1 #include <petsc/private/petscimpl.h> 2 #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3 #include <../src/ksp/pc/impls/bddc/bddcstructs.h> 4 5 #undef __FUNCT__ 6 #define __FUNCT__ "PCBDDCGraphGetDirichletDofsB" 7 PetscErrorCode PCBDDCGraphGetDirichletDofsB(PCBDDCGraph graph, IS* dirdofs) 8 { 9 PetscErrorCode ierr; 10 11 PetscFunctionBegin; 12 if (graph->dirdofsB) { 13 ierr = PetscObjectReference((PetscObject)graph->dirdofsB);CHKERRQ(ierr); 14 } else if (graph->has_dirichlet) { 15 PetscInt i,size; 16 PetscInt *dirdofs_idxs; 17 18 size = 0; 19 for (i=0;i<graph->nvtxs;i++) { 20 if (graph->count[i] && graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 21 } 22 23 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 24 size = 0; 25 for (i=0;i<graph->nvtxs;i++) { 26 if (graph->count[i] && graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 27 } 28 ierr = ISCreateGeneral(PETSC_COMM_SELF,size,dirdofs_idxs,PETSC_OWN_POINTER,&graph->dirdofsB);CHKERRQ(ierr); 29 ierr = PetscObjectReference((PetscObject)graph->dirdofsB);CHKERRQ(ierr); 30 } 31 *dirdofs = graph->dirdofsB; 32 PetscFunctionReturn(0); 33 } 34 35 #undef __FUNCT__ 36 #define __FUNCT__ "PCBDDCGraphGetDirichletDofs" 37 PetscErrorCode PCBDDCGraphGetDirichletDofs(PCBDDCGraph graph, IS* dirdofs) 38 { 39 PetscErrorCode ierr; 40 41 PetscFunctionBegin; 42 if (graph->dirdofs) { 43 ierr = PetscObjectReference((PetscObject)graph->dirdofs);CHKERRQ(ierr); 44 } else if (graph->has_dirichlet) { 45 PetscInt i,size; 46 PetscInt *dirdofs_idxs; 47 48 size = 0; 49 for (i=0;i<graph->nvtxs;i++) { 50 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 51 } 52 53 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 54 size = 0; 55 for (i=0;i<graph->nvtxs;i++) { 56 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 57 } 58 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)graph->l2gmap),size,dirdofs_idxs,PETSC_OWN_POINTER,&graph->dirdofs);CHKERRQ(ierr); 59 ierr = PetscObjectReference((PetscObject)graph->dirdofs);CHKERRQ(ierr); 60 } 61 *dirdofs = graph->dirdofs; 62 PetscFunctionReturn(0); 63 } 64 65 #undef __FUNCT__ 66 #define __FUNCT__ "PCBDDCGraphASCIIView" 67 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 68 { 69 PetscInt i,j,tabs; 70 PetscInt* queue_in_global_numbering; 71 PetscErrorCode ierr; 72 73 PetscFunctionBegin; 74 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 75 ierr = PetscViewerASCIIGetTab(viewer,&tabs);CHKERRQ(ierr); 76 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 77 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 78 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 79 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 80 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 81 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Max count %d\n",graph->maxcount);CHKERRQ(ierr); 82 if (verbosity_level > 2) { 83 for (i=0;i<graph->nvtxs;i++) { 84 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 85 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 86 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 87 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 88 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 89 if (graph->count[i]) { 90 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 91 for (j=0;j<graph->count[i];j++) { 92 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 93 } 94 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 95 } 96 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 97 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 98 if (graph->mirrors) { 99 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 100 if (graph->mirrors[i]) { 101 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 102 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 103 for (j=0;j<graph->mirrors[i];j++) { 104 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 105 } 106 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 107 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 108 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 109 } 110 } 111 if (verbosity_level > 3) { 112 if (graph->xadj && graph->adjncy) { 113 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 114 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 115 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 116 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 117 } 118 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 119 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 120 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 121 } else { 122 ierr = PetscViewerASCIISynchronizedPrintf(viewer," no adj info\n");CHKERRQ(ierr); 123 } 124 } 125 if (graph->n_local_subs) { 126 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local sub id: %d\n",graph->local_subs[i]);CHKERRQ(ierr); 127 } 128 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 129 if (graph->subset[i] && graph->subset_ncc) { 130 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 131 } 132 } 133 } 134 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 135 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 136 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 137 for (i=0;i<graph->ncc;i++) { 138 PetscInt node_num=graph->queue[graph->cptr[i]]; 139 PetscBool printcc = PETSC_FALSE; 140 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 141 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 142 for (j=0;j<graph->count[node_num];j++) { 143 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 144 } 145 if (verbosity_level > 1) { 146 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 147 if (graph->twodim || graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 148 printcc = PETSC_TRUE; 149 } 150 if (printcc) { 151 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 152 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 153 } 154 } 155 } else { 156 ierr = PetscViewerASCIISynchronizedPrintf(viewer,")");CHKERRQ(ierr); 157 } 158 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 159 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 160 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 161 } 162 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 163 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 164 PetscFunctionReturn(0); 165 } 166 167 #undef __FUNCT__ 168 #define __FUNCT__ "PCBDDCGraphRestoreCandidatesIS" 169 PetscErrorCode PCBDDCGraphRestoreCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 170 { 171 PetscInt i; 172 PetscErrorCode ierr; 173 174 PetscFunctionBegin; 175 if (n_faces) { 176 if (FacesIS) { 177 for (i=0;i<*n_faces;i++) { 178 ierr = ISDestroy(&((*FacesIS)[i]));CHKERRQ(ierr); 179 } 180 ierr = PetscFree(*FacesIS);CHKERRQ(ierr); 181 } 182 *n_faces = 0; 183 } 184 if (n_edges) { 185 if (EdgesIS) { 186 for (i=0;i<*n_edges;i++) { 187 ierr = ISDestroy(&((*EdgesIS)[i]));CHKERRQ(ierr); 188 } 189 ierr = PetscFree(*EdgesIS);CHKERRQ(ierr); 190 } 191 *n_edges = 0; 192 } 193 if (VerticesIS) { 194 ierr = ISDestroy(VerticesIS);CHKERRQ(ierr); 195 } 196 PetscFunctionReturn(0); 197 } 198 199 #undef __FUNCT__ 200 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 201 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 202 { 203 IS *ISForFaces,*ISForEdges,ISForVertices; 204 PetscInt i,nfc,nec,nvc,*idx,*mark; 205 PetscErrorCode ierr; 206 207 PetscFunctionBegin; 208 ierr = PetscCalloc1(graph->ncc,&mark);CHKERRQ(ierr); 209 /* loop on ccs to evalute number of faces, edges and vertices */ 210 nfc = 0; 211 nec = 0; 212 nvc = 0; 213 for (i=0;i<graph->ncc;i++) { 214 PetscInt repdof = graph->queue[graph->cptr[i]]; 215 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size && graph->count[repdof] < graph->maxcount) { 216 if (!graph->twodim && graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 217 nfc++; 218 mark[i] = 2; 219 } else { 220 nec++; 221 mark[i] = 1; 222 } 223 } else { 224 nvc += graph->cptr[i+1]-graph->cptr[i]; 225 } 226 } 227 228 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 229 if (FacesIS) { 230 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 231 } 232 if (EdgesIS) { 233 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 234 } 235 if (VerticesIS) { 236 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 237 } 238 239 /* loop on ccs to compute index sets for faces and edges */ 240 if (!graph->queue_sorted) { 241 PetscInt *queue_global; 242 243 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 244 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 245 for (i=0;i<graph->ncc;i++) { 246 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 247 } 248 ierr = PetscFree(queue_global);CHKERRQ(ierr); 249 graph->queue_sorted = PETSC_TRUE; 250 } 251 nfc = 0; 252 nec = 0; 253 for (i=0;i<graph->ncc;i++) { 254 PetscInt repdof = graph->queue[graph->cptr[i]]; 255 if (mark[i] == 2) { 256 if (FacesIS) { 257 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_USE_POINTER,&ISForFaces[nfc]);CHKERRQ(ierr); 258 } 259 nfc++; 260 } else if (mark[i] == 1) { 261 if (EdgesIS) { 262 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_USE_POINTER,&ISForEdges[nec]);CHKERRQ(ierr); 263 } 264 nec++; 265 } 266 } 267 268 /* index set for vertices */ 269 if (VerticesIS) { 270 nvc = 0; 271 for (i=0;i<graph->ncc;i++) { 272 if (!mark[i]) { 273 PetscInt j; 274 275 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 276 idx[nvc]=graph->queue[j]; 277 nvc++; 278 } 279 } 280 } 281 /* sort vertex set (by local ordering) */ 282 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 283 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 284 } 285 ierr = PetscFree(mark);CHKERRQ(ierr); 286 287 /* get back info */ 288 if (n_faces) *n_faces = nfc; 289 if (FacesIS) *FacesIS = ISForFaces; 290 if (n_edges) *n_edges = nec; 291 if (EdgesIS) *EdgesIS = ISForEdges; 292 if (VerticesIS) *VerticesIS = ISForVertices; 293 PetscFunctionReturn(0); 294 } 295 296 #undef __FUNCT__ 297 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 298 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 299 { 300 PetscBool adapt_interface_reduced; 301 MPI_Comm interface_comm; 302 PetscMPIInt size; 303 PetscInt i; 304 PetscBool twodim; 305 PetscErrorCode ierr; 306 307 PetscFunctionBegin; 308 /* compute connected components locally */ 309 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 310 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 311 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 312 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 313 adapt_interface_reduced = PETSC_FALSE; 314 if (size > 1) { 315 PetscInt i; 316 PetscBool adapt_interface = PETSC_FALSE; 317 for (i=0;i<graph->n_subsets;i++) { 318 /* We are not sure that on a given subset of the local interface, 319 with two connected components, the latters be the same among sharing subdomains */ 320 if (graph->subset_ncc[i] > 1) { 321 adapt_interface = PETSC_TRUE; 322 break; 323 } 324 } 325 ierr = MPIU_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 326 } 327 328 if (graph->n_subsets && adapt_interface_reduced) { 329 PetscBT subset_cc_adapt; 330 MPI_Request *send_requests,*recv_requests; 331 PetscInt *send_buffer,*recv_buffer; 332 PetscInt sum_requests,start_of_recv,start_of_send; 333 PetscInt *cum_recv_counts; 334 PetscInt *labels; 335 PetscInt ncc,cum_queue,mss,mns,j,k,s; 336 PetscInt **refine_buffer=NULL,*private_labels = NULL; 337 338 ierr = PetscMalloc1(graph->nvtxs,&labels);CHKERRQ(ierr); 339 ierr = PetscMemzero(labels,graph->nvtxs*sizeof(*labels));CHKERRQ(ierr); 340 for (i=0;i<graph->ncc;i++) 341 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) 342 labels[graph->queue[j]] = i; 343 344 /* allocate some space */ 345 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 346 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 347 348 /* first count how many neighbours per connected component I will receive from */ 349 cum_recv_counts[0] = 0; 350 for (i=0;i<graph->n_subsets;i++) cum_recv_counts[i+1] = cum_recv_counts[i]+graph->count[graph->subset_idxs[i][0]]; 351 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_buffer);CHKERRQ(ierr); 352 ierr = PetscMalloc2(cum_recv_counts[graph->n_subsets],&send_requests,cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 353 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 354 send_requests[i] = MPI_REQUEST_NULL; 355 recv_requests[i] = MPI_REQUEST_NULL; 356 } 357 358 /* exchange with my neighbours the number of my connected components on the subset of interface */ 359 sum_requests = 0; 360 for (i=0;i<graph->n_subsets;i++) { 361 PetscMPIInt neigh,tag; 362 PetscInt count,*neighs; 363 364 count = graph->count[graph->subset_idxs[i][0]]; 365 neighs = graph->neighbours_set[graph->subset_idxs[i][0]]; 366 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 367 for (k=0;k<count;k++) { 368 ierr = PetscMPIIntCast(neighs[k],&neigh);CHKERRQ(ierr); 369 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 370 ierr = MPI_Irecv(&recv_buffer[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 371 sum_requests++; 372 } 373 } 374 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 375 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 376 377 /* determine the subsets I have to adapt (those having more than 1 cc) */ 378 ierr = PetscBTCreate(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 379 ierr = PetscBTMemzero(graph->n_subsets,subset_cc_adapt);CHKERRQ(ierr); 380 for (i=0;i<graph->n_subsets;i++) { 381 if (graph->subset_ncc[i] > 1) { 382 ierr = PetscBTSet(subset_cc_adapt,i);CHKERRQ(ierr); 383 continue; 384 } 385 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 386 if (recv_buffer[j] > 1) { 387 ierr = PetscBTSet(subset_cc_adapt,i); 388 break; 389 } 390 } 391 } 392 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 393 394 /* determine send/recv buffers sizes */ 395 j = 0; 396 mss = 0; 397 for (i=0;i<graph->n_subsets;i++) { 398 if (PetscBTLookup(subset_cc_adapt,i)) { 399 j += graph->subset_size[i]; 400 mss = PetscMax(graph->subset_size[i],mss); 401 } 402 } 403 k = 0; 404 mns = 0; 405 for (i=0;i<graph->n_subsets;i++) { 406 if (PetscBTLookup(subset_cc_adapt,i)) { 407 k += (cum_recv_counts[i+1]-cum_recv_counts[i])*graph->subset_size[i]; 408 mns = PetscMax(cum_recv_counts[i+1]-cum_recv_counts[i],mns); 409 } 410 } 411 ierr = PetscMalloc2(j,&send_buffer,k,&recv_buffer);CHKERRQ(ierr); 412 413 /* fill send buffer (order matters: subset_idxs ordered by global ordering) */ 414 j = 0; 415 for (i=0;i<graph->n_subsets;i++) 416 if (PetscBTLookup(subset_cc_adapt,i)) 417 for (k=0;k<graph->subset_size[i];k++) 418 send_buffer[j++] = labels[graph->subset_idxs[i][k]]; 419 420 /* now exchange the data */ 421 start_of_recv = 0; 422 start_of_send = 0; 423 sum_requests = 0; 424 for (i=0;i<graph->n_subsets;i++) { 425 if (PetscBTLookup(subset_cc_adapt,i)) { 426 PetscMPIInt neigh,tag; 427 PetscInt size_of_send = graph->subset_size[i]; 428 429 j = graph->subset_idxs[i][0]; 430 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 431 for (k=0;k<graph->count[j];k++) { 432 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 433 ierr = MPI_Isend(&send_buffer[start_of_send],size_of_send,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 434 ierr = MPI_Irecv(&recv_buffer[start_of_recv],size_of_send,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 435 start_of_recv += size_of_send; 436 sum_requests++; 437 } 438 start_of_send += size_of_send; 439 } 440 } 441 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 442 443 /* refine connected components */ 444 start_of_recv = 0; 445 /* allocate some temporary space */ 446 if (mss) { 447 ierr = PetscMalloc1(mss,&refine_buffer);CHKERRQ(ierr); 448 ierr = PetscMalloc2(mss*(mns+1),&refine_buffer[0],mss,&private_labels);CHKERRQ(ierr); 449 } 450 ncc = 0; 451 cum_queue = 0; 452 graph->cptr[0] = 0; 453 for (i=0;i<graph->n_subsets;i++) { 454 if (PetscBTLookup(subset_cc_adapt,i)) { 455 PetscInt subset_counter = 0; 456 PetscInt sharingprocs = cum_recv_counts[i+1]-cum_recv_counts[i]+1; /* count myself */ 457 PetscInt buffer_size = graph->subset_size[i]; 458 459 /* compute pointers */ 460 for (j=1;j<buffer_size;j++) refine_buffer[j] = refine_buffer[j-1] + sharingprocs; 461 /* analyze contributions from subdomains that share the i-th subset 462 The stricture of refine_buffer is suitable to find intersections of ccs among sharingprocs. 463 supposing the current subset is shared by 3 processes and has dimension 5 with global dofs 0,1,2,3,4 (local 0,4,3,1,2) 464 sharing procs connected components: 465 neigh 0: [0 1 4], [2 3], labels [4,7] (2 connected components) 466 neigh 1: [0 1], [2 3 4], labels [3 2] (2 connected components) 467 neigh 2: [0 4], [1], [2 3], labels [1 5 6] (3 connected components) 468 refine_buffer will be filled as: 469 [ 4, 3, 1; 470 4, 2, 1; 471 7, 2, 6; 472 4, 3, 5; 473 7, 2, 6; ]; 474 The connected components in local ordering are [0], [1], [2 3], [4] */ 475 /* fill temp_buffer */ 476 for (k=0;k<buffer_size;k++) refine_buffer[k][0] = labels[graph->subset_idxs[i][k]]; 477 for (j=0;j<sharingprocs-1;j++) { 478 for (k=0;k<buffer_size;k++) refine_buffer[k][j+1] = recv_buffer[start_of_recv+k]; 479 start_of_recv += buffer_size; 480 } 481 ierr = PetscMemzero(private_labels,buffer_size*sizeof(PetscInt));CHKERRQ(ierr); 482 for (j=0;j<buffer_size;j++) { 483 if (!private_labels[j]) { /* found a new cc */ 484 PetscBool same_set; 485 486 graph->cptr[ncc] = cum_queue; 487 ncc++; 488 subset_counter++; 489 private_labels[j] = subset_counter; 490 graph->queue[cum_queue++] = graph->subset_idxs[i][j]; 491 for (k=j+1;k<buffer_size;k++) { /* check for other nodes in new cc */ 492 same_set = PETSC_TRUE; 493 for (s=0;s<sharingprocs;s++) { 494 if (refine_buffer[j][s] != refine_buffer[k][s]) { 495 same_set = PETSC_FALSE; 496 break; 497 } 498 } 499 if (same_set) { 500 private_labels[k] = subset_counter; 501 graph->queue[cum_queue++] = graph->subset_idxs[i][k]; 502 } 503 } 504 } 505 } 506 graph->cptr[ncc] = cum_queue; 507 graph->subset_ncc[i] = subset_counter; 508 graph->queue_sorted = PETSC_FALSE; 509 } else { /* this subset does not need to be adapted */ 510 ierr = PetscMemcpy(graph->queue+cum_queue,graph->subset_idxs[i],graph->subset_size[i]*sizeof(PetscInt));CHKERRQ(ierr); 511 ncc++; 512 cum_queue += graph->subset_size[i]; 513 graph->cptr[ncc] = cum_queue; 514 } 515 } 516 graph->cptr[ncc] = cum_queue; 517 graph->ncc = ncc; 518 if (mss) { 519 ierr = PetscFree2(refine_buffer[0],private_labels);CHKERRQ(ierr); 520 ierr = PetscFree(refine_buffer);CHKERRQ(ierr); 521 } 522 ierr = PetscFree(labels);CHKERRQ(ierr); 523 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 524 ierr = PetscFree2(send_requests,recv_requests);CHKERRQ(ierr); 525 ierr = PetscFree2(send_buffer,recv_buffer);CHKERRQ(ierr); 526 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 527 ierr = PetscBTDestroy(&subset_cc_adapt);CHKERRQ(ierr); 528 } 529 530 /* Determine if we are in 2D or 3D */ 531 twodim = PETSC_TRUE; 532 for (i=0;i<graph->ncc;i++) { 533 PetscInt repdof = graph->queue[graph->cptr[i]]; 534 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 535 if (graph->count[repdof] > 1 || graph->special_dof[repdof] == PCBDDCGRAPH_NEUMANN_MARK) { 536 twodim = PETSC_FALSE; 537 break; 538 } 539 } 540 } 541 ierr = MPIU_Allreduce(&twodim,&graph->twodim,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 542 PetscFunctionReturn(0); 543 } 544 545 546 #undef __FUNCT__ 547 #define __FUNCT__ "PCBDDCGraphComputeCC_Private" 548 PETSC_STATIC_INLINE PetscErrorCode PCBDDCGraphComputeCC_Private(PCBDDCGraph graph,PetscInt pid,PetscInt* queue_tip,PetscInt n_prev,PetscInt* n_added) 549 { 550 PetscInt i,j,n; 551 PetscInt *xadj = graph->xadj,*adjncy = graph->adjncy; 552 PetscBT touched = graph->touched; 553 PetscBool havecsr = (PetscBool)(xadj && adjncy); 554 PetscBool havesubs = (PetscBool)(!!graph->n_local_subs); 555 PetscErrorCode ierr; 556 557 PetscFunctionBegin; 558 n = 0; 559 if (havecsr && !havesubs) { 560 for (i=-n_prev;i<0;i++) { 561 PetscInt start_dof = queue_tip[i]; 562 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 563 PetscInt dof = adjncy[j]; 564 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 565 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 566 queue_tip[n] = dof; 567 n++; 568 } 569 } 570 } 571 } else if (havecsr && havesubs) { 572 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 573 for (i=-n_prev;i<0;i++) { 574 PetscInt start_dof = queue_tip[i]; 575 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 576 PetscInt dof = adjncy[j]; 577 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 578 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 579 queue_tip[n] = dof; 580 n++; 581 } 582 } 583 } 584 } else { /* sub info only */ 585 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 586 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 587 PetscInt dof = graph->subset_idxs[pid-1][j]; 588 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 589 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 590 queue_tip[n] = dof; 591 n++; 592 } 593 } 594 } 595 *n_added = n; 596 PetscFunctionReturn(0); 597 } 598 599 #undef __FUNCT__ 600 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 601 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 602 { 603 PetscInt ncc,cum_queue,n; 604 PetscMPIInt commsize; 605 PetscErrorCode ierr; 606 607 PetscFunctionBegin; 608 if (!graph->setupcalled) SETERRQ(PetscObjectComm((PetscObject)graph->l2gmap),PETSC_ERR_ORDER,"PCBDDCGraphSetUp should be called first"); 609 /* quiet return if there isn't any local info */ 610 if ((!graph->xadj || !graph->adjncy) && !graph->n_local_subs) { 611 PetscFunctionReturn(0); 612 } 613 614 /* reset any previous search of connected components */ 615 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 616 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&commsize);CHKERRQ(ierr); 617 if (commsize > graph->commsizelimit) { 618 PetscInt i; 619 for (i=0;i<graph->nvtxs;i++) { 620 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 621 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 622 } 623 } 624 } 625 626 /* begin search for connected components */ 627 cum_queue = 0; 628 ncc = 0; 629 for (n=0;n<graph->n_subsets;n++) { 630 PetscInt pid = n+1; /* partition labeled by 0 is discarded */ 631 PetscInt found = 0,prev = 0,first = 0,ncc_pid = 0; 632 while (found != graph->subset_size[n]) { 633 PetscInt added = 0; 634 if (!prev) { /* search for new starting dof */ 635 while (PetscBTLookup(graph->touched,graph->subset_idxs[n][first])) first++; 636 ierr = PetscBTSet(graph->touched,graph->subset_idxs[n][first]);CHKERRQ(ierr); 637 graph->queue[cum_queue] = graph->subset_idxs[n][first]; 638 graph->cptr[ncc] = cum_queue; 639 prev = 1; 640 cum_queue++; 641 found++; 642 ncc_pid++; 643 ncc++; 644 } 645 ierr = PCBDDCGraphComputeCC_Private(graph,pid,graph->queue + cum_queue,prev,&added);CHKERRQ(ierr); 646 if (!added) { 647 graph->subset_ncc[n] = ncc_pid; 648 graph->cptr[ncc] = cum_queue; 649 } 650 prev = added; 651 found += added; 652 cum_queue += added; 653 if (added && found == graph->subset_size[n]) { 654 graph->subset_ncc[n] = ncc_pid; 655 graph->cptr[ncc] = cum_queue; 656 } 657 } 658 } 659 graph->ncc = ncc; 660 graph->queue_sorted = PETSC_FALSE; 661 PetscFunctionReturn(0); 662 } 663 664 #undef __FUNCT__ 665 #define __FUNCT__ "PCBDDCGraphSetUp" 666 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 667 { 668 VecScatter scatter_ctx; 669 Vec local_vec,local_vec2,global_vec; 670 IS to,from,subset,subset_n; 671 MPI_Comm comm; 672 PetscScalar *array,*array2; 673 const PetscInt *is_indices; 674 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global; 675 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 676 PetscMPIInt commsize; 677 PetscBool same_set,mirrors_found; 678 PetscErrorCode ierr; 679 680 PetscFunctionBegin; 681 graph->has_dirichlet = PETSC_FALSE; 682 if (dirichlet_is) { 683 PetscCheckSameComm(graph->l2gmap,1,dirichlet_is,4); 684 graph->has_dirichlet = PETSC_TRUE; 685 } 686 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 687 ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 688 689 /* custom_minimal_size */ 690 graph->custom_minimal_size = custom_minimal_size; 691 /* get info l2gmap and allocate work vectors */ 692 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 693 /* check if we have any local periodic nodes (periodic BCs) */ 694 mirrors_found = PETSC_FALSE; 695 if (graph->nvtxs && n_neigh) { 696 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 697 for (i=0; i<n_shared[0]; i++) { 698 if (graph->count[shared[0][i]] > 1) { 699 mirrors_found = PETSC_TRUE; 700 break; 701 } 702 } 703 } 704 /* create some workspace objects */ 705 local_vec = NULL; 706 local_vec2 = NULL; 707 global_vec = NULL; 708 to = NULL; 709 from = NULL; 710 scatter_ctx = NULL; 711 if (n_ISForDofs || dirichlet_is || neumann_is || custom_primal_vertices) { 712 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 713 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 714 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 715 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 716 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 717 ierr = VecSetSizes(global_vec,PETSC_DECIDE,graph->nvtxs_global);CHKERRQ(ierr); 718 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 719 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 720 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 721 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 722 } else if (mirrors_found) { 723 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 724 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 725 } 726 /* compute local mirrors (if any) */ 727 if (mirrors_found) { 728 PetscInt *local_indices,*global_indices; 729 /* get arrays of local and global indices */ 730 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 731 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 732 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 733 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 734 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 735 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 736 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 737 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 738 /* allocate space for mirrors */ 739 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 740 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 741 graph->mirrors_set[0] = 0; 742 743 k=0; 744 for (i=0;i<n_shared[0];i++) { 745 j=shared[0][i]; 746 if (graph->count[j] > 1) { 747 graph->mirrors[j]++; 748 k++; 749 } 750 } 751 /* allocate space for set of mirrors */ 752 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 753 for (i=1;i<graph->nvtxs;i++) 754 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 755 756 /* fill arrays */ 757 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 758 for (j=0;j<n_shared[0];j++) { 759 i=shared[0][j]; 760 if (graph->count[i] > 1) 761 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 762 } 763 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 764 for (i=0;i<graph->nvtxs;i++) { 765 if (graph->mirrors[i] > 0) { 766 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 767 j = global_indices[k]; 768 while ( k > 0 && global_indices[k-1] == j) k--; 769 for (j=0;j<graph->mirrors[i];j++) { 770 graph->mirrors_set[i][j]=local_indices[k+j]; 771 } 772 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 773 } 774 } 775 ierr = PetscFree(local_indices);CHKERRQ(ierr); 776 ierr = PetscFree(global_indices);CHKERRQ(ierr); 777 } 778 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 779 ierr = ISDestroy(&to);CHKERRQ(ierr); 780 ierr = ISDestroy(&from);CHKERRQ(ierr); 781 782 /* Count total number of neigh per node */ 783 k = 0; 784 for (i=1;i<n_neigh;i++) { 785 k += n_shared[i]; 786 for (j=0;j<n_shared[i];j++) { 787 graph->count[shared[i][j]] += 1; 788 } 789 } 790 /* Allocate space for storing the set of neighbours for each node */ 791 if (graph->nvtxs) { 792 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 793 } 794 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 795 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 796 } 797 /* Get information for sharing subdomains */ 798 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 799 for (i=1;i<n_neigh;i++) { /* dont count myself */ 800 s = n_shared[i]; 801 for (j=0;j<s;j++) { 802 k = shared[i][j]; 803 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 804 graph->count[k] += 1; 805 } 806 } 807 /* sort set of sharing subdomains */ 808 for (i=0;i<graph->nvtxs;i++) { 809 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 810 } 811 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 812 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 813 814 /* 815 Get info for dofs splitting 816 User can specify just a subset; an additional field is considered as a complementary field 817 */ 818 for (i=0;i<graph->nvtxs;i++) graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 819 if (n_ISForDofs) { 820 ierr = VecSet(local_vec,-1.0);CHKERRQ(ierr); 821 } 822 for (i=0;i<n_ISForDofs;i++) { 823 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 824 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 825 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 826 for (j=0;j<is_size;j++) { 827 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 828 graph->which_dof[is_indices[j]] = i; 829 array[is_indices[j]] = 1.*i; 830 } 831 } 832 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 833 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 834 } 835 /* Check consistency among neighbours */ 836 if (n_ISForDofs) { 837 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 838 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 839 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec2,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 840 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec2,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 841 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 842 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 843 for (i=0;i<graph->nvtxs;i++){ 844 PetscInt field1,field2; 845 846 field1 = (PetscInt)PetscRealPart(array[i]); 847 field2 = (PetscInt)PetscRealPart(array2[i]); 848 if (field1 != field2) SETERRQ3(comm,PETSC_ERR_USER,"Local node %D have been assigned two different field ids %D and %D at the same time\n",i,field1,field2); 849 } 850 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 851 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 852 } 853 /* Take into account Neumann nodes */ 854 if (neumann_is) { 855 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 856 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 857 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 858 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 859 for (i=0;i<is_size;i++) { 860 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 861 array[is_indices[i]] = 1.0; 862 } 863 } 864 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 865 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 866 /* Neumann nodes: impose consistency among neighbours */ 867 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 868 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 869 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 870 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 871 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 872 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 873 for (i=0;i<graph->nvtxs;i++) { 874 if (PetscRealPart(array[i]) > 0.1) { 875 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 876 } 877 } 878 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 879 } 880 /* Take into account Dirichlet nodes (they overwrite any Neumann boundary node previously set) */ 881 if (dirichlet_is) { 882 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 883 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 884 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 885 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 886 for (i=0;i<is_size;i++){ 887 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 888 array[is_indices[i]] = 1.0; 889 } 890 } 891 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 892 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 893 /* Dirichlet nodes: impose consistency among neighbours */ 894 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 895 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 896 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 897 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 898 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 899 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 900 for (i=0;i<graph->nvtxs;i++) { 901 if (PetscRealPart(array[i]) > 0.1) { 902 if (commsize > graph->commsizelimit) { /* dirichlet nodes treated as internal */ 903 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 904 graph->subset[i] = 0; 905 } 906 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 907 } 908 } 909 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 910 } 911 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 912 if (graph->mirrors) { 913 for (i=0;i<graph->nvtxs;i++) 914 if (graph->mirrors[i]) 915 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 916 917 if (graph->xadj && graph->adjncy) { 918 PetscInt *new_xadj,*new_adjncy; 919 /* sort CSR graph */ 920 for (i=0;i<graph->nvtxs;i++) 921 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 922 923 /* adapt local CSR graph in case of local periodicity */ 924 k = 0; 925 for (i=0;i<graph->nvtxs;i++) 926 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 927 k += graph->mirrors[graph->adjncy[j]]; 928 929 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 930 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 931 new_xadj[0] = 0; 932 for (i=0;i<graph->nvtxs;i++) { 933 k = graph->xadj[i+1]-graph->xadj[i]; 934 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 935 new_xadj[i+1] = new_xadj[i]+k; 936 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 937 k = graph->mirrors[graph->adjncy[j]]; 938 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 939 new_xadj[i+1] += k; 940 } 941 k = new_xadj[i+1]-new_xadj[i]; 942 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 943 new_xadj[i+1] = new_xadj[i]+k; 944 } 945 /* set new CSR into graph */ 946 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 947 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 948 graph->xadj = new_xadj; 949 graph->adjncy = new_adjncy; 950 } 951 } 952 953 /* mark special nodes (if any) -> each will become a single node equivalence class */ 954 if (custom_primal_vertices) { 955 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 956 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 957 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 958 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 959 for (i=0;i<is_size;i++){ 960 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 961 array[is_indices[i]] = 1.0; 962 } 963 } 964 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 965 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 966 /* special nodes: impose consistency among neighbours */ 967 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 968 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 969 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 970 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 971 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 972 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 973 j = 0; 974 for (i=0;i<graph->nvtxs;i++) { 975 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 976 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 977 j++; 978 } 979 } 980 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 981 } 982 983 /* mark interior nodes (if commsize > graph->commsizelimit) as touched and belonging to partition number 0 */ 984 if (commsize > graph->commsizelimit) { 985 for (i=0;i<graph->nvtxs;i++) { 986 if (!graph->count[i]) { 987 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 988 graph->subset[i] = 0; 989 } 990 } 991 } 992 993 /* init graph structure and compute default subsets */ 994 nodes_touched = 0; 995 for (i=0;i<graph->nvtxs;i++) { 996 if (PetscBTLookup(graph->touched,i)) { 997 nodes_touched++; 998 } 999 } 1000 i = 0; 1001 graph->ncc = 0; 1002 total_counts = 0; 1003 1004 /* allocated space for queues */ 1005 if (commsize == graph->commsizelimit) { 1006 ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,graph->nvtxs,&graph->queue);CHKERRQ(ierr); 1007 } else { 1008 PetscInt nused = graph->nvtxs - nodes_touched; 1009 ierr = PetscMalloc2(nused+1,&graph->cptr,nused,&graph->queue);CHKERRQ(ierr); 1010 } 1011 1012 while (nodes_touched<graph->nvtxs) { 1013 /* find first untouched node in local ordering */ 1014 while (PetscBTLookup(graph->touched,i)) i++; 1015 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1016 graph->subset[i] = graph->ncc+1; 1017 graph->cptr[graph->ncc] = total_counts; 1018 graph->queue[total_counts] = i; 1019 total_counts++; 1020 nodes_touched++; 1021 /* now find all other nodes having the same set of sharing subdomains */ 1022 for (j=i+1;j<graph->nvtxs;j++) { 1023 /* check for same number of sharing subdomains, dof number and same special mark */ 1024 if (!PetscBTLookup(graph->touched,j) && graph->count[i] == graph->count[j] && graph->which_dof[i] == graph->which_dof[j] && graph->special_dof[i] == graph->special_dof[j]) { 1025 /* check for same set of sharing subdomains */ 1026 same_set = PETSC_TRUE; 1027 for (k=0;k<graph->count[j];k++){ 1028 if (graph->neighbours_set[i][k] != graph->neighbours_set[j][k]) { 1029 same_set = PETSC_FALSE; 1030 } 1031 } 1032 /* I found a friend of mine */ 1033 if (same_set) { 1034 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1035 graph->subset[j] = graph->ncc+1; 1036 nodes_touched++; 1037 graph->queue[total_counts] = j; 1038 total_counts++; 1039 } 1040 } 1041 } 1042 graph->ncc++; 1043 } 1044 /* set default number of subsets (at this point no info on csr and/or local_subs has been taken into account, so n_subsets = ncc */ 1045 graph->n_subsets = graph->ncc; 1046 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1047 for (i=0;i<graph->n_subsets;i++) { 1048 graph->subset_ncc[i] = 1; 1049 } 1050 /* final pointer */ 1051 graph->cptr[graph->ncc] = total_counts; 1052 1053 /* For consistency reasons (among neighbours), I need to sort (by global ordering) each connected component */ 1054 /* Get a reference node (min index in global ordering) for each subset for tagging messages */ 1055 ierr = PetscMalloc1(graph->ncc,&graph->subset_ref_node);CHKERRQ(ierr); 1056 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1057 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1058 for (j=0;j<graph->ncc;j++) { 1059 ierr = PetscSortIntWithArray(graph->cptr[j+1]-graph->cptr[j],&queue_global[graph->cptr[j]],&graph->queue[graph->cptr[j]]);CHKERRQ(ierr); 1060 graph->subset_ref_node[j] = graph->queue[graph->cptr[j]]; 1061 } 1062 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1063 graph->queue_sorted = PETSC_TRUE; 1064 /* save information on subsets (needed when analyzing the connected components) */ 1065 if (graph->ncc) { 1066 ierr = PetscMalloc2(graph->ncc,&graph->subset_size,graph->ncc,&graph->subset_idxs);CHKERRQ(ierr); 1067 ierr = PetscMalloc1(graph->cptr[graph->ncc],&graph->subset_idxs[0]);CHKERRQ(ierr); 1068 ierr = PetscMemzero(graph->subset_idxs[0],graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1069 for (j=1;j<graph->ncc;j++) { 1070 graph->subset_size[j-1] = graph->cptr[j] - graph->cptr[j-1]; 1071 graph->subset_idxs[j] = graph->subset_idxs[j-1] + graph->subset_size[j-1]; 1072 } 1073 graph->subset_size[graph->ncc-1] = graph->cptr[graph->ncc] - graph->cptr[graph->ncc-1]; 1074 ierr = PetscMemcpy(graph->subset_idxs[0],graph->queue,graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1075 } 1076 1077 /* renumber reference nodes */ 1078 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->ncc,graph->subset_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 1079 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,subset_n,&subset);CHKERRQ(ierr); 1080 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1081 ierr = ISRenumber(subset,NULL,NULL,&subset_n);CHKERRQ(ierr); 1082 ierr = ISDestroy(&subset);CHKERRQ(ierr); 1083 ierr = ISGetLocalSize(subset_n,&k);CHKERRQ(ierr); 1084 if (k != graph->ncc) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",k,graph->ncc); 1085 ierr = ISGetIndices(subset_n,&is_indices);CHKERRQ(ierr); 1086 ierr = PetscMemcpy(graph->subset_ref_node,is_indices,graph->ncc*sizeof(PetscInt));CHKERRQ(ierr); 1087 ierr = ISRestoreIndices(subset_n,&is_indices);CHKERRQ(ierr); 1088 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1089 1090 /* free workspace */ 1091 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1092 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1093 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1094 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1095 graph->setupcalled = PETSC_TRUE; 1096 PetscFunctionReturn(0); 1097 } 1098 1099 #undef __FUNCT__ 1100 #define __FUNCT__ "PCBDDCGraphResetCSR" 1101 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1102 { 1103 PetscErrorCode ierr; 1104 1105 PetscFunctionBegin; 1106 if (graph->freecsr) { 1107 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1108 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1109 } else { 1110 graph->xadj = NULL; 1111 graph->adjncy = NULL; 1112 } 1113 graph->freecsr = PETSC_FALSE; 1114 graph->nvtxs_csr = 0; 1115 PetscFunctionReturn(0); 1116 } 1117 1118 #undef __FUNCT__ 1119 #define __FUNCT__ "PCBDDCGraphReset" 1120 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1121 { 1122 PetscErrorCode ierr; 1123 1124 PetscFunctionBegin; 1125 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1126 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1127 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1128 if (graph->nvtxs) { 1129 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1130 } 1131 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1132 ierr = PetscFree5(graph->count, 1133 graph->neighbours_set, 1134 graph->subset, 1135 graph->which_dof, 1136 graph->special_dof);CHKERRQ(ierr); 1137 ierr = PetscFree2(graph->cptr,graph->queue);CHKERRQ(ierr); 1138 if (graph->mirrors) { 1139 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1140 } 1141 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1142 if (graph->subset_idxs) { 1143 ierr = PetscFree(graph->subset_idxs[0]);CHKERRQ(ierr); 1144 } 1145 ierr = PetscFree2(graph->subset_size,graph->subset_idxs);CHKERRQ(ierr); 1146 ierr = ISDestroy(&graph->dirdofs);CHKERRQ(ierr); 1147 ierr = ISDestroy(&graph->dirdofsB);CHKERRQ(ierr); 1148 if (graph->n_local_subs) { 1149 ierr = PetscFree(graph->local_subs);CHKERRQ(ierr); 1150 } 1151 graph->has_dirichlet = PETSC_FALSE; 1152 graph->nvtxs = 0; 1153 graph->nvtxs_global = 0; 1154 graph->n_subsets = 0; 1155 graph->custom_minimal_size = 1; 1156 graph->n_local_subs = 0; 1157 graph->maxcount = PETSC_MAX_INT; 1158 graph->setupcalled = PETSC_FALSE; 1159 PetscFunctionReturn(0); 1160 } 1161 1162 #undef __FUNCT__ 1163 #define __FUNCT__ "PCBDDCGraphInit" 1164 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap, PetscInt N, PetscInt maxcount) 1165 { 1166 PetscInt n; 1167 PetscErrorCode ierr; 1168 1169 PetscFunctionBegin; 1170 PetscValidPointer(graph,1); 1171 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1172 PetscValidLogicalCollectiveInt(l2gmap,N,3); 1173 PetscValidLogicalCollectiveInt(l2gmap,maxcount,4); 1174 /* raise an error if already allocated */ 1175 if (graph->nvtxs_global) SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1176 /* set number of vertices */ 1177 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1178 graph->l2gmap = l2gmap; 1179 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1180 graph->nvtxs = n; 1181 graph->nvtxs_global = N; 1182 /* allocate used space */ 1183 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1184 ierr = PetscMalloc5(graph->nvtxs,&graph->count, 1185 graph->nvtxs,&graph->neighbours_set, 1186 graph->nvtxs,&graph->subset, 1187 graph->nvtxs,&graph->which_dof, 1188 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1189 /* zeroes memory */ 1190 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1191 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1192 /* use -1 as a default value for which_dof array */ 1193 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1194 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1195 /* zeroes first pointer to neighbour set */ 1196 if (graph->nvtxs) { 1197 graph->neighbours_set[0] = 0; 1198 } 1199 /* zeroes workspace for values of ncc */ 1200 graph->subset_ncc = 0; 1201 graph->subset_ref_node = 0; 1202 /* default flag for csr */ 1203 graph->freecsr = PETSC_FALSE; 1204 /* maxcount for cc */ 1205 graph->maxcount = maxcount; 1206 PetscFunctionReturn(0); 1207 } 1208 1209 #undef __FUNCT__ 1210 #define __FUNCT__ "PCBDDCGraphDestroy" 1211 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1212 { 1213 PetscErrorCode ierr; 1214 1215 PetscFunctionBegin; 1216 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1217 ierr = PetscFree(*graph);CHKERRQ(ierr); 1218 PetscFunctionReturn(0); 1219 } 1220 1221 #undef __FUNCT__ 1222 #define __FUNCT__ "PCBDDCGraphCreate" 1223 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1224 { 1225 PCBDDCGraph new_graph; 1226 PetscErrorCode ierr; 1227 1228 PetscFunctionBegin; 1229 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1230 new_graph->custom_minimal_size = 1; 1231 new_graph->commsizelimit = 1; 1232 *graph = new_graph; 1233 PetscFunctionReturn(0); 1234 } 1235