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