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