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__ "PCBDDCGraphGetDirichletDofs" 7 PetscErrorCode PCBDDCGraphGetDirichletDofs(PCBDDCGraph graph,IS* dirdofs) 8 { 9 PetscInt i,size,sizer; 10 PetscErrorCode ierr; 11 12 PetscFunctionBegin; 13 size = 0; 14 for (i=0;i<graph->nvtxs;i++) { 15 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 16 } 17 ierr = MPI_Allreduce(&size,&sizer,1,MPIU_INT,MPI_LOR,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 18 if (sizer) { 19 PetscInt *dirdofs_idxs; 20 21 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 22 size = 0; 23 for (i=0;i<graph->nvtxs;i++) { 24 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 25 } 26 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)graph->l2gmap),size,dirdofs_idxs,PETSC_OWN_POINTER,dirdofs);CHKERRQ(ierr); 27 } else { 28 *dirdofs = NULL; 29 } 30 PetscFunctionReturn(0); 31 } 32 33 #undef __FUNCT__ 34 #define __FUNCT__ "PCBDDCGraphASCIIView" 35 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 36 { 37 PetscInt i,j,tabs; 38 PetscInt* queue_in_global_numbering; 39 PetscErrorCode ierr; 40 41 PetscFunctionBegin; 42 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 43 ierr = PetscViewerASCIIGetTab(viewer,&tabs);CHKERRQ(ierr); 44 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 46 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 47 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 48 if (verbosity_level > 1) { 49 for (i=0;i<graph->nvtxs;i++) { 50 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 51 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 52 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 53 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 54 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 55 if (graph->count[i]) { 56 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 57 for (j=0;j<graph->count[i];j++) { 58 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 59 } 60 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 61 } 62 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 63 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 64 if (graph->mirrors) { 65 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 66 if (graph->mirrors[i]) { 67 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 68 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 69 for (j=0;j<graph->mirrors[i];j++) { 70 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 71 } 72 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 73 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 74 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 75 } 76 } 77 if (verbosity_level > 2) { 78 if (graph->xadj && graph->adjncy) { 79 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 80 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 81 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 82 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 83 } 84 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 85 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 86 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 87 } 88 } 89 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 90 if (graph->subset[i] && graph->subset_ncc) { 91 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 92 } 93 } 94 } 95 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 96 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 97 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 98 for (i=0;i<graph->ncc;i++) { 99 PetscInt node_num=graph->queue[graph->cptr[i]]; 100 PetscBool printcc = PETSC_FALSE; 101 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 102 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 103 for (j=0;j<graph->count[node_num];j++) { 104 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 105 } 106 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 107 if (verbosity_level > 1) { 108 printcc = PETSC_TRUE; 109 } else if (graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 110 printcc = PETSC_TRUE; 111 } 112 if (printcc) { 113 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 114 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 115 } 116 } 117 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 118 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 119 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 120 } 121 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 122 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 123 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 124 } 125 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 126 PetscFunctionReturn(0); 127 } 128 129 #undef __FUNCT__ 130 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 131 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 132 { 133 IS *ISForFaces,*ISForEdges,ISForVertices; 134 PetscInt i,j,nfc,nec,nvc,*idx; 135 PetscBool twodim_flag; 136 PetscErrorCode ierr; 137 138 PetscFunctionBegin; 139 /* loop on ccs to evalute number of faces, edges and vertices */ 140 ISForFaces = NULL; 141 ISForEdges = NULL; 142 ISForVertices = NULL; 143 nfc = 0; 144 nec = 0; 145 nvc = 0; 146 twodim_flag = PETSC_FALSE; 147 for (i=0;i<graph->ncc;i++) { 148 PetscInt repdof = graph->queue[graph->cptr[i]]; 149 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 150 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 151 nfc++; 152 } else { /* note that nec will be zero in 2d */ 153 nec++; 154 } 155 } else { 156 nvc += graph->cptr[i+1]-graph->cptr[i]; 157 } 158 } 159 j=0; 160 ierr = MPI_Allreduce(&nec,&j,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 161 if (!j) { /* we are in a 2D case -> no faces, only edges */ 162 nec = nfc; 163 nfc = 0; 164 twodim_flag = PETSC_TRUE; 165 } 166 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 167 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 168 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 169 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 170 /* loop on ccs to compute index sets for faces and edges */ 171 nfc = 0; 172 nec = 0; 173 for (i=0;i<graph->ncc;i++) { 174 PetscInt repdof = graph->queue[graph->cptr[i]]; 175 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 176 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 177 if (twodim_flag) { 178 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 179 nec++; 180 } else { 181 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 182 nfc++; 183 } 184 } else { 185 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 186 nec++; 187 } 188 } 189 } 190 /* index set for vertices */ 191 if (nvc) { 192 nvc = 0; 193 for (i=0;i<graph->ncc;i++) { 194 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 195 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 196 idx[nvc]=graph->queue[j]; 197 nvc++; 198 } 199 } 200 } 201 /* sort vertex set (by local ordering) */ 202 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 203 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 204 } 205 /* get back info */ 206 if (n_faces) *n_faces = nfc; 207 if (FacesIS) { 208 *FacesIS = ISForFaces; 209 } else { 210 for (i=0;i<nfc;i++) { 211 ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 212 } 213 ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 214 } 215 if (n_edges) *n_edges = nec; 216 if (EdgesIS) { 217 *EdgesIS = ISForEdges; 218 } else { 219 for (i=0;i<nec;i++) { 220 ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 221 } 222 ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 223 } 224 if (VerticesIS) { 225 *VerticesIS = ISForVertices; 226 } else { 227 ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 228 } 229 PetscFunctionReturn(0); 230 } 231 232 #undef __FUNCT__ 233 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 234 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 235 { 236 PetscBool adapt_interface_reduced; 237 MPI_Comm interface_comm; 238 PetscMPIInt size; 239 PetscErrorCode ierr; 240 241 PetscFunctionBegin; 242 /* compute connected components locally */ 243 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 244 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 245 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 246 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 247 adapt_interface_reduced = PETSC_FALSE; 248 if (size > 1) { 249 PetscInt i; 250 PetscBool adapt_interface = PETSC_FALSE; 251 for (i=0;i<graph->n_subsets;i++) { 252 /* We are not sure that on a given subset of the local interface, 253 with two connected components, the latters be the same among sharing subdomains */ 254 if (graph->subset_ncc[i] > 1) { 255 adapt_interface = PETSC_TRUE; 256 break; 257 } 258 } 259 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 260 } 261 262 if (graph->n_subsets && adapt_interface_reduced) { 263 MPI_Request *send_requests; 264 MPI_Request *recv_requests; 265 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 266 PetscInt *old_xadj,*old_adjncy; 267 PetscInt i,j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 268 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 269 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 270 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 271 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 272 PetscBool *subset_cc_adapt,same_set; 273 274 /* Retrict adjacency graph using information from previously computed connected components */ 275 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 276 for (i=0;i<graph->nvtxs;i++) { 277 aux_new_xadj[i]=1; 278 } 279 for (i=0;i<graph->ncc;i++) { 280 k = graph->cptr[i+1]-graph->cptr[i]; 281 for (j=0;j<k;j++) { 282 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 283 } 284 } 285 j = 0; 286 for (i=0;i<graph->nvtxs;i++) { 287 j += aux_new_xadj[i]; 288 } 289 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 290 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 291 new_xadj[0]=0; 292 for (i=0;i<graph->nvtxs;i++) { 293 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 294 if (aux_new_xadj[i]==1) { 295 new_adjncy[new_xadj[i]]=i; 296 } 297 } 298 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 299 for (i=0;i<graph->ncc;i++) { 300 k = graph->cptr[i+1]-graph->cptr[i]; 301 for (j=0;j<k;j++) { 302 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 303 } 304 } 305 /* set temporarly new CSR into graph */ 306 old_xadj = graph->xadj; 307 old_adjncy = graph->adjncy; 308 graph->xadj = new_xadj; 309 graph->adjncy = new_adjncy; 310 /* allocate some space */ 311 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 312 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 313 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 314 ierr = PetscMalloc1(graph->n_subsets,&subset_to_nodes_indices);CHKERRQ(ierr); 315 /* first count how many neighbours per connected component I will receive from */ 316 cum_recv_counts[0]=0; 317 for (i=1;i<graph->n_subsets+1;i++) { 318 j = 0; 319 while (graph->subset[j] != i) { 320 j++; 321 } 322 subset_to_nodes_indices[i-1]=j; 323 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 324 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 325 } 326 ierr = PetscMalloc1(2*cum_recv_counts[graph->n_subsets],&recv_buffer_subset);CHKERRQ(ierr); 327 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&send_requests);CHKERRQ(ierr); 328 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 329 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 330 send_requests[i]=MPI_REQUEST_NULL; 331 recv_requests[i]=MPI_REQUEST_NULL; 332 } 333 /* exchange with my neighbours the number of my connected components on the shared interface */ 334 sum_requests = 0; 335 for (i=0;i<graph->n_subsets;i++) { 336 j = subset_to_nodes_indices[i]; 337 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 338 for (k=0;k<graph->count[j];k++) { 339 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 340 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 341 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 342 sum_requests++; 343 } 344 } 345 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 346 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 347 /* determine the connected component I need to adapt */ 348 ierr = PetscMalloc1(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 349 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 350 for (i=0;i<graph->n_subsets;i++) { 351 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 352 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 353 if (graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1) { 354 subset_cc_adapt[i] = PETSC_TRUE; 355 break; 356 } 357 } 358 } 359 buffer_size = 0; 360 for (i=0;i<graph->n_subsets;i++) { 361 if (subset_cc_adapt[i]) { 362 for (j=i;j<graph->ncc;j++) { 363 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 364 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 365 } 366 } 367 } 368 } 369 ierr = PetscMalloc1(buffer_size,&send_buffer);CHKERRQ(ierr); 370 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 371 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 372 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 373 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 374 ierr = PetscMalloc1(graph->n_subsets,&sizes_of_sends);CHKERRQ(ierr); 375 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 376 sum_requests = 0; 377 start_of_send = 0; 378 start_of_recv = cum_recv_counts[graph->n_subsets]; 379 for (i=0;i<graph->n_subsets;i++) { 380 if (subset_cc_adapt[i]) { 381 size_of_send = 0; 382 for (j=i;j<graph->ncc;j++) { 383 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 384 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 385 size_of_send += 1; 386 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 387 &queue_in_global_numbering[graph->cptr[j]], 388 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 389 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 390 } 391 } 392 j = subset_to_nodes_indices[i]; 393 sizes_of_sends[i] = size_of_send; 394 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 395 for (k=0;k<graph->count[j];k++) { 396 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 397 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 398 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 399 sum_requests++; 400 } 401 start_of_send += size_of_send; 402 } 403 } 404 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 405 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 406 buffer_size = 0; 407 for (k=0;k<sum_requests;k++) { 408 buffer_size += recv_buffer_subset[start_of_recv+k]; 409 } 410 ierr = PetscMalloc1(buffer_size,&recv_buffer);CHKERRQ(ierr); 411 /* now exchange the data */ 412 start_of_recv = 0; 413 start_of_send = 0; 414 sum_requests = 0; 415 for (i=0;i<graph->n_subsets;i++) { 416 if (subset_cc_adapt[i]) { 417 size_of_send = sizes_of_sends[i]; 418 j = subset_to_nodes_indices[i]; 419 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+2,&tag);CHKERRQ(ierr); 420 for (k=0;k<graph->count[j];k++) { 421 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 422 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 423 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 424 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 425 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 426 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 427 start_of_recv += size_of_recv; 428 sum_requests++; 429 } 430 start_of_send += size_of_send; 431 } 432 } 433 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 434 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 435 for (j=0;j<buffer_size;) { 436 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_MASK,recv_buffer[j],&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 437 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 438 if (graph->mirrors) { 439 PetscBool mirrored_found=PETSC_FALSE; 440 for (k=0;k<recv_buffer[j];k++) { 441 if (graph->mirrors[recv_buffer[j+k+1]]) { 442 mirrored_found=PETSC_TRUE; 443 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 444 } 445 } 446 if (mirrored_found) { 447 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 448 k=0; 449 while (k<recv_buffer[j]) { 450 for (s=1;s<graph->mirrors[recv_buffer[j+1+k]];s++) { 451 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 452 } 453 k+=graph->mirrors[recv_buffer[j+1+k]]+s; 454 } 455 } 456 } 457 k = recv_buffer[j]+1; 458 j += k; 459 } 460 sum_requests = cum_recv_counts[graph->n_subsets]; 461 start_of_recv = 0; 462 ierr = PetscMalloc1(graph->nvtxs,&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 463 global_subset_counter = 0; 464 for (i=0;i<graph->n_subsets;i++) { 465 if (subset_cc_adapt[i]) { 466 temp_buffer_size = 0; 467 /* find nodes on the shared interface we need to adapt */ 468 for (j=0;j<graph->nvtxs;j++) { 469 if (graph->subset[j]==i+1) { 470 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 471 temp_buffer_size++; 472 } else { 473 nodes_to_temp_buffer_indices[j] = -1; 474 } 475 } 476 /* allocate some temporary space */ 477 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 478 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 479 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 480 for (j=1;j<temp_buffer_size;j++) { 481 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 482 } 483 /* analyze contributions from neighbouring subdomains for i-th conn comp 484 temp buffer structure: 485 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 486 3 neighs procs with structured connected components: 487 neigh 0: [0 1 4], [2 3]; (2 connected components) 488 neigh 1: [0 1], [2 3 4]; (2 connected components) 489 neigh 2: [0 4], [1], [2 3]; (3 connected components) 490 tempbuffer (row-oriented) will be filled as: 491 [ 0, 0, 0; 492 0, 0, 1; 493 1, 1, 2; 494 1, 1, 2; 495 0, 1, 0; ]; 496 This way we can simply find intersections of ccs among neighs. 497 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 498 */ 499 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 500 ins_val = 0; 501 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 502 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 503 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 504 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 505 } 506 buffer_size += k; 507 ins_val++; 508 } 509 start_of_recv += size_of_recv; 510 sum_requests++; 511 } 512 ierr = PetscMalloc1(temp_buffer_size,&add_to_subset);CHKERRQ(ierr); 513 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 514 for (j=0;j<temp_buffer_size;j++) { 515 if (!add_to_subset[j]) { /* found a new cc */ 516 global_subset_counter++; 517 add_to_subset[j] = global_subset_counter; 518 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 519 same_set = PETSC_TRUE; 520 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 521 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 522 same_set = PETSC_FALSE; 523 break; 524 } 525 } 526 if (same_set) { 527 add_to_subset[k] = global_subset_counter; 528 } 529 } 530 } 531 } 532 /* insert new data in subset array */ 533 temp_buffer_size = 0; 534 for (j=0;j<graph->nvtxs;j++) { 535 if (graph->subset[j]==i+1) { 536 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 537 temp_buffer_size++; 538 } 539 } 540 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 541 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 542 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 543 } 544 } 545 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 546 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 547 ierr = PetscFree(send_requests);CHKERRQ(ierr); 548 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 549 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 550 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 551 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 552 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 553 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 554 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 555 /* We are ready to find for connected components consistent among neighbouring subdomains */ 556 if (global_subset_counter) { 557 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 558 global_subset_counter = 0; 559 for (i=0;i<graph->nvtxs;i++) { 560 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 561 global_subset_counter++; 562 for (j=i+1;j<graph->nvtxs;j++) { 563 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 564 graph->subset[j] = global_subset_counter; 565 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 566 } 567 } 568 graph->subset[i] = global_subset_counter; 569 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 570 } 571 } 572 /* refine connected components locally */ 573 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 574 } 575 /* restore original CSR graph of dofs */ 576 ierr = PetscFree(new_xadj);CHKERRQ(ierr); 577 ierr = PetscFree(new_adjncy);CHKERRQ(ierr); 578 graph->xadj = old_xadj; 579 graph->adjncy = old_adjncy; 580 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 581 } 582 PetscFunctionReturn(0); 583 } 584 585 /* The following code has been adapted from function IsConnectedSubdomain contained 586 in source file contig.c of METIS library (version 5.0.1) 587 It finds connected components for each subset */ 588 #undef __FUNCT__ 589 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 590 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 591 { 592 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 593 PetscInt *queue_global; 594 PetscMPIInt size; 595 PetscErrorCode ierr; 596 597 PetscFunctionBegin; 598 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&size);CHKERRQ(ierr); 599 /* quiet return if no csr info is available */ 600 if (!graph->xadj || !graph->adjncy) { 601 PetscFunctionReturn(0); 602 } 603 604 /* reset any previous search of connected components */ 605 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 606 graph->n_subsets = 0; 607 if (size == 1) { 608 if (graph->nvtxs) { 609 graph->n_subsets = 1; 610 for (i=0;i<graph->nvtxs;i++) { 611 graph->subset[i] = 1; 612 } 613 } 614 } else { 615 for (i=0;i<graph->nvtxs;i++) { 616 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 617 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 618 graph->subset[i] = 0; 619 } 620 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 621 } 622 } 623 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 624 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 625 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 626 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 627 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 628 629 /* begin search for connected components */ 630 cum_queue = 0; 631 ncc = 0; 632 for (n=0;n<graph->n_subsets;n++) { 633 pid = n+1; /* partition labeled by 0 is discarded */ 634 nleft = 0; 635 for (i=0;i<graph->nvtxs;i++) { 636 if (graph->subset[i] == pid) { 637 nleft++; 638 } 639 } 640 for (i=0; i<graph->nvtxs; i++) { 641 if (graph->subset[i] == pid) { 642 break; 643 } 644 } 645 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 646 graph->queue[cum_queue] = i; 647 first = 0; 648 last = 1; 649 graph->cptr[ncc] = cum_queue; 650 ncc_pid = 0; 651 while (first != nleft) { 652 if (first == last) { 653 graph->cptr[++ncc] = first+cum_queue; 654 ncc_pid++; 655 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 656 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 657 break; 658 } 659 } 660 graph->queue[cum_queue+last] = i; 661 last++; 662 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 663 } 664 i = graph->queue[cum_queue+first]; 665 first++; 666 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 667 k = graph->adjncy[j]; 668 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 669 graph->queue[cum_queue+last] = k; 670 last++; 671 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 672 } 673 } 674 } 675 graph->cptr[++ncc] = first+cum_queue; 676 ncc_pid++; 677 cum_queue = graph->cptr[ncc]; 678 graph->subset_ncc[n] = ncc_pid; 679 } 680 graph->ncc = ncc; 681 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 682 if (size > 1) { 683 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 684 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 685 for (i=0;i<graph->ncc;i++) { 686 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 687 } 688 ierr = PetscFree(queue_global);CHKERRQ(ierr); 689 } 690 PetscFunctionReturn(0); 691 } 692 693 #undef __FUNCT__ 694 #define __FUNCT__ "PCBDDCGraphSetUp" 695 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 696 { 697 VecScatter scatter_ctx; 698 Vec local_vec,local_vec2,global_vec; 699 IS to,from; 700 MPI_Comm comm; 701 PetscScalar *array,*array2; 702 const PetscInt *is_indices; 703 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global,*subset_ref_node_global; 704 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 705 PetscErrorCode ierr; 706 PetscBool same_set,mirrors_found; 707 708 PetscFunctionBegin; 709 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 710 /* custom_minimal_size */ 711 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 712 /* get info l2gmap and allocate work vectors */ 713 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 714 ierr = ISLocalToGlobalMappingGetIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 715 j = 0; 716 for (i=0;i<graph->nvtxs;i++) { 717 j = PetscMax(j,is_indices[i]); 718 } 719 ierr = MPI_Allreduce(&j,&i,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 720 i++; 721 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 722 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 723 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 724 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 725 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 726 ierr = VecSetSizes(global_vec,PETSC_DECIDE,i);CHKERRQ(ierr); 727 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 728 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 729 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 730 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 731 732 /* get local periodic nodes */ 733 mirrors_found = PETSC_FALSE; 734 if (graph->nvtxs && n_neigh) { 735 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 736 for (i=0; i<n_shared[0]; i++) { 737 if (graph->count[shared[0][i]] > 1) { 738 mirrors_found = PETSC_TRUE; 739 break; 740 } 741 } 742 } 743 /* compute local mirrors (if any) */ 744 if (mirrors_found) { 745 PetscInt *local_indices,*global_indices; 746 /* get arrays of local and global indices */ 747 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 748 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 749 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 750 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 751 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 752 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 753 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 754 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 755 /* allocate space for mirrors */ 756 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 757 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 758 graph->mirrors_set[0] = 0; 759 760 k=0; 761 for (i=0;i<n_shared[0];i++) { 762 j=shared[0][i]; 763 if (graph->count[j] > 1) { 764 graph->mirrors[j]++; 765 k++; 766 } 767 } 768 /* allocate space for set of mirrors */ 769 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 770 for (i=1;i<graph->nvtxs;i++) 771 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 772 773 /* fill arrays */ 774 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 775 for (j=0;j<n_shared[0];j++) { 776 i=shared[0][j]; 777 if (graph->count[i] > 1) 778 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 779 } 780 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 781 for (i=0;i<graph->nvtxs;i++) { 782 if (graph->mirrors[i] > 0) { 783 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 784 j = global_indices[k]; 785 while ( k > 0 && global_indices[k-1] == j) k--; 786 for (j=0;j<graph->mirrors[i];j++) { 787 graph->mirrors_set[i][j]=local_indices[k+j]; 788 } 789 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 790 } 791 } 792 ierr = PetscFree(local_indices);CHKERRQ(ierr); 793 ierr = PetscFree(global_indices);CHKERRQ(ierr); 794 } 795 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 796 ierr = ISDestroy(&to);CHKERRQ(ierr); 797 ierr = ISDestroy(&from);CHKERRQ(ierr); 798 799 /* Count total number of neigh per node */ 800 k=0; 801 for (i=1;i<n_neigh;i++) { 802 k += n_shared[i]; 803 for (j=0;j<n_shared[i];j++) { 804 graph->count[shared[i][j]] += 1; 805 } 806 } 807 /* Allocate space for storing the set of neighbours for each node */ 808 if (graph->nvtxs) { 809 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 810 } 811 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 812 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 813 } 814 /* Get information for sharing subdomains */ 815 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 816 for (i=1;i<n_neigh;i++) { /* dont count myself */ 817 s = n_shared[i]; 818 for (j=0;j<s;j++) { 819 k = shared[i][j]; 820 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 821 graph->count[k] += 1; 822 } 823 } 824 /* sort set of sharing subdomains */ 825 for (i=0;i<graph->nvtxs;i++) { 826 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 827 } 828 /* 829 Get info for dofs splitting 830 User can specify only a subset; an additional field is considered as a complementary set 831 */ 832 for (i=0;i<graph->nvtxs;i++) { 833 graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 834 } 835 for (i=0;i<n_ISForDofs;i++) { 836 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 837 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 838 for (j=0;j<is_size;j++) { 839 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 840 graph->which_dof[is_indices[j]] = i; 841 } 842 } 843 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 844 } 845 /* Take into account Neumann nodes */ 846 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 847 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 848 if (neumann_is) { 849 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 850 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 851 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 852 for (i=0;i<is_size;i++) { 853 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 854 array[is_indices[i]] = 1.0; 855 } 856 } 857 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 858 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 859 } 860 /* Neumann nodes: impose consistency among neighbours */ 861 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 862 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 863 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 864 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 865 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 866 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 867 for (i=0;i<graph->nvtxs;i++) { 868 if (PetscRealPart(array[i]) > 0.1) { 869 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 870 } 871 } 872 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 873 /* Take into account Dirichlet nodes */ 874 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 875 if (dirichlet_is) { 876 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 877 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 878 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 879 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 880 for (i=0;i<is_size;i++){ 881 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 882 k = is_indices[i]; 883 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 884 if (PetscRealPart(array[k]) > 0.1) { 885 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); 886 } 887 array2[k] = 1.0; 888 } 889 } 890 } 891 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 892 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 893 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 894 } 895 /* Dirichlet nodes: impose consistency among neighbours */ 896 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 897 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 898 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 899 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 900 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 901 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 902 for (i=0;i<graph->nvtxs;i++) { 903 if (PetscRealPart(array[i]) > 0.1) { 904 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 905 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 906 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 907 } 908 } 909 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 910 911 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 912 if (graph->mirrors) { 913 for (i=0;i<graph->nvtxs;i++) 914 if (graph->mirrors[i]) 915 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 916 917 if (graph->xadj && graph->adjncy) { 918 PetscInt *new_xadj,*new_adjncy; 919 /* sort CSR graph */ 920 for (i=0;i<graph->nvtxs;i++) 921 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 922 923 /* adapt local CSR graph in case of local periodicity */ 924 k=0; 925 for (i=0;i<graph->nvtxs;i++) 926 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 927 k += graph->mirrors[graph->adjncy[j]]; 928 929 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 930 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 931 new_xadj[0]=0; 932 for (i=0;i<graph->nvtxs;i++) { 933 k = graph->xadj[i+1]-graph->xadj[i]; 934 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 935 new_xadj[i+1]=new_xadj[i]+k; 936 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 937 k = graph->mirrors[graph->adjncy[j]]; 938 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 939 new_xadj[i+1]+=k; 940 } 941 k = new_xadj[i+1]-new_xadj[i]; 942 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 943 new_xadj[i+1]=new_xadj[i]+k; 944 } 945 /* set new CSR into graph */ 946 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 947 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 948 graph->xadj = new_xadj; 949 graph->adjncy = new_adjncy; 950 } 951 } 952 953 /* mark special nodes -> each will become a single node equivalence class */ 954 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 955 if (custom_primal_vertices) { 956 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 957 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 958 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 959 for (i=0;i<is_size;i++){ 960 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 961 array[is_indices[i]] = 1.0; 962 } 963 } 964 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 965 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 966 } 967 /* special nodes: impose consistency among neighbours */ 968 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 969 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 970 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 971 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 972 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 973 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 974 j = 0; 975 for (i=0;i<graph->nvtxs;i++) { 976 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 977 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 978 j++; 979 } 980 } 981 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 982 983 /* mark interior nodes as touched and belonging to partition number 0 */ 984 for (i=0;i<graph->nvtxs;i++) { 985 if (!graph->count[i]) { 986 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 987 graph->subset[i] = 0; 988 } 989 } 990 /* init graph structure and compute default subsets */ 991 nodes_touched=0; 992 for (i=0;i<graph->nvtxs;i++) { 993 if (PetscBTLookup(graph->touched,i)) { 994 nodes_touched++; 995 } 996 } 997 i = 0; 998 graph->ncc = 0; 999 total_counts = 0; 1000 while (nodes_touched<graph->nvtxs) { 1001 /* find first untouched node in local ordering */ 1002 while (PetscBTLookup(graph->touched,i)) { 1003 i++; 1004 } 1005 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1006 graph->subset[i] = graph->ncc+1; 1007 graph->cptr[graph->ncc] = total_counts; 1008 graph->queue[total_counts] = i; 1009 total_counts++; 1010 nodes_touched++; 1011 /* now find all other nodes having the same set of sharing subdomains */ 1012 for (j=i+1;j<graph->nvtxs;j++) { 1013 /* check for same number of sharing subdomains, dof number and same special mark */ 1014 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]) { 1015 /* check for same set of sharing subdomains */ 1016 same_set=PETSC_TRUE; 1017 for (k=0;k<graph->count[j];k++){ 1018 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 1019 same_set=PETSC_FALSE; 1020 } 1021 } 1022 /* I found a friend of mine */ 1023 if (same_set) { 1024 graph->subset[j]=graph->ncc+1; 1025 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1026 nodes_touched++; 1027 graph->queue[total_counts] = j; 1028 total_counts++; 1029 } 1030 } 1031 } 1032 graph->ncc++; 1033 } 1034 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 1035 graph->n_subsets = graph->ncc; 1036 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1037 for (i=0;i<graph->n_subsets;i++) { 1038 graph->subset_ncc[i] = 1; 1039 } 1040 /* final pointer */ 1041 graph->cptr[graph->ncc] = total_counts; 1042 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 1043 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 1044 /* get a reference node (min index in global ordering) for each subset */ 1045 ierr = PetscMalloc1(graph->ncc,&subset_ref_node_global);CHKERRQ(ierr); 1046 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1047 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1048 for (i=0;i<graph->ncc;i++) { 1049 PetscInt minval = queue_global[graph->cptr[i]]; 1050 PetscInt minloc = graph->queue[graph->cptr[i]]; 1051 for (j=graph->cptr[i]+1;j<graph->cptr[i+1];j++) { 1052 if (minval > queue_global[j]) { 1053 minval = queue_global[j]; 1054 minloc = graph->queue[j]; 1055 } 1056 } 1057 subset_ref_node_global[i] = minloc; 1058 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 1059 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 1060 } 1061 /* renumber reference nodes */ 1062 ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->l2gmap,graph->ncc,subset_ref_node_global,NULL,&k,&graph->subset_ref_node);CHKERRQ(ierr); 1063 ierr = PetscFree(subset_ref_node_global);CHKERRQ(ierr); 1064 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1065 /* free objects */ 1066 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1067 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1068 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1069 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1070 PetscFunctionReturn(0); 1071 } 1072 1073 #undef __FUNCT__ 1074 #define __FUNCT__ "PCBDDCGraphResetCSR" 1075 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1076 { 1077 PetscErrorCode ierr; 1078 1079 PetscFunctionBegin; 1080 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1081 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1082 graph->nvtxs_csr = 0; 1083 PetscFunctionReturn(0); 1084 } 1085 1086 #undef __FUNCT__ 1087 #define __FUNCT__ "PCBDDCGraphReset" 1088 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1089 { 1090 PetscErrorCode ierr; 1091 1092 PetscFunctionBegin; 1093 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1094 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1095 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1096 if (graph->nvtxs) { 1097 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1098 } 1099 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1100 ierr = PetscFree7(graph->count, 1101 graph->neighbours_set, 1102 graph->subset, 1103 graph->which_dof, 1104 graph->cptr, 1105 graph->queue, 1106 graph->special_dof);CHKERRQ(ierr); 1107 if (graph->mirrors) { 1108 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1109 } 1110 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1111 graph->nvtxs = 0; 1112 graph->n_subsets = 0; 1113 graph->custom_minimal_size = 1; 1114 PetscFunctionReturn(0); 1115 } 1116 1117 #undef __FUNCT__ 1118 #define __FUNCT__ "PCBDDCGraphInit" 1119 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap) 1120 { 1121 PetscInt n; 1122 PetscErrorCode ierr; 1123 1124 PetscFunctionBegin; 1125 PetscValidPointer(graph,1); 1126 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1127 /* raise an error if already allocated */ 1128 if (graph->nvtxs) { 1129 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1130 } 1131 /* set number of vertices */ 1132 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1133 graph->l2gmap = l2gmap; 1134 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1135 graph->nvtxs = n; 1136 /* allocate used space */ 1137 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1138 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1139 graph->nvtxs,&graph->neighbours_set, 1140 graph->nvtxs,&graph->subset, 1141 graph->nvtxs,&graph->which_dof, 1142 graph->nvtxs+1,&graph->cptr, 1143 graph->nvtxs,&graph->queue, 1144 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1145 /* zeroes memory */ 1146 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1147 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1148 /* use -1 as a default value for which_dof array */ 1149 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1150 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1151 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1152 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1153 /* zeroes first pointer to neighbour set */ 1154 if (graph->nvtxs) { 1155 graph->neighbours_set[0] = 0; 1156 } 1157 /* zeroes workspace for values of ncc */ 1158 graph->subset_ncc = 0; 1159 graph->subset_ref_node = 0; 1160 PetscFunctionReturn(0); 1161 } 1162 1163 #undef __FUNCT__ 1164 #define __FUNCT__ "PCBDDCGraphDestroy" 1165 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1166 { 1167 PetscErrorCode ierr; 1168 1169 PetscFunctionBegin; 1170 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1171 ierr = PetscFree(*graph);CHKERRQ(ierr); 1172 PetscFunctionReturn(0); 1173 } 1174 1175 #undef __FUNCT__ 1176 #define __FUNCT__ "PCBDDCGraphCreate" 1177 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1178 { 1179 PCBDDCGraph new_graph; 1180 PetscErrorCode ierr; 1181 1182 PetscFunctionBegin; 1183 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1184 /* local to global mapping of dofs */ 1185 new_graph->l2gmap = 0; 1186 /* vertex size */ 1187 new_graph->nvtxs = 0; 1188 new_graph->n_subsets = 0; 1189 new_graph->custom_minimal_size = 1; 1190 /* zeroes ponters */ 1191 new_graph->mirrors = 0; 1192 new_graph->mirrors_set = 0; 1193 new_graph->neighbours_set = 0; 1194 new_graph->subset = 0; 1195 new_graph->which_dof = 0; 1196 new_graph->special_dof = 0; 1197 new_graph->cptr = 0; 1198 new_graph->queue = 0; 1199 new_graph->count = 0; 1200 new_graph->subset_ncc = 0; 1201 new_graph->subset_ref_node = 0; 1202 new_graph->touched = 0; 1203 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1204 new_graph->nvtxs_csr = 0; 1205 new_graph->xadj = 0; 1206 new_graph->adjncy = 0; 1207 *graph = new_graph; 1208 PetscFunctionReturn(0); 1209 } 1210