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