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