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__ "PCBDDCGraphGetDirichletDofs" 7 PetscErrorCode PCBDDCGraphGetDirichletDofs(PCBDDCGraph graph,IS* dirdofs) 8 { 9 PetscInt i,size,sizer; 10 PetscErrorCode ierr; 11 12 PetscFunctionBegin; 13 size = 0; 14 for (i=0;i<graph->nvtxs;i++) { 15 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 16 } 17 ierr = MPI_Allreduce(&size,&sizer,1,MPIU_INT,MPI_LOR,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 18 if (sizer) { 19 PetscInt *dirdofs_idxs; 20 21 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 22 size = 0; 23 for (i=0;i<graph->nvtxs;i++) { 24 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 25 } 26 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)graph->l2gmap),size,dirdofs_idxs,PETSC_OWN_POINTER,dirdofs);CHKERRQ(ierr); 27 } else { 28 *dirdofs = NULL; 29 } 30 PetscFunctionReturn(0); 31 } 32 33 #undef __FUNCT__ 34 #define __FUNCT__ "PCBDDCGraphASCIIView" 35 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 36 { 37 PetscInt i,j,tabs; 38 PetscInt* queue_in_global_numbering; 39 PetscErrorCode ierr; 40 41 PetscFunctionBegin; 42 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 43 ierr = PetscViewerASCIIGetTab(viewer,&tabs);CHKERRQ(ierr); 44 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 46 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 47 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 48 if (verbosity_level > 1) { 49 for (i=0;i<graph->nvtxs;i++) { 50 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 51 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 52 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 53 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 54 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 55 if (graph->count[i]) { 56 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 57 for (j=0;j<graph->count[i];j++) { 58 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 59 } 60 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 61 } 62 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 63 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 64 if (graph->mirrors) { 65 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 66 if (graph->mirrors[i]) { 67 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 68 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 69 for (j=0;j<graph->mirrors[i];j++) { 70 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 71 } 72 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 73 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 74 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 75 } 76 } 77 if (verbosity_level > 2) { 78 if (graph->xadj && graph->adjncy) { 79 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 80 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 81 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 82 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 83 } 84 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 85 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 86 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 87 } else { 88 ierr = PetscViewerASCIISynchronizedPrintf(viewer," no adj info\n");CHKERRQ(ierr); 89 } 90 } 91 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 92 if (graph->subset[i] && graph->subset_ncc) { 93 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 94 } 95 } 96 } 97 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 98 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 99 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 100 for (i=0;i<graph->ncc;i++) { 101 PetscInt node_num=graph->queue[graph->cptr[i]]; 102 PetscBool printcc = PETSC_FALSE; 103 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 104 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 105 for (j=0;j<graph->count[node_num];j++) { 106 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 107 } 108 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 109 if (verbosity_level > 1) { 110 printcc = PETSC_TRUE; 111 } else if (graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 112 printcc = PETSC_TRUE; 113 } 114 if (printcc) { 115 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 116 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 117 } 118 } 119 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 120 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 121 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 122 } 123 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 124 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 125 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 126 } 127 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 128 PetscFunctionReturn(0); 129 } 130 131 #undef __FUNCT__ 132 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 133 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 134 { 135 IS *ISForFaces,*ISForEdges,ISForVertices; 136 PetscInt i,nfc,nec,nvc,*idx; 137 PetscErrorCode ierr; 138 139 PetscFunctionBegin; 140 /* loop on ccs to evalute number of faces, edges and vertices */ 141 nfc = 0; 142 nec = 0; 143 nvc = 0; 144 for (i=0;i<graph->ncc;i++) { 145 PetscInt repdof = graph->queue[graph->cptr[i]]; 146 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 147 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 148 nfc++; 149 } else { /* note that nec will be zero in 2d */ 150 nec++; 151 } 152 } else { 153 nvc += graph->cptr[i+1]-graph->cptr[i]; 154 } 155 } 156 157 /* check if we are in 2D or 3D */ 158 if (graph->twodim) { /* we are in a 2D case -> edges are shared by 2 subregions and faces don't exist */ 159 nec = nfc; 160 nfc = 0; 161 } 162 163 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 164 if (FacesIS) { 165 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 166 } 167 if (EdgesIS) { 168 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 169 } 170 if (VerticesIS) { 171 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 172 } 173 174 /* loop on ccs to compute index sets for faces and edges */ 175 nfc = 0; 176 nec = 0; 177 for (i=0;i<graph->ncc;i++) { 178 PetscInt repdof = graph->queue[graph->cptr[i]]; 179 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 180 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 181 if (graph->twodim) { 182 if (EdgesIS) { 183 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 184 } 185 nec++; 186 } else { 187 if (FacesIS) { 188 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 189 } 190 nfc++; 191 } 192 } else { 193 if (EdgesIS) { 194 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 195 } 196 nec++; 197 } 198 } 199 } 200 /* index set for vertices */ 201 if (VerticesIS) { 202 nvc = 0; 203 for (i=0;i<graph->ncc;i++) { 204 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 205 PetscInt j; 206 207 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 208 idx[nvc]=graph->queue[j]; 209 nvc++; 210 } 211 } 212 } 213 /* sort vertex set (by local ordering) */ 214 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 215 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 216 } 217 /* get back info */ 218 if (n_faces) *n_faces = nfc; 219 if (FacesIS) { 220 *FacesIS = ISForFaces; 221 } 222 if (n_edges) *n_edges = nec; 223 if (EdgesIS) { 224 *EdgesIS = ISForEdges; 225 } 226 if (VerticesIS) { 227 *VerticesIS = ISForVertices; 228 } 229 PetscFunctionReturn(0); 230 } 231 232 #undef __FUNCT__ 233 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 234 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 235 { 236 PetscBool adapt_interface_reduced; 237 MPI_Comm interface_comm; 238 PetscMPIInt size; 239 PetscInt i; 240 PetscBool twodim; 241 PetscErrorCode ierr; 242 243 PetscFunctionBegin; 244 /* compute connected components locally */ 245 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 246 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 247 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 248 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 249 adapt_interface_reduced = PETSC_FALSE; 250 if (size > 1) { 251 PetscInt i; 252 PetscBool adapt_interface = PETSC_FALSE; 253 for (i=0;i<graph->n_subsets;i++) { 254 /* We are not sure that on a given subset of the local interface, 255 with two connected components, the latters be the same among sharing subdomains */ 256 if (graph->subset_ncc[i] > 1) { 257 adapt_interface = PETSC_TRUE; 258 break; 259 } 260 } 261 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 262 } 263 264 if (graph->n_subsets && adapt_interface_reduced) { 265 /* old csr stuff */ 266 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy; 267 PetscInt *old_xadj,*old_adjncy; 268 PetscBT subset_cc_adapt; 269 /* MPI */ 270 MPI_Request *send_requests,*recv_requests; 271 PetscInt *send_buffer,*recv_buffer; 272 PetscInt sum_requests,start_of_recv,start_of_send; 273 PetscInt *cum_recv_counts; 274 /* others */ 275 PetscInt *labels; 276 PetscInt global_subset_counter; 277 PetscInt j,k,s; 278 279 /* Retrict adjacency graph using information from previously computed connected components */ 280 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 281 for (i=0;i<graph->nvtxs;i++) { 282 aux_new_xadj[i]=1; 283 } 284 for (i=0;i<graph->ncc;i++) { 285 k = graph->cptr[i+1]-graph->cptr[i]; 286 for (j=0;j<k;j++) { 287 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 288 } 289 } 290 j = 0; 291 for (i=0;i<graph->nvtxs;i++) { 292 j += aux_new_xadj[i]; 293 } 294 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 295 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 296 new_xadj[0]=0; 297 for (i=0;i<graph->nvtxs;i++) { 298 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 299 if (aux_new_xadj[i]==1) { 300 new_adjncy[new_xadj[i]]=i; 301 } 302 } 303 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 304 ierr = PetscMalloc1(graph->nvtxs,&labels);CHKERRQ(ierr); 305 ierr = PetscMemzero(labels,graph->nvtxs*sizeof(*labels));CHKERRQ(ierr); 306 for (i=0;i<graph->ncc;i++) { 307 k = graph->cptr[i+1]-graph->cptr[i]; 308 for (j=0;j<k;j++) { 309 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 310 labels[graph->queue[graph->cptr[i]+j]] = i; 311 } 312 } 313 /* set temporarly new CSR into graph */ 314 old_xadj = graph->xadj; 315 old_adjncy = graph->adjncy; 316 graph->xadj = new_xadj; 317 graph->adjncy = new_adjncy; 318 /* allocate some space */ 319 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 320 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 321 /* first count how many neighbours per connected component I will receive from */ 322 cum_recv_counts[0]=0; 323 for (i=0;i<graph->n_subsets;i++) { 324 cum_recv_counts[i+1]=cum_recv_counts[i]+graph->count[graph->subsets[i][0]]; 325 } 326 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_buffer);CHKERRQ(ierr); 327 ierr = PetscMalloc2(cum_recv_counts[graph->n_subsets],&send_requests,cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 328 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 329 send_requests[i]=MPI_REQUEST_NULL; 330 recv_requests[i]=MPI_REQUEST_NULL; 331 } 332 /* exchange with my neighbours the number of my connected components on the subset of interface */ 333 sum_requests = 0; 334 for (i=0;i<graph->n_subsets;i++) { 335 PetscMPIInt neigh,tag; 336 337 j = graph->subsets[i][0]; 338 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 339 for (k=0;k<graph->count[j];k++) { 340 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 341 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 342 ierr = MPI_Irecv(&recv_buffer[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 343 sum_requests++; 344 } 345 } 346 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 347 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 348 /* determine the subsets I have to adapt */ 349 ierr = PetscBTCreate(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 350 ierr = PetscBTMemzero(graph->n_subsets,subset_cc_adapt);CHKERRQ(ierr); 351 for (i=0;i<graph->n_subsets;i++) { 352 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 353 /* adapt if more than one cc is present */ 354 if (graph->subset_ncc[i] > 1 || recv_buffer[j] > 1) { 355 ierr = PetscBTSet(subset_cc_adapt,i); 356 break; 357 } 358 } 359 } 360 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 361 /* determine send/recv buffers sizes */ 362 j = 0; 363 for (i=0;i<graph->n_subsets;i++) { 364 if (PetscBTLookup(subset_cc_adapt,i)) { 365 j += graph->subsets_size[i]; 366 } 367 } 368 k = 0; 369 for (i=0;i<graph->n_subsets;i++) { 370 if (PetscBTLookup(subset_cc_adapt,i)) { 371 k += (cum_recv_counts[i+1]-cum_recv_counts[i])*graph->subsets_size[i]; 372 } 373 } 374 ierr = PetscMalloc2(j,&send_buffer,k,&recv_buffer);CHKERRQ(ierr); 375 376 /* fill send buffer */ 377 j = 0; 378 for (i=0;i<graph->n_subsets;i++) { 379 if (PetscBTLookup(subset_cc_adapt,i)) { 380 for (k=0;k<graph->subsets_size[i];k++) { 381 send_buffer[j++] = labels[graph->subsets[i][k]]; 382 } 383 } 384 } 385 ierr = PetscFree(labels);CHKERRQ(ierr); 386 387 /* now exchange the data */ 388 start_of_recv = 0; 389 start_of_send = 0; 390 sum_requests = 0; 391 for (i=0;i<graph->n_subsets;i++) { 392 if (PetscBTLookup(subset_cc_adapt,i)) { 393 PetscMPIInt neigh,tag; 394 PetscInt size_of_send = graph->subsets_size[i]; 395 396 j = graph->subsets[i][0]; 397 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 398 for (k=0;k<graph->count[j];k++) { 399 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 400 ierr = MPI_Isend(&send_buffer[start_of_send],size_of_send,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 401 ierr = MPI_Irecv(&recv_buffer[start_of_recv],size_of_send,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 402 start_of_recv += size_of_send; 403 sum_requests++; 404 } 405 start_of_send += size_of_send; 406 } 407 } 408 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 409 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 410 ierr = PetscFree2(send_requests,recv_requests);CHKERRQ(ierr); 411 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 412 413 start_of_recv = 0; 414 global_subset_counter = 0; 415 for (i=0;i<graph->n_subsets;i++) { 416 if (PetscBTLookup(subset_cc_adapt,i)) { 417 PetscInt **temp_buffer,temp_buffer_size,*private_label; 418 419 /* allocate some temporary space */ 420 temp_buffer_size = graph->subsets_size[i]; 421 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 422 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 423 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 424 for (j=1;j<temp_buffer_size;j++) { 425 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 426 } 427 /* analyze contributions from neighbouring subdomains for i-th subset 428 temp buffer structure: 429 supposing part of the interface has dimension 5 430 e.g with global dofs 0,1,2,3,4, locally ordered on the current process as 0,4,3,1,2 431 3 neighs procs having connected components: 432 neigh 0: [0 1 4], [2 3], labels [4,7] (2 connected components) 433 neigh 1: [0 1], [2 3 4], labels [3 2] (2 connected components) 434 neigh 2: [0 4], [1], [2 3], labels [1 5 6] (3 connected components) 435 tempbuffer (row-oriented) will be filled as: 436 [ 4, 3, 1; 437 4, 2, 1; 438 7, 2, 6; 439 4, 3, 5; 440 7, 2, 6; ]; 441 This way we can simply find intersections of ccs among neighs. 442 The graph->subset array will need to be modified. The output for the example is [0], [1], [2 3], [4]; 443 */ 444 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 445 for (k=0;k<temp_buffer_size;k++) { 446 temp_buffer[k][j] = recv_buffer[start_of_recv+k]; 447 } 448 start_of_recv += temp_buffer_size; 449 } 450 ierr = PetscMalloc1(temp_buffer_size,&private_label);CHKERRQ(ierr); 451 ierr = PetscMemzero(private_label,temp_buffer_size*sizeof(*private_label));CHKERRQ(ierr); 452 for (j=0;j<temp_buffer_size;j++) { 453 if (!private_label[j]) { /* found a new cc */ 454 PetscBool same_set; 455 456 global_subset_counter++; 457 private_label[j] = global_subset_counter; 458 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 459 same_set = PETSC_TRUE; 460 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 461 if (temp_buffer[j][s] != temp_buffer[k][s]) { 462 same_set = PETSC_FALSE; 463 break; 464 } 465 } 466 if (same_set) { 467 private_label[k] = global_subset_counter; 468 } 469 } 470 } 471 } 472 /* insert private labels in graph structure */ 473 for (j=0;j<graph->subsets_size[i];j++) { 474 graph->subset[graph->subsets[i][j]] = graph->n_subsets+private_label[j]; 475 } 476 ierr = PetscFree(private_label);CHKERRQ(ierr); 477 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 478 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 479 } 480 } 481 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 482 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 483 ierr = PetscBTDestroy(&subset_cc_adapt);CHKERRQ(ierr); 484 /* We are ready to find for connected components which are consistent among neighbouring subdomains */ 485 if (global_subset_counter) { 486 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 487 global_subset_counter = 0; 488 for (i=0;i<graph->nvtxs;i++) { 489 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 490 global_subset_counter++; 491 for (j=i+1;j<graph->nvtxs;j++) { 492 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 493 graph->subset[j] = global_subset_counter; 494 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 495 } 496 } 497 graph->subset[i] = global_subset_counter; 498 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 499 } 500 } 501 /* refine connected components locally */ 502 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 503 } 504 /* restore original CSR graph of dofs */ 505 ierr = PetscFree(new_xadj);CHKERRQ(ierr); 506 ierr = PetscFree(new_adjncy);CHKERRQ(ierr); 507 graph->xadj = old_xadj; 508 graph->adjncy = old_adjncy; 509 } 510 511 /* Determine if we are in 2D or 3D */ 512 twodim = PETSC_TRUE; 513 for (i=0;i<graph->ncc;i++) { 514 PetscInt repdof = graph->queue[graph->cptr[i]]; 515 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 516 if (graph->count[repdof] > 1 || graph->special_dof[repdof] == PCBDDCGRAPH_NEUMANN_MARK) { 517 twodim = PETSC_FALSE; 518 break; 519 } 520 } 521 } 522 ierr = MPI_Allreduce(&twodim,&graph->twodim,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 523 PetscFunctionReturn(0); 524 } 525 526 /* The following code has been adapted from function IsConnectedSubdomain contained 527 in source file contig.c of METIS library (version 5.0.1) 528 It finds connected components for each subset */ 529 #undef __FUNCT__ 530 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 531 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 532 { 533 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 534 PetscInt *queue_global; 535 PetscMPIInt size; 536 PetscErrorCode ierr; 537 538 PetscFunctionBegin; 539 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&size);CHKERRQ(ierr); 540 /* quiet return if no csr info is available */ 541 if (!graph->xadj || !graph->adjncy) { 542 PetscFunctionReturn(0); 543 } 544 545 /* reset any previous search of connected components */ 546 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 547 graph->n_subsets = 0; 548 if (size == 1) { 549 if (graph->nvtxs) { 550 graph->n_subsets = 1; 551 for (i=0;i<graph->nvtxs;i++) { 552 graph->subset[i] = 1; 553 } 554 } 555 } else { 556 for (i=0;i<graph->nvtxs;i++) { 557 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 558 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 559 graph->subset[i] = 0; 560 } 561 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 562 } 563 } 564 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 565 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 566 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 567 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 568 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 569 570 /* begin search for connected components */ 571 cum_queue = 0; 572 ncc = 0; 573 for (n=0;n<graph->n_subsets;n++) { 574 pid = n+1; /* partition labeled by 0 is discarded */ 575 nleft = 0; 576 for (i=0;i<graph->nvtxs;i++) { 577 if (graph->subset[i] == pid) { 578 nleft++; 579 } 580 } 581 for (i=0; i<graph->nvtxs; i++) { 582 if (graph->subset[i] == pid) { 583 break; 584 } 585 } 586 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 587 graph->queue[cum_queue] = i; 588 first = 0; 589 last = 1; 590 graph->cptr[ncc] = cum_queue; 591 ncc_pid = 0; 592 while (first != nleft) { 593 if (first == last) { 594 graph->cptr[++ncc] = first+cum_queue; 595 ncc_pid++; 596 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 597 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 598 break; 599 } 600 } 601 graph->queue[cum_queue+last] = i; 602 last++; 603 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 604 } 605 i = graph->queue[cum_queue+first]; 606 first++; 607 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 608 k = graph->adjncy[j]; 609 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 610 graph->queue[cum_queue+last] = k; 611 last++; 612 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 613 } 614 } 615 } 616 graph->cptr[++ncc] = first+cum_queue; 617 ncc_pid++; 618 cum_queue = graph->cptr[ncc]; 619 graph->subset_ncc[n] = ncc_pid; 620 } 621 graph->ncc = ncc; 622 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 623 if (size > 1) { 624 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 625 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 626 for (i=0;i<graph->ncc;i++) { 627 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 628 } 629 ierr = PetscFree(queue_global);CHKERRQ(ierr); 630 } 631 PetscFunctionReturn(0); 632 } 633 634 #undef __FUNCT__ 635 #define __FUNCT__ "PCBDDCGraphSetUp" 636 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 637 { 638 VecScatter scatter_ctx; 639 Vec local_vec,local_vec2,global_vec; 640 IS to,from; 641 MPI_Comm comm; 642 PetscScalar *array,*array2; 643 const PetscInt *is_indices; 644 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global,*subset_ref_node_global; 645 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 646 PetscErrorCode ierr; 647 PetscBool same_set,mirrors_found; 648 649 PetscFunctionBegin; 650 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 651 /* custom_minimal_size */ 652 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 653 /* get info l2gmap and allocate work vectors */ 654 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 655 /* check if we have any local periodic nodes (periodic BCs) */ 656 mirrors_found = PETSC_FALSE; 657 if (graph->nvtxs && n_neigh) { 658 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 659 for (i=0; i<n_shared[0]; i++) { 660 if (graph->count[shared[0][i]] > 1) { 661 mirrors_found = PETSC_TRUE; 662 break; 663 } 664 } 665 } 666 /* create some workspace objects */ 667 local_vec = NULL; 668 local_vec2 = NULL; 669 global_vec = NULL; 670 to = NULL; 671 from = NULL; 672 scatter_ctx = NULL; 673 if (n_ISForDofs || dirichlet_is || neumann_is || custom_primal_vertices) { 674 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 675 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 676 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 677 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 678 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 679 ierr = VecSetSizes(global_vec,PETSC_DECIDE,graph->nvtxs_global);CHKERRQ(ierr); 680 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 681 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 682 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 683 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 684 } else if (mirrors_found) { 685 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 686 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 687 } 688 /* compute local mirrors (if any) */ 689 if (mirrors_found) { 690 PetscInt *local_indices,*global_indices; 691 /* get arrays of local and global indices */ 692 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 693 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 694 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 695 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 696 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 697 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 698 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 699 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 700 /* allocate space for mirrors */ 701 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 702 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 703 graph->mirrors_set[0] = 0; 704 705 k=0; 706 for (i=0;i<n_shared[0];i++) { 707 j=shared[0][i]; 708 if (graph->count[j] > 1) { 709 graph->mirrors[j]++; 710 k++; 711 } 712 } 713 /* allocate space for set of mirrors */ 714 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 715 for (i=1;i<graph->nvtxs;i++) 716 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 717 718 /* fill arrays */ 719 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 720 for (j=0;j<n_shared[0];j++) { 721 i=shared[0][j]; 722 if (graph->count[i] > 1) 723 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 724 } 725 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 726 for (i=0;i<graph->nvtxs;i++) { 727 if (graph->mirrors[i] > 0) { 728 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 729 j = global_indices[k]; 730 while ( k > 0 && global_indices[k-1] == j) k--; 731 for (j=0;j<graph->mirrors[i];j++) { 732 graph->mirrors_set[i][j]=local_indices[k+j]; 733 } 734 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 735 } 736 } 737 ierr = PetscFree(local_indices);CHKERRQ(ierr); 738 ierr = PetscFree(global_indices);CHKERRQ(ierr); 739 } 740 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 741 ierr = ISDestroy(&to);CHKERRQ(ierr); 742 ierr = ISDestroy(&from);CHKERRQ(ierr); 743 744 /* Count total number of neigh per node */ 745 k=0; 746 for (i=1;i<n_neigh;i++) { 747 k += n_shared[i]; 748 for (j=0;j<n_shared[i];j++) { 749 graph->count[shared[i][j]] += 1; 750 } 751 } 752 /* Allocate space for storing the set of neighbours for each node */ 753 if (graph->nvtxs) { 754 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 755 } 756 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 757 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 758 } 759 /* Get information for sharing subdomains */ 760 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 761 for (i=1;i<n_neigh;i++) { /* dont count myself */ 762 s = n_shared[i]; 763 for (j=0;j<s;j++) { 764 k = shared[i][j]; 765 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 766 graph->count[k] += 1; 767 } 768 } 769 /* sort set of sharing subdomains */ 770 for (i=0;i<graph->nvtxs;i++) { 771 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 772 } 773 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 774 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 775 776 /* 777 Get info for dofs splitting 778 User can specify just a subset; an additional field is considered as a complementary field 779 */ 780 for (i=0;i<graph->nvtxs;i++) { 781 graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 782 } 783 if (n_ISForDofs) { 784 ierr = VecSet(local_vec,-1.0);CHKERRQ(ierr); 785 } 786 for (i=0;i<n_ISForDofs;i++) { 787 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 788 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 789 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 790 for (j=0;j<is_size;j++) { 791 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 792 graph->which_dof[is_indices[j]] = i; 793 array[is_indices[j]] = 1.*i; 794 } 795 } 796 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 797 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 798 } 799 /* Check consistency among neighbours */ 800 if (n_ISForDofs) { 801 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 802 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 803 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec2,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 804 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec2,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 805 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 806 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 807 for (i=0;i<graph->nvtxs;i++){ 808 PetscInt field1,field2; 809 810 field1 = (PetscInt)PetscRealPart(array[i]); 811 field2 = (PetscInt)PetscRealPart(array2[i]); 812 if (field1 != field2) { 813 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); 814 } 815 } 816 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 817 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 818 } 819 if (neumann_is || dirichlet_is) { 820 /* Take into account Neumann nodes */ 821 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 822 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 823 if (neumann_is) { 824 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 825 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 826 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 827 for (i=0;i<is_size;i++) { 828 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 829 array[is_indices[i]] = 1.0; 830 } 831 } 832 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 833 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 834 } 835 /* Neumann nodes: impose consistency among neighbours */ 836 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 837 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 838 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 839 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 840 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 841 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 842 for (i=0;i<graph->nvtxs;i++) { 843 if (PetscRealPart(array[i]) > 0.1) { 844 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 845 } 846 } 847 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 848 /* Take into account Dirichlet nodes */ 849 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 850 if (dirichlet_is) { 851 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 852 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 853 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 854 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 855 for (i=0;i<is_size;i++){ 856 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 857 k = is_indices[i]; 858 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 859 if (PetscRealPart(array[k]) > 0.1) { 860 SETERRQ1(comm,PETSC_ERR_USER,"BDDC cannot have boundary nodes which are marked Neumann and Dirichlet at the same time! Local node %d is wrong\n",k); 861 } 862 array2[k] = 1.0; 863 } 864 } 865 } 866 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 867 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 868 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 869 } 870 /* Dirichlet nodes: impose consistency among neighbours */ 871 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 872 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 873 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 874 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 875 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 876 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 877 for (i=0;i<graph->nvtxs;i++) { 878 if (PetscRealPart(array[i]) > 0.1) { 879 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 880 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 881 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 882 } 883 } 884 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 885 } 886 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 887 if (graph->mirrors) { 888 for (i=0;i<graph->nvtxs;i++) 889 if (graph->mirrors[i]) 890 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 891 892 if (graph->xadj && graph->adjncy) { 893 PetscInt *new_xadj,*new_adjncy; 894 /* sort CSR graph */ 895 for (i=0;i<graph->nvtxs;i++) 896 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 897 898 /* adapt local CSR graph in case of local periodicity */ 899 k=0; 900 for (i=0;i<graph->nvtxs;i++) 901 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 902 k += graph->mirrors[graph->adjncy[j]]; 903 904 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 905 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 906 new_xadj[0]=0; 907 for (i=0;i<graph->nvtxs;i++) { 908 k = graph->xadj[i+1]-graph->xadj[i]; 909 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 910 new_xadj[i+1]=new_xadj[i]+k; 911 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 912 k = graph->mirrors[graph->adjncy[j]]; 913 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 914 new_xadj[i+1]+=k; 915 } 916 k = new_xadj[i+1]-new_xadj[i]; 917 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 918 new_xadj[i+1]=new_xadj[i]+k; 919 } 920 /* set new CSR into graph */ 921 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 922 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 923 graph->xadj = new_xadj; 924 graph->adjncy = new_adjncy; 925 } 926 } 927 928 /* mark special nodes (if any) -> each will become a single node equivalence class */ 929 if (custom_primal_vertices) { 930 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 931 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 932 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 933 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 934 for (i=0;i<is_size;i++){ 935 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 936 array[is_indices[i]] = 1.0; 937 } 938 } 939 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 940 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 941 /* special nodes: impose consistency among neighbours */ 942 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 943 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 944 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 945 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 946 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 947 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 948 j = 0; 949 for (i=0;i<graph->nvtxs;i++) { 950 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 951 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 952 j++; 953 } 954 } 955 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 956 } 957 958 /* mark interior nodes as touched and belonging to partition number 0 */ 959 for (i=0;i<graph->nvtxs;i++) { 960 if (!graph->count[i]) { 961 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 962 graph->subset[i] = 0; 963 } 964 } 965 /* init graph structure and compute default subsets */ 966 nodes_touched=0; 967 for (i=0;i<graph->nvtxs;i++) { 968 if (PetscBTLookup(graph->touched,i)) { 969 nodes_touched++; 970 } 971 } 972 i = 0; 973 graph->ncc = 0; 974 total_counts = 0; 975 while (nodes_touched<graph->nvtxs) { 976 /* find first untouched node in local ordering */ 977 while (PetscBTLookup(graph->touched,i)) { 978 i++; 979 } 980 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 981 graph->subset[i] = graph->ncc+1; 982 graph->cptr[graph->ncc] = total_counts; 983 graph->queue[total_counts] = i; 984 total_counts++; 985 nodes_touched++; 986 /* now find all other nodes having the same set of sharing subdomains */ 987 for (j=i+1;j<graph->nvtxs;j++) { 988 /* check for same number of sharing subdomains, dof number and same special mark */ 989 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]) { 990 /* check for same set of sharing subdomains */ 991 same_set=PETSC_TRUE; 992 for (k=0;k<graph->count[j];k++){ 993 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 994 same_set=PETSC_FALSE; 995 } 996 } 997 /* I found a friend of mine */ 998 if (same_set) { 999 graph->subset[j]=graph->ncc+1; 1000 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1001 nodes_touched++; 1002 graph->queue[total_counts] = j; 1003 total_counts++; 1004 } 1005 } 1006 } 1007 graph->ncc++; 1008 } 1009 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 1010 graph->n_subsets = graph->ncc; 1011 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1012 for (i=0;i<graph->n_subsets;i++) { 1013 graph->subset_ncc[i] = 1; 1014 } 1015 /* final pointer */ 1016 graph->cptr[graph->ncc] = total_counts; 1017 1018 /* save information on subsets (needed if we have to adapt the connected components later) */ 1019 /* For consistency reasons (among neighbours), I need to sort (by global ordering) each subset */ 1020 /* Get a reference node (min index in global ordering) for each subset for tagging messages */ 1021 ierr = PetscMalloc2(graph->ncc,&subset_ref_node_global,graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1022 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1023 for (j=0;j<graph->ncc;j++) { 1024 ierr = PetscSortIntWithArray(graph->cptr[j+1]-graph->cptr[j],&queue_global[graph->cptr[j]],&graph->queue[graph->cptr[j]]);CHKERRQ(ierr); 1025 subset_ref_node_global[j] = graph->queue[graph->cptr[j]]; 1026 } 1027 ierr = PetscMalloc2(graph->ncc,&graph->subsets_size,graph->ncc,&graph->subsets);CHKERRQ(ierr); 1028 ierr = PetscMalloc1(graph->cptr[graph->ncc],&graph->subsets[0]);CHKERRQ(ierr); 1029 ierr = PetscMemzero(graph->subsets[0],graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1030 for (j=1;j<graph->ncc;j++) { 1031 graph->subsets_size[j-1] = graph->cptr[j] - graph->cptr[j-1]; 1032 graph->subsets[j] = graph->subsets[j-1] + graph->subsets_size[j-1]; 1033 } 1034 graph->subsets_size[graph->ncc-1] = graph->cptr[graph->ncc] - graph->cptr[graph->ncc-1]; 1035 for (j=0;j<graph->ncc;j++) { 1036 ierr = PetscMemcpy(graph->subsets[j],&graph->queue[graph->cptr[j]],graph->subsets_size[j]*sizeof(PetscInt));CHKERRQ(ierr); 1037 } 1038 1039 /* renumber reference nodes */ 1040 ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->l2gmap,graph->ncc,subset_ref_node_global,NULL,&k,&graph->subset_ref_node);CHKERRQ(ierr); 1041 1042 /* free workspace */ 1043 ierr = PetscFree2(subset_ref_node_global,queue_global);CHKERRQ(ierr); 1044 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1045 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1046 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1047 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1048 PetscFunctionReturn(0); 1049 } 1050 1051 #undef __FUNCT__ 1052 #define __FUNCT__ "PCBDDCGraphResetCSR" 1053 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1054 { 1055 PetscErrorCode ierr; 1056 1057 PetscFunctionBegin; 1058 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1059 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1060 graph->nvtxs_csr = 0; 1061 PetscFunctionReturn(0); 1062 } 1063 1064 #undef __FUNCT__ 1065 #define __FUNCT__ "PCBDDCGraphReset" 1066 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1067 { 1068 PetscErrorCode ierr; 1069 1070 PetscFunctionBegin; 1071 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1072 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1073 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1074 if (graph->nvtxs) { 1075 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1076 } 1077 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1078 ierr = PetscFree7(graph->count, 1079 graph->neighbours_set, 1080 graph->subset, 1081 graph->which_dof, 1082 graph->cptr, 1083 graph->queue, 1084 graph->special_dof);CHKERRQ(ierr); 1085 if (graph->mirrors) { 1086 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1087 } 1088 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1089 if (graph->subsets) { 1090 ierr = PetscFree(graph->subsets[0]);CHKERRQ(ierr); 1091 } 1092 ierr = PetscFree2(graph->subsets_size,graph->subsets);CHKERRQ(ierr); 1093 graph->nvtxs = 0; 1094 graph->n_subsets = 0; 1095 graph->custom_minimal_size = 1; 1096 PetscFunctionReturn(0); 1097 } 1098 1099 #undef __FUNCT__ 1100 #define __FUNCT__ "PCBDDCGraphInit" 1101 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap, PetscInt N) 1102 { 1103 PetscInt n; 1104 PetscErrorCode ierr; 1105 1106 PetscFunctionBegin; 1107 PetscValidPointer(graph,1); 1108 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1109 /* raise an error if already allocated */ 1110 if (graph->nvtxs) { 1111 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1112 } 1113 /* set number of vertices */ 1114 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1115 graph->l2gmap = l2gmap; 1116 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1117 graph->nvtxs = n; 1118 graph->nvtxs_global = N; 1119 /* allocate used space */ 1120 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1121 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1122 graph->nvtxs,&graph->neighbours_set, 1123 graph->nvtxs,&graph->subset, 1124 graph->nvtxs,&graph->which_dof, 1125 graph->nvtxs+1,&graph->cptr, 1126 graph->nvtxs,&graph->queue, 1127 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1128 /* zeroes memory */ 1129 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1130 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1131 /* use -1 as a default value for which_dof array */ 1132 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1133 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1134 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1135 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1136 /* zeroes first pointer to neighbour set */ 1137 if (graph->nvtxs) { 1138 graph->neighbours_set[0] = 0; 1139 } 1140 /* zeroes workspace for values of ncc */ 1141 graph->subset_ncc = 0; 1142 graph->subset_ref_node = 0; 1143 PetscFunctionReturn(0); 1144 } 1145 1146 #undef __FUNCT__ 1147 #define __FUNCT__ "PCBDDCGraphDestroy" 1148 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1149 { 1150 PetscErrorCode ierr; 1151 1152 PetscFunctionBegin; 1153 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1154 ierr = PetscFree(*graph);CHKERRQ(ierr); 1155 PetscFunctionReturn(0); 1156 } 1157 1158 #undef __FUNCT__ 1159 #define __FUNCT__ "PCBDDCGraphCreate" 1160 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1161 { 1162 PCBDDCGraph new_graph; 1163 PetscErrorCode ierr; 1164 1165 PetscFunctionBegin; 1166 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1167 new_graph->custom_minimal_size = 1; 1168 *graph = new_graph; 1169 PetscFunctionReturn(0); 1170 } 1171