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