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