | } |
---|
| 929 | |
---|
| 930 | Mapper::~Mapper() |
---|
| 931 | { |
---|
[1328] | 932 | delete [] remapMatrix; |
---|
| 933 | delete [] srcAddress; |
---|
| 934 | delete [] srcRank; |
---|
| 935 | delete [] dstAddress; |
---|
| 936 | if (neighbourElements) delete [] neighbourElements; |
---|
[688] | 937 | } |
---|
| 938 | |
---|
[1328] | 939 | } |
---|