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