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 } 88 } 89 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 90 if (graph->subset[i] && graph->subset_ncc) { 91 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 92 } 93 } 94 } 95 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 96 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 97 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 98 for (i=0;i<graph->ncc;i++) { 99 PetscInt node_num=graph->queue[graph->cptr[i]]; 100 PetscBool printcc = PETSC_FALSE; 101 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 102 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 103 for (j=0;j<graph->count[node_num];j++) { 104 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 105 } 106 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 107 if (verbosity_level > 1) { 108 printcc = PETSC_TRUE; 109 } else if (graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 110 printcc = PETSC_TRUE; 111 } 112 if (printcc) { 113 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 114 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 115 } 116 } 117 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 118 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 119 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 120 } 121 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 122 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 123 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 124 } 125 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 126 PetscFunctionReturn(0); 127 } 128 129 #undef __FUNCT__ 130 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 131 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 132 { 133 IS *ISForFaces,*ISForEdges,ISForVertices; 134 PetscInt i,nfc,nec,nvc,*idx; 135 PetscErrorCode ierr; 136 137 PetscFunctionBegin; 138 /* loop on ccs to evalute number of faces, edges and vertices */ 139 nfc = 0; 140 nec = 0; 141 nvc = 0; 142 for (i=0;i<graph->ncc;i++) { 143 PetscInt repdof = graph->queue[graph->cptr[i]]; 144 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 145 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 146 nfc++; 147 } else { /* note that nec will be zero in 2d */ 148 nec++; 149 } 150 } else { 151 nvc += graph->cptr[i+1]-graph->cptr[i]; 152 } 153 } 154 155 /* check if we are in 2D or 3D */ 156 if (graph->twodim) { /* we are in a 2D case -> edges are shared by 2 subregions and faces don't exist */ 157 nec = nfc; 158 nfc = 0; 159 } 160 161 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 162 if (FacesIS) { 163 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 164 } 165 if (EdgesIS) { 166 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 167 } 168 if (VerticesIS) { 169 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 170 } 171 172 /* loop on ccs to compute index sets for faces and edges */ 173 nfc = 0; 174 nec = 0; 175 for (i=0;i<graph->ncc;i++) { 176 PetscInt repdof = graph->queue[graph->cptr[i]]; 177 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 178 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 179 if (graph->twodim) { 180 if (EdgesIS) { 181 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 182 } 183 nec++; 184 } else { 185 if (FacesIS) { 186 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 187 } 188 nfc++; 189 } 190 } else { 191 if (EdgesIS) { 192 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 193 } 194 nec++; 195 } 196 } 197 } 198 /* index set for vertices */ 199 if (VerticesIS) { 200 nvc = 0; 201 for (i=0;i<graph->ncc;i++) { 202 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 203 PetscInt j; 204 205 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 206 idx[nvc]=graph->queue[j]; 207 nvc++; 208 } 209 } 210 } 211 /* sort vertex set (by local ordering) */ 212 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 213 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 214 } 215 /* get back info */ 216 if (n_faces) *n_faces = nfc; 217 if (FacesIS) { 218 *FacesIS = ISForFaces; 219 } 220 if (n_edges) *n_edges = nec; 221 if (EdgesIS) { 222 *EdgesIS = ISForEdges; 223 } 224 if (VerticesIS) { 225 *VerticesIS = ISForVertices; 226 } 227 PetscFunctionReturn(0); 228 } 229 230 #undef __FUNCT__ 231 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 232 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 233 { 234 PetscBool adapt_interface_reduced; 235 MPI_Comm interface_comm; 236 PetscMPIInt size; 237 PetscInt i; 238 PetscBool twodim; 239 PetscErrorCode ierr; 240 241 PetscFunctionBegin; 242 /* compute connected components locally */ 243 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 244 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 245 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 246 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 247 adapt_interface_reduced = PETSC_FALSE; 248 if (size > 1) { 249 PetscInt i; 250 PetscBool adapt_interface = PETSC_FALSE; 251 for (i=0;i<graph->n_subsets;i++) { 252 /* We are not sure that on a given subset of the local interface, 253 with two connected components, the latters be the same among sharing subdomains */ 254 if (graph->subset_ncc[i] > 1) { 255 adapt_interface = PETSC_TRUE; 256 break; 257 } 258 } 259 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 260 } 261 262 if (graph->n_subsets && adapt_interface_reduced) { 263 MPI_Request *send_requests; 264 MPI_Request *recv_requests; 265 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 266 PetscInt *old_xadj,*old_adjncy; 267 PetscInt j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 268 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 269 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 270 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 271 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 272 PetscBool *subset_cc_adapt,same_set; 273 274 /* Retrict adjacency graph using information from previously computed connected components */ 275 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 276 for (i=0;i<graph->nvtxs;i++) { 277 aux_new_xadj[i]=1; 278 } 279 for (i=0;i<graph->ncc;i++) { 280 k = graph->cptr[i+1]-graph->cptr[i]; 281 for (j=0;j<k;j++) { 282 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 283 } 284 } 285 j = 0; 286 for (i=0;i<graph->nvtxs;i++) { 287 j += aux_new_xadj[i]; 288 } 289 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 290 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 291 new_xadj[0]=0; 292 for (i=0;i<graph->nvtxs;i++) { 293 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 294 if (aux_new_xadj[i]==1) { 295 new_adjncy[new_xadj[i]]=i; 296 } 297 } 298 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 299 for (i=0;i<graph->ncc;i++) { 300 k = graph->cptr[i+1]-graph->cptr[i]; 301 for (j=0;j<k;j++) { 302 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 303 } 304 } 305 /* set temporarly new CSR into graph */ 306 old_xadj = graph->xadj; 307 old_adjncy = graph->adjncy; 308 graph->xadj = new_xadj; 309 graph->adjncy = new_adjncy; 310 /* allocate some space */ 311 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 312 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 313 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 314 ierr = PetscMalloc1(graph->n_subsets,&subset_to_nodes_indices);CHKERRQ(ierr); 315 /* first count how many neighbours per connected component I will receive from */ 316 cum_recv_counts[0]=0; 317 for (i=1;i<graph->n_subsets+1;i++) { 318 j = 0; 319 while (graph->subset[j] != i) { 320 j++; 321 } 322 subset_to_nodes_indices[i-1]=j; 323 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 324 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 325 } 326 ierr = PetscMalloc1(2*cum_recv_counts[graph->n_subsets],&recv_buffer_subset);CHKERRQ(ierr); 327 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&send_requests);CHKERRQ(ierr); 328 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 329 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 330 send_requests[i]=MPI_REQUEST_NULL; 331 recv_requests[i]=MPI_REQUEST_NULL; 332 } 333 /* exchange with my neighbours the number of my connected components on the shared interface */ 334 sum_requests = 0; 335 for (i=0;i<graph->n_subsets;i++) { 336 j = subset_to_nodes_indices[i]; 337 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 338 for (k=0;k<graph->count[j];k++) { 339 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 340 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 341 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 342 sum_requests++; 343 } 344 } 345 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 346 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 347 /* determine the connected component I need to adapt */ 348 ierr = PetscMalloc1(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 349 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 350 for (i=0;i<graph->n_subsets;i++) { 351 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 352 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 353 if (graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1) { 354 subset_cc_adapt[i] = PETSC_TRUE; 355 break; 356 } 357 } 358 } 359 buffer_size = 0; 360 for (i=0;i<graph->n_subsets;i++) { 361 if (subset_cc_adapt[i]) { 362 for (j=i;j<graph->ncc;j++) { 363 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 364 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 365 } 366 } 367 } 368 } 369 ierr = PetscMalloc1(buffer_size,&send_buffer);CHKERRQ(ierr); 370 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 371 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 372 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 373 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 374 ierr = PetscMalloc1(graph->n_subsets,&sizes_of_sends);CHKERRQ(ierr); 375 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 376 sum_requests = 0; 377 start_of_send = 0; 378 start_of_recv = cum_recv_counts[graph->n_subsets]; 379 for (i=0;i<graph->n_subsets;i++) { 380 if (subset_cc_adapt[i]) { 381 size_of_send = 0; 382 for (j=i;j<graph->ncc;j++) { 383 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 384 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 385 size_of_send += 1; 386 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 387 &queue_in_global_numbering[graph->cptr[j]], 388 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 389 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 390 } 391 } 392 j = subset_to_nodes_indices[i]; 393 sizes_of_sends[i] = size_of_send; 394 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 395 for (k=0;k<graph->count[j];k++) { 396 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 397 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 398 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 399 sum_requests++; 400 } 401 start_of_send += size_of_send; 402 } 403 } 404 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 405 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 406 buffer_size = 0; 407 for (k=0;k<sum_requests;k++) { 408 buffer_size += recv_buffer_subset[start_of_recv+k]; 409 } 410 ierr = PetscMalloc1(buffer_size,&recv_buffer);CHKERRQ(ierr); 411 /* now exchange the data */ 412 start_of_recv = 0; 413 start_of_send = 0; 414 sum_requests = 0; 415 for (i=0;i<graph->n_subsets;i++) { 416 if (subset_cc_adapt[i]) { 417 size_of_send = sizes_of_sends[i]; 418 j = subset_to_nodes_indices[i]; 419 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+2,&tag);CHKERRQ(ierr); 420 for (k=0;k<graph->count[j];k++) { 421 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 422 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 423 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 424 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 425 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 426 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 427 start_of_recv += size_of_recv; 428 sum_requests++; 429 } 430 start_of_send += size_of_send; 431 } 432 } 433 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 434 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 435 for (j=0;j<buffer_size;) { 436 PetscInt rbs = recv_buffer[j]; 437 438 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_DROP,rbs,&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 439 if (rbs != recv_buffer[j]) { 440 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Mapped received buffer size %d differs from input size %d\n",rbs,recv_buffer[j]); 441 } 442 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 443 if (graph->mirrors) { 444 PetscBool mirrored_found=PETSC_FALSE; 445 for (k=0;k<recv_buffer[j];k++) { 446 if (graph->mirrors[recv_buffer[j+k+1]]) { 447 mirrored_found=PETSC_TRUE; 448 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 449 } 450 } 451 if (mirrored_found) { 452 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 453 k=0; 454 while (k<recv_buffer[j]) { 455 PetscInt nmir = graph->mirrors[recv_buffer[j+1+k]]; 456 for (s=1;s<nmir;s++) { 457 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 458 } 459 k += nmir; 460 } 461 } 462 } 463 k = recv_buffer[j]+1; 464 j += k; 465 } 466 sum_requests = cum_recv_counts[graph->n_subsets]; 467 start_of_recv = 0; 468 ierr = PetscMalloc1(graph->nvtxs,&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 469 global_subset_counter = 0; 470 for (i=0;i<graph->n_subsets;i++) { 471 if (subset_cc_adapt[i]) { 472 temp_buffer_size = 0; 473 /* find nodes on the shared interface we need to adapt */ 474 for (j=0;j<graph->nvtxs;j++) { 475 if (graph->subset[j]==i+1) { 476 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 477 temp_buffer_size++; 478 } else { 479 nodes_to_temp_buffer_indices[j] = -1; 480 } 481 } 482 /* allocate some temporary space */ 483 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 484 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 485 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 486 for (j=1;j<temp_buffer_size;j++) { 487 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 488 } 489 /* analyze contributions from neighbouring subdomains for i-th conn comp 490 temp buffer structure: 491 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 492 3 neighs procs with structured connected components: 493 neigh 0: [0 1 4], [2 3]; (2 connected components) 494 neigh 1: [0 1], [2 3 4]; (2 connected components) 495 neigh 2: [0 4], [1], [2 3]; (3 connected components) 496 tempbuffer (row-oriented) will be filled as: 497 [ 0, 0, 0; 498 0, 0, 1; 499 1, 1, 2; 500 1, 1, 2; 501 0, 1, 0; ]; 502 This way we can simply find intersections of ccs among neighs. 503 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 504 */ 505 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 506 ins_val = 0; 507 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 508 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 509 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 510 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 511 } 512 buffer_size += k; 513 ins_val++; 514 } 515 start_of_recv += size_of_recv; 516 sum_requests++; 517 } 518 ierr = PetscMalloc1(temp_buffer_size,&add_to_subset);CHKERRQ(ierr); 519 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 520 for (j=0;j<temp_buffer_size;j++) { 521 if (!add_to_subset[j]) { /* found a new cc */ 522 global_subset_counter++; 523 add_to_subset[j] = global_subset_counter; 524 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 525 same_set = PETSC_TRUE; 526 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 527 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 528 same_set = PETSC_FALSE; 529 break; 530 } 531 } 532 if (same_set) { 533 add_to_subset[k] = global_subset_counter; 534 } 535 } 536 } 537 } 538 /* insert new data in subset array */ 539 temp_buffer_size = 0; 540 for (j=0;j<graph->nvtxs;j++) { 541 if (graph->subset[j]==i+1) { 542 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 543 temp_buffer_size++; 544 } 545 } 546 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 547 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 548 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 549 } 550 } 551 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 552 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 553 ierr = PetscFree(send_requests);CHKERRQ(ierr); 554 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 555 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 556 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 557 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 558 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 559 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 560 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 561 /* We are ready to find for connected components consistent among neighbouring subdomains */ 562 if (global_subset_counter) { 563 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 564 global_subset_counter = 0; 565 for (i=0;i<graph->nvtxs;i++) { 566 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 567 global_subset_counter++; 568 for (j=i+1;j<graph->nvtxs;j++) { 569 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 570 graph->subset[j] = global_subset_counter; 571 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 572 } 573 } 574 graph->subset[i] = global_subset_counter; 575 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 576 } 577 } 578 /* refine connected components locally */ 579 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 580 } 581 /* restore original CSR graph of dofs */ 582 ierr = PetscFree(new_xadj);CHKERRQ(ierr); 583 ierr = PetscFree(new_adjncy);CHKERRQ(ierr); 584 graph->xadj = old_xadj; 585 graph->adjncy = old_adjncy; 586 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 587 } 588 589 /* Determine if we are in 2D or 3D */ 590 twodim = PETSC_TRUE; 591 for (i=0;i<graph->ncc;i++) { 592 PetscInt repdof = graph->queue[graph->cptr[i]]; 593 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 594 if (graph->count[repdof] > 1 || graph->special_dof[repdof] == PCBDDCGRAPH_NEUMANN_MARK) { 595 twodim = PETSC_FALSE; 596 break; 597 } 598 } 599 } 600 ierr = MPI_Allreduce(&twodim,&graph->twodim,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 601 PetscFunctionReturn(0); 602 } 603 604 /* The following code has been adapted from function IsConnectedSubdomain contained 605 in source file contig.c of METIS library (version 5.0.1) 606 It finds connected components for each subset */ 607 #undef __FUNCT__ 608 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 609 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 610 { 611 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 612 PetscInt *queue_global; 613 PetscMPIInt size; 614 PetscErrorCode ierr; 615 616 PetscFunctionBegin; 617 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&size);CHKERRQ(ierr); 618 /* quiet return if no csr info is available */ 619 if (!graph->xadj || !graph->adjncy) { 620 PetscFunctionReturn(0); 621 } 622 623 /* reset any previous search of connected components */ 624 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 625 graph->n_subsets = 0; 626 if (size == 1) { 627 if (graph->nvtxs) { 628 graph->n_subsets = 1; 629 for (i=0;i<graph->nvtxs;i++) { 630 graph->subset[i] = 1; 631 } 632 } 633 } else { 634 for (i=0;i<graph->nvtxs;i++) { 635 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 636 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 637 graph->subset[i] = 0; 638 } 639 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 640 } 641 } 642 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 643 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 644 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 645 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 646 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 647 648 /* begin search for connected components */ 649 cum_queue = 0; 650 ncc = 0; 651 for (n=0;n<graph->n_subsets;n++) { 652 pid = n+1; /* partition labeled by 0 is discarded */ 653 nleft = 0; 654 for (i=0;i<graph->nvtxs;i++) { 655 if (graph->subset[i] == pid) { 656 nleft++; 657 } 658 } 659 for (i=0; i<graph->nvtxs; i++) { 660 if (graph->subset[i] == pid) { 661 break; 662 } 663 } 664 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 665 graph->queue[cum_queue] = i; 666 first = 0; 667 last = 1; 668 graph->cptr[ncc] = cum_queue; 669 ncc_pid = 0; 670 while (first != nleft) { 671 if (first == last) { 672 graph->cptr[++ncc] = first+cum_queue; 673 ncc_pid++; 674 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 675 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 676 break; 677 } 678 } 679 graph->queue[cum_queue+last] = i; 680 last++; 681 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 682 } 683 i = graph->queue[cum_queue+first]; 684 first++; 685 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 686 k = graph->adjncy[j]; 687 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 688 graph->queue[cum_queue+last] = k; 689 last++; 690 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 691 } 692 } 693 } 694 graph->cptr[++ncc] = first+cum_queue; 695 ncc_pid++; 696 cum_queue = graph->cptr[ncc]; 697 graph->subset_ncc[n] = ncc_pid; 698 } 699 graph->ncc = ncc; 700 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 701 if (size > 1) { 702 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 703 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 704 for (i=0;i<graph->ncc;i++) { 705 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 706 } 707 ierr = PetscFree(queue_global);CHKERRQ(ierr); 708 } 709 PetscFunctionReturn(0); 710 } 711 712 #undef __FUNCT__ 713 #define __FUNCT__ "PCBDDCGraphSetUp" 714 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 715 { 716 VecScatter scatter_ctx; 717 Vec local_vec,local_vec2,global_vec; 718 IS to,from; 719 MPI_Comm comm; 720 PetscScalar *array,*array2; 721 const PetscInt *is_indices; 722 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global,*subset_ref_node_global; 723 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 724 PetscErrorCode ierr; 725 PetscBool same_set,mirrors_found; 726 727 PetscFunctionBegin; 728 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 729 /* custom_minimal_size */ 730 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 731 /* get info l2gmap and allocate work vectors */ 732 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 733 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 734 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 735 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 736 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 737 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 738 ierr = VecSetSizes(global_vec,PETSC_DECIDE,graph->nvtxs_global);CHKERRQ(ierr); 739 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 740 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 741 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 742 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 743 744 /* get local periodic nodes */ 745 mirrors_found = PETSC_FALSE; 746 if (graph->nvtxs && n_neigh) { 747 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 748 for (i=0; i<n_shared[0]; i++) { 749 if (graph->count[shared[0][i]] > 1) { 750 mirrors_found = PETSC_TRUE; 751 break; 752 } 753 } 754 } 755 /* compute local mirrors (if any) */ 756 if (mirrors_found) { 757 PetscInt *local_indices,*global_indices; 758 /* get arrays of local and global indices */ 759 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 760 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 761 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 762 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 763 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 764 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 765 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 766 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 767 /* allocate space for mirrors */ 768 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 769 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 770 graph->mirrors_set[0] = 0; 771 772 k=0; 773 for (i=0;i<n_shared[0];i++) { 774 j=shared[0][i]; 775 if (graph->count[j] > 1) { 776 graph->mirrors[j]++; 777 k++; 778 } 779 } 780 /* allocate space for set of mirrors */ 781 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 782 for (i=1;i<graph->nvtxs;i++) 783 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 784 785 /* fill arrays */ 786 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 787 for (j=0;j<n_shared[0];j++) { 788 i=shared[0][j]; 789 if (graph->count[i] > 1) 790 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 791 } 792 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 793 for (i=0;i<graph->nvtxs;i++) { 794 if (graph->mirrors[i] > 0) { 795 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 796 j = global_indices[k]; 797 while ( k > 0 && global_indices[k-1] == j) k--; 798 for (j=0;j<graph->mirrors[i];j++) { 799 graph->mirrors_set[i][j]=local_indices[k+j]; 800 } 801 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 802 } 803 } 804 ierr = PetscFree(local_indices);CHKERRQ(ierr); 805 ierr = PetscFree(global_indices);CHKERRQ(ierr); 806 } 807 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 808 ierr = ISDestroy(&to);CHKERRQ(ierr); 809 ierr = ISDestroy(&from);CHKERRQ(ierr); 810 811 /* Count total number of neigh per node */ 812 k=0; 813 for (i=1;i<n_neigh;i++) { 814 k += n_shared[i]; 815 for (j=0;j<n_shared[i];j++) { 816 graph->count[shared[i][j]] += 1; 817 } 818 } 819 /* Allocate space for storing the set of neighbours for each node */ 820 if (graph->nvtxs) { 821 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 822 } 823 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 824 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 825 } 826 /* Get information for sharing subdomains */ 827 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 828 for (i=1;i<n_neigh;i++) { /* dont count myself */ 829 s = n_shared[i]; 830 for (j=0;j<s;j++) { 831 k = shared[i][j]; 832 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 833 graph->count[k] += 1; 834 } 835 } 836 /* sort set of sharing subdomains */ 837 for (i=0;i<graph->nvtxs;i++) { 838 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 839 } 840 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 841 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 842 843 /* 844 Get info for dofs splitting 845 User can specify only a subset; an additional field is considered as a complementary set 846 */ 847 for (i=0;i<graph->nvtxs;i++) { 848 graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 849 } 850 for (i=0;i<n_ISForDofs;i++) { 851 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 852 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 853 for (j=0;j<is_size;j++) { 854 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 855 graph->which_dof[is_indices[j]] = i; 856 } 857 } 858 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 859 } 860 if (neumann_is || dirichlet_is) { 861 /* Take into account Neumann nodes */ 862 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 863 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 864 if (neumann_is) { 865 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 866 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 867 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 868 for (i=0;i<is_size;i++) { 869 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 870 array[is_indices[i]] = 1.0; 871 } 872 } 873 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 874 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 875 } 876 /* Neumann nodes: impose consistency among neighbours */ 877 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 878 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 879 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 880 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 881 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 882 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 883 for (i=0;i<graph->nvtxs;i++) { 884 if (PetscRealPart(array[i]) > 0.1) { 885 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 886 } 887 } 888 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 889 /* Take into account Dirichlet nodes */ 890 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 891 if (dirichlet_is) { 892 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 893 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 894 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 895 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 896 for (i=0;i<is_size;i++){ 897 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 898 k = is_indices[i]; 899 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 900 if (PetscRealPart(array[k]) > 0.1) { 901 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); 902 } 903 array2[k] = 1.0; 904 } 905 } 906 } 907 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 908 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 909 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 910 } 911 /* Dirichlet nodes: impose consistency among neighbours */ 912 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 913 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 914 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 915 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 916 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 917 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 918 for (i=0;i<graph->nvtxs;i++) { 919 if (PetscRealPart(array[i]) > 0.1) { 920 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 921 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 922 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 923 } 924 } 925 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 926 } 927 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 928 if (graph->mirrors) { 929 for (i=0;i<graph->nvtxs;i++) 930 if (graph->mirrors[i]) 931 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 932 933 if (graph->xadj && graph->adjncy) { 934 PetscInt *new_xadj,*new_adjncy; 935 /* sort CSR graph */ 936 for (i=0;i<graph->nvtxs;i++) 937 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 938 939 /* adapt local CSR graph in case of local periodicity */ 940 k=0; 941 for (i=0;i<graph->nvtxs;i++) 942 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 943 k += graph->mirrors[graph->adjncy[j]]; 944 945 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 946 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 947 new_xadj[0]=0; 948 for (i=0;i<graph->nvtxs;i++) { 949 k = graph->xadj[i+1]-graph->xadj[i]; 950 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 951 new_xadj[i+1]=new_xadj[i]+k; 952 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 953 k = graph->mirrors[graph->adjncy[j]]; 954 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 955 new_xadj[i+1]+=k; 956 } 957 k = new_xadj[i+1]-new_xadj[i]; 958 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 959 new_xadj[i+1]=new_xadj[i]+k; 960 } 961 /* set new CSR into graph */ 962 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 963 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 964 graph->xadj = new_xadj; 965 graph->adjncy = new_adjncy; 966 } 967 } 968 969 /* mark special nodes -> each will become a single node equivalence class */ 970 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 971 if (custom_primal_vertices) { 972 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 973 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 974 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 975 for (i=0;i<is_size;i++){ 976 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 977 array[is_indices[i]] = 1.0; 978 } 979 } 980 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 981 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 982 } 983 /* special nodes: impose consistency among neighbours */ 984 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 985 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 986 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 987 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 988 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 989 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 990 j = 0; 991 for (i=0;i<graph->nvtxs;i++) { 992 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 993 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 994 j++; 995 } 996 } 997 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 998 999 /* mark interior nodes as touched and belonging to partition number 0 */ 1000 for (i=0;i<graph->nvtxs;i++) { 1001 if (!graph->count[i]) { 1002 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1003 graph->subset[i] = 0; 1004 } 1005 } 1006 /* init graph structure and compute default subsets */ 1007 nodes_touched=0; 1008 for (i=0;i<graph->nvtxs;i++) { 1009 if (PetscBTLookup(graph->touched,i)) { 1010 nodes_touched++; 1011 } 1012 } 1013 i = 0; 1014 graph->ncc = 0; 1015 total_counts = 0; 1016 while (nodes_touched<graph->nvtxs) { 1017 /* find first untouched node in local ordering */ 1018 while (PetscBTLookup(graph->touched,i)) { 1019 i++; 1020 } 1021 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1022 graph->subset[i] = graph->ncc+1; 1023 graph->cptr[graph->ncc] = total_counts; 1024 graph->queue[total_counts] = i; 1025 total_counts++; 1026 nodes_touched++; 1027 /* now find all other nodes having the same set of sharing subdomains */ 1028 for (j=i+1;j<graph->nvtxs;j++) { 1029 /* check for same number of sharing subdomains, dof number and same special mark */ 1030 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]) { 1031 /* check for same set of sharing subdomains */ 1032 same_set=PETSC_TRUE; 1033 for (k=0;k<graph->count[j];k++){ 1034 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 1035 same_set=PETSC_FALSE; 1036 } 1037 } 1038 /* I found a friend of mine */ 1039 if (same_set) { 1040 graph->subset[j]=graph->ncc+1; 1041 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1042 nodes_touched++; 1043 graph->queue[total_counts] = j; 1044 total_counts++; 1045 } 1046 } 1047 } 1048 graph->ncc++; 1049 } 1050 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 1051 graph->n_subsets = graph->ncc; 1052 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1053 for (i=0;i<graph->n_subsets;i++) { 1054 graph->subset_ncc[i] = 1; 1055 } 1056 /* final pointer */ 1057 graph->cptr[graph->ncc] = total_counts; 1058 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 1059 /* get a reference node (min index in global ordering) for each subset */ 1060 ierr = PetscMalloc1(graph->ncc,&subset_ref_node_global);CHKERRQ(ierr); 1061 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1062 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1063 for (i=0;i<graph->ncc;i++) { 1064 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 1065 subset_ref_node_global[i] = graph->queue[graph->cptr[i]]; 1066 } 1067 /* renumber reference nodes */ 1068 ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->l2gmap,graph->ncc,subset_ref_node_global,NULL,&k,&graph->subset_ref_node);CHKERRQ(ierr); 1069 ierr = PetscFree(subset_ref_node_global);CHKERRQ(ierr); 1070 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1071 /* free objects */ 1072 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1073 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1074 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1075 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1076 PetscFunctionReturn(0); 1077 } 1078 1079 #undef __FUNCT__ 1080 #define __FUNCT__ "PCBDDCGraphResetCSR" 1081 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1082 { 1083 PetscErrorCode ierr; 1084 1085 PetscFunctionBegin; 1086 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1087 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1088 graph->nvtxs_csr = 0; 1089 PetscFunctionReturn(0); 1090 } 1091 1092 #undef __FUNCT__ 1093 #define __FUNCT__ "PCBDDCGraphReset" 1094 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1095 { 1096 PetscErrorCode ierr; 1097 1098 PetscFunctionBegin; 1099 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1100 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1101 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1102 if (graph->nvtxs) { 1103 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1104 } 1105 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1106 ierr = PetscFree7(graph->count, 1107 graph->neighbours_set, 1108 graph->subset, 1109 graph->which_dof, 1110 graph->cptr, 1111 graph->queue, 1112 graph->special_dof);CHKERRQ(ierr); 1113 if (graph->mirrors) { 1114 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1115 } 1116 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1117 graph->nvtxs = 0; 1118 graph->n_subsets = 0; 1119 graph->custom_minimal_size = 1; 1120 PetscFunctionReturn(0); 1121 } 1122 1123 #undef __FUNCT__ 1124 #define __FUNCT__ "PCBDDCGraphInit" 1125 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap, PetscInt N) 1126 { 1127 PetscInt n; 1128 PetscErrorCode ierr; 1129 1130 PetscFunctionBegin; 1131 PetscValidPointer(graph,1); 1132 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1133 /* raise an error if already allocated */ 1134 if (graph->nvtxs) { 1135 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1136 } 1137 /* set number of vertices */ 1138 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1139 graph->l2gmap = l2gmap; 1140 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1141 graph->nvtxs = n; 1142 graph->nvtxs_global = N; 1143 /* allocate used space */ 1144 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1145 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1146 graph->nvtxs,&graph->neighbours_set, 1147 graph->nvtxs,&graph->subset, 1148 graph->nvtxs,&graph->which_dof, 1149 graph->nvtxs+1,&graph->cptr, 1150 graph->nvtxs,&graph->queue, 1151 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1152 /* zeroes memory */ 1153 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1154 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1155 /* use -1 as a default value for which_dof array */ 1156 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1157 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1158 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1159 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1160 /* zeroes first pointer to neighbour set */ 1161 if (graph->nvtxs) { 1162 graph->neighbours_set[0] = 0; 1163 } 1164 /* zeroes workspace for values of ncc */ 1165 graph->subset_ncc = 0; 1166 graph->subset_ref_node = 0; 1167 PetscFunctionReturn(0); 1168 } 1169 1170 #undef __FUNCT__ 1171 #define __FUNCT__ "PCBDDCGraphDestroy" 1172 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1173 { 1174 PetscErrorCode ierr; 1175 1176 PetscFunctionBegin; 1177 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1178 ierr = PetscFree(*graph);CHKERRQ(ierr); 1179 PetscFunctionReturn(0); 1180 } 1181 1182 #undef __FUNCT__ 1183 #define __FUNCT__ "PCBDDCGraphCreate" 1184 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1185 { 1186 PCBDDCGraph new_graph; 1187 PetscErrorCode ierr; 1188 1189 PetscFunctionBegin; 1190 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1191 /* local to global mapping of dofs */ 1192 new_graph->l2gmap = 0; 1193 /* vertex size */ 1194 new_graph->nvtxs = 0; 1195 new_graph->n_subsets = 0; 1196 new_graph->custom_minimal_size = 1; 1197 /* zeroes ponters */ 1198 new_graph->mirrors = 0; 1199 new_graph->mirrors_set = 0; 1200 new_graph->neighbours_set = 0; 1201 new_graph->subset = 0; 1202 new_graph->which_dof = 0; 1203 new_graph->special_dof = 0; 1204 new_graph->cptr = 0; 1205 new_graph->queue = 0; 1206 new_graph->count = 0; 1207 new_graph->subset_ncc = 0; 1208 new_graph->subset_ref_node = 0; 1209 new_graph->touched = 0; 1210 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1211 new_graph->nvtxs_csr = 0; 1212 new_graph->xadj = 0; 1213 new_graph->adjncy = 0; 1214 *graph = new_graph; 1215 PetscFunctionReturn(0); 1216 } 1217