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,nfc,nec,nvc,*idx; 135 PetscBool twodim,twodimr; 136 PetscErrorCode ierr; 137 138 PetscFunctionBegin; 139 /* loop on ccs to evalute number of faces, edges and vertices */ 140 nfc = 0; 141 nec = 0; 142 nvc = 0; 143 for (i=0;i<graph->ncc;i++) { 144 PetscInt repdof = graph->queue[graph->cptr[i]]; 145 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 146 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 147 nfc++; 148 } else { /* note that nec will be zero in 2d */ 149 nec++; 150 } 151 } else { 152 nvc += graph->cptr[i+1]-graph->cptr[i]; 153 } 154 } 155 156 /* determine if we are in 2D or 3D */ 157 twodim = (PetscBool)!nec; 158 twodimr = PETSC_FALSE; 159 ierr = MPI_Allreduce(&twodim,&twodimr,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 160 if (twodimr) { /* we are in a 2D case -> edge are shared by 2 subregions and faces don't exist */ 161 nec = nfc; 162 nfc = 0; 163 } 164 165 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 166 if (FacesIS) { 167 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 168 } 169 if (EdgesIS) { 170 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 171 } 172 if (VerticesIS) { 173 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 174 } 175 176 /* loop on ccs to compute index sets for faces and edges */ 177 nfc = 0; 178 nec = 0; 179 for (i=0;i<graph->ncc;i++) { 180 PetscInt repdof = graph->queue[graph->cptr[i]]; 181 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 182 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 183 if (twodimr) { 184 if (EdgesIS) { 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 } 187 nec++; 188 } else { 189 if (FacesIS) { 190 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 191 } 192 nfc++; 193 } 194 } else { 195 if (EdgesIS) { 196 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 197 } 198 nec++; 199 } 200 } 201 } 202 /* index set for vertices */ 203 if (VerticesIS) { 204 nvc = 0; 205 for (i=0;i<graph->ncc;i++) { 206 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 207 PetscInt j; 208 209 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 210 idx[nvc]=graph->queue[j]; 211 nvc++; 212 } 213 } 214 } 215 /* sort vertex set (by local ordering) */ 216 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 217 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 218 } 219 /* get back info */ 220 if (n_faces) *n_faces = nfc; 221 if (FacesIS) { 222 *FacesIS = ISForFaces; 223 } 224 if (n_edges) *n_edges = nec; 225 if (EdgesIS) { 226 *EdgesIS = ISForEdges; 227 } 228 if (VerticesIS) { 229 *VerticesIS = ISForVertices; 230 } 231 PetscFunctionReturn(0); 232 } 233 234 #undef __FUNCT__ 235 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 236 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 237 { 238 PetscBool adapt_interface_reduced; 239 MPI_Comm interface_comm; 240 PetscMPIInt size; 241 PetscErrorCode ierr; 242 243 PetscFunctionBegin; 244 /* compute connected components locally */ 245 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 246 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 247 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 248 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 249 adapt_interface_reduced = PETSC_FALSE; 250 if (size > 1) { 251 PetscInt i; 252 PetscBool adapt_interface = PETSC_FALSE; 253 for (i=0;i<graph->n_subsets;i++) { 254 /* We are not sure that on a given subset of the local interface, 255 with two connected components, the latters be the same among sharing subdomains */ 256 if (graph->subset_ncc[i] > 1) { 257 adapt_interface = PETSC_TRUE; 258 break; 259 } 260 } 261 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 262 } 263 264 if (graph->n_subsets && adapt_interface_reduced) { 265 MPI_Request *send_requests; 266 MPI_Request *recv_requests; 267 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 268 PetscInt *old_xadj,*old_adjncy; 269 PetscInt i,j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 270 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 271 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 272 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 273 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 274 PetscBool *subset_cc_adapt,same_set; 275 276 /* Retrict adjacency graph using information from previously computed connected components */ 277 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 278 for (i=0;i<graph->nvtxs;i++) { 279 aux_new_xadj[i]=1; 280 } 281 for (i=0;i<graph->ncc;i++) { 282 k = graph->cptr[i+1]-graph->cptr[i]; 283 for (j=0;j<k;j++) { 284 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 285 } 286 } 287 j = 0; 288 for (i=0;i<graph->nvtxs;i++) { 289 j += aux_new_xadj[i]; 290 } 291 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 292 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 293 new_xadj[0]=0; 294 for (i=0;i<graph->nvtxs;i++) { 295 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 296 if (aux_new_xadj[i]==1) { 297 new_adjncy[new_xadj[i]]=i; 298 } 299 } 300 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 301 for (i=0;i<graph->ncc;i++) { 302 k = graph->cptr[i+1]-graph->cptr[i]; 303 for (j=0;j<k;j++) { 304 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 305 } 306 } 307 /* set temporarly new CSR into graph */ 308 old_xadj = graph->xadj; 309 old_adjncy = graph->adjncy; 310 graph->xadj = new_xadj; 311 graph->adjncy = new_adjncy; 312 /* allocate some space */ 313 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 314 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 315 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 316 ierr = PetscMalloc1(graph->n_subsets,&subset_to_nodes_indices);CHKERRQ(ierr); 317 /* first count how many neighbours per connected component I will receive from */ 318 cum_recv_counts[0]=0; 319 for (i=1;i<graph->n_subsets+1;i++) { 320 j = 0; 321 while (graph->subset[j] != i) { 322 j++; 323 } 324 subset_to_nodes_indices[i-1]=j; 325 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 326 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 327 } 328 ierr = PetscMalloc1(2*cum_recv_counts[graph->n_subsets],&recv_buffer_subset);CHKERRQ(ierr); 329 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&send_requests);CHKERRQ(ierr); 330 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 331 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 332 send_requests[i]=MPI_REQUEST_NULL; 333 recv_requests[i]=MPI_REQUEST_NULL; 334 } 335 /* exchange with my neighbours the number of my connected components on the shared interface */ 336 sum_requests = 0; 337 for (i=0;i<graph->n_subsets;i++) { 338 j = subset_to_nodes_indices[i]; 339 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 340 for (k=0;k<graph->count[j];k++) { 341 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 342 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 343 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 344 sum_requests++; 345 } 346 } 347 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 348 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 349 /* determine the connected component I need to adapt */ 350 ierr = PetscMalloc1(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 351 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 352 for (i=0;i<graph->n_subsets;i++) { 353 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 354 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 355 if (graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1) { 356 subset_cc_adapt[i] = PETSC_TRUE; 357 break; 358 } 359 } 360 } 361 buffer_size = 0; 362 for (i=0;i<graph->n_subsets;i++) { 363 if (subset_cc_adapt[i]) { 364 for (j=i;j<graph->ncc;j++) { 365 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 366 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 367 } 368 } 369 } 370 } 371 ierr = PetscMalloc1(buffer_size,&send_buffer);CHKERRQ(ierr); 372 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 373 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 374 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 375 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 376 ierr = PetscMalloc1(graph->n_subsets,&sizes_of_sends);CHKERRQ(ierr); 377 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 378 sum_requests = 0; 379 start_of_send = 0; 380 start_of_recv = cum_recv_counts[graph->n_subsets]; 381 for (i=0;i<graph->n_subsets;i++) { 382 if (subset_cc_adapt[i]) { 383 size_of_send = 0; 384 for (j=i;j<graph->ncc;j++) { 385 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 386 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 387 size_of_send += 1; 388 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 389 &queue_in_global_numbering[graph->cptr[j]], 390 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 391 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 392 } 393 } 394 j = subset_to_nodes_indices[i]; 395 sizes_of_sends[i] = size_of_send; 396 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 397 for (k=0;k<graph->count[j];k++) { 398 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 399 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 400 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 401 sum_requests++; 402 } 403 start_of_send += size_of_send; 404 } 405 } 406 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 407 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 408 buffer_size = 0; 409 for (k=0;k<sum_requests;k++) { 410 buffer_size += recv_buffer_subset[start_of_recv+k]; 411 } 412 ierr = PetscMalloc1(buffer_size,&recv_buffer);CHKERRQ(ierr); 413 /* now exchange the data */ 414 start_of_recv = 0; 415 start_of_send = 0; 416 sum_requests = 0; 417 for (i=0;i<graph->n_subsets;i++) { 418 if (subset_cc_adapt[i]) { 419 size_of_send = sizes_of_sends[i]; 420 j = subset_to_nodes_indices[i]; 421 ierr = PetscMPIIntCast(3*graph->subset_ref_node[i]+2,&tag);CHKERRQ(ierr); 422 for (k=0;k<graph->count[j];k++) { 423 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 424 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 425 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 426 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 427 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 428 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 429 start_of_recv += size_of_recv; 430 sum_requests++; 431 } 432 start_of_send += size_of_send; 433 } 434 } 435 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 436 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 437 for (j=0;j<buffer_size;) { 438 PetscInt rbs = recv_buffer[j]; 439 440 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_DROP,rbs,&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 441 if (rbs != recv_buffer[j]) { 442 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Mapped received buffer size %d differs from input size %d\n",rbs,recv_buffer[j]); 443 } 444 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 445 if (graph->mirrors) { 446 PetscBool mirrored_found=PETSC_FALSE; 447 for (k=0;k<recv_buffer[j];k++) { 448 if (graph->mirrors[recv_buffer[j+k+1]]) { 449 mirrored_found=PETSC_TRUE; 450 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 451 } 452 } 453 if (mirrored_found) { 454 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 455 k=0; 456 while (k<recv_buffer[j]) { 457 PetscInt nmir = graph->mirrors[recv_buffer[j+1+k]]; 458 for (s=1;s<nmir;s++) { 459 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 460 } 461 k += nmir; 462 } 463 } 464 } 465 k = recv_buffer[j]+1; 466 j += k; 467 } 468 sum_requests = cum_recv_counts[graph->n_subsets]; 469 start_of_recv = 0; 470 ierr = PetscMalloc1(graph->nvtxs,&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 471 global_subset_counter = 0; 472 for (i=0;i<graph->n_subsets;i++) { 473 if (subset_cc_adapt[i]) { 474 temp_buffer_size = 0; 475 /* find nodes on the shared interface we need to adapt */ 476 for (j=0;j<graph->nvtxs;j++) { 477 if (graph->subset[j]==i+1) { 478 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 479 temp_buffer_size++; 480 } else { 481 nodes_to_temp_buffer_indices[j] = -1; 482 } 483 } 484 /* allocate some temporary space */ 485 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 486 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 487 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 488 for (j=1;j<temp_buffer_size;j++) { 489 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 490 } 491 /* analyze contributions from neighbouring subdomains for i-th conn comp 492 temp buffer structure: 493 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 494 3 neighs procs with structured connected components: 495 neigh 0: [0 1 4], [2 3]; (2 connected components) 496 neigh 1: [0 1], [2 3 4]; (2 connected components) 497 neigh 2: [0 4], [1], [2 3]; (3 connected components) 498 tempbuffer (row-oriented) will be filled as: 499 [ 0, 0, 0; 500 0, 0, 1; 501 1, 1, 2; 502 1, 1, 2; 503 0, 1, 0; ]; 504 This way we can simply find intersections of ccs among neighs. 505 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 506 */ 507 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 508 ins_val = 0; 509 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 510 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 511 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 512 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 513 } 514 buffer_size += k; 515 ins_val++; 516 } 517 start_of_recv += size_of_recv; 518 sum_requests++; 519 } 520 ierr = PetscMalloc1(temp_buffer_size,&add_to_subset);CHKERRQ(ierr); 521 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 522 for (j=0;j<temp_buffer_size;j++) { 523 if (!add_to_subset[j]) { /* found a new cc */ 524 global_subset_counter++; 525 add_to_subset[j] = global_subset_counter; 526 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 527 same_set = PETSC_TRUE; 528 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 529 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 530 same_set = PETSC_FALSE; 531 break; 532 } 533 } 534 if (same_set) { 535 add_to_subset[k] = global_subset_counter; 536 } 537 } 538 } 539 } 540 /* insert new data in subset array */ 541 temp_buffer_size = 0; 542 for (j=0;j<graph->nvtxs;j++) { 543 if (graph->subset[j]==i+1) { 544 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 545 temp_buffer_size++; 546 } 547 } 548 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 549 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 550 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 551 } 552 } 553 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 554 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 555 ierr = PetscFree(send_requests);CHKERRQ(ierr); 556 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 557 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 558 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 559 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 560 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 561 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 562 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 563 /* We are ready to find for connected components consistent among neighbouring subdomains */ 564 if (global_subset_counter) { 565 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 566 global_subset_counter = 0; 567 for (i=0;i<graph->nvtxs;i++) { 568 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 569 global_subset_counter++; 570 for (j=i+1;j<graph->nvtxs;j++) { 571 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 572 graph->subset[j] = global_subset_counter; 573 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 574 } 575 } 576 graph->subset[i] = global_subset_counter; 577 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 578 } 579 } 580 /* refine connected components locally */ 581 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 582 } 583 /* restore original CSR graph of dofs */ 584 ierr = PetscFree(new_xadj);CHKERRQ(ierr); 585 ierr = PetscFree(new_adjncy);CHKERRQ(ierr); 586 graph->xadj = old_xadj; 587 graph->adjncy = old_adjncy; 588 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 589 } 590 PetscFunctionReturn(0); 591 } 592 593 /* The following code has been adapted from function IsConnectedSubdomain contained 594 in source file contig.c of METIS library (version 5.0.1) 595 It finds connected components for each subset */ 596 #undef __FUNCT__ 597 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 598 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 599 { 600 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 601 PetscInt *queue_global; 602 PetscMPIInt size; 603 PetscErrorCode ierr; 604 605 PetscFunctionBegin; 606 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&size);CHKERRQ(ierr); 607 /* quiet return if no csr info is available */ 608 if (!graph->xadj || !graph->adjncy) { 609 PetscFunctionReturn(0); 610 } 611 612 /* reset any previous search of connected components */ 613 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 614 graph->n_subsets = 0; 615 if (size == 1) { 616 if (graph->nvtxs) { 617 graph->n_subsets = 1; 618 for (i=0;i<graph->nvtxs;i++) { 619 graph->subset[i] = 1; 620 } 621 } 622 } else { 623 for (i=0;i<graph->nvtxs;i++) { 624 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 625 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 626 graph->subset[i] = 0; 627 } 628 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 629 } 630 } 631 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 632 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 633 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 634 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 635 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 636 637 /* begin search for connected components */ 638 cum_queue = 0; 639 ncc = 0; 640 for (n=0;n<graph->n_subsets;n++) { 641 pid = n+1; /* partition labeled by 0 is discarded */ 642 nleft = 0; 643 for (i=0;i<graph->nvtxs;i++) { 644 if (graph->subset[i] == pid) { 645 nleft++; 646 } 647 } 648 for (i=0; i<graph->nvtxs; i++) { 649 if (graph->subset[i] == pid) { 650 break; 651 } 652 } 653 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 654 graph->queue[cum_queue] = i; 655 first = 0; 656 last = 1; 657 graph->cptr[ncc] = cum_queue; 658 ncc_pid = 0; 659 while (first != nleft) { 660 if (first == last) { 661 graph->cptr[++ncc] = first+cum_queue; 662 ncc_pid++; 663 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 664 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 665 break; 666 } 667 } 668 graph->queue[cum_queue+last] = i; 669 last++; 670 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 671 } 672 i = graph->queue[cum_queue+first]; 673 first++; 674 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 675 k = graph->adjncy[j]; 676 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 677 graph->queue[cum_queue+last] = k; 678 last++; 679 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 680 } 681 } 682 } 683 graph->cptr[++ncc] = first+cum_queue; 684 ncc_pid++; 685 cum_queue = graph->cptr[ncc]; 686 graph->subset_ncc[n] = ncc_pid; 687 } 688 graph->ncc = ncc; 689 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 690 if (size > 1) { 691 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 692 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 693 for (i=0;i<graph->ncc;i++) { 694 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 695 } 696 ierr = PetscFree(queue_global);CHKERRQ(ierr); 697 } 698 PetscFunctionReturn(0); 699 } 700 701 #undef __FUNCT__ 702 #define __FUNCT__ "PCBDDCGraphSetUp" 703 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 704 { 705 VecScatter scatter_ctx; 706 Vec local_vec,local_vec2,global_vec; 707 IS to,from; 708 MPI_Comm comm; 709 PetscScalar *array,*array2; 710 const PetscInt *is_indices; 711 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global,*subset_ref_node_global; 712 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 713 PetscErrorCode ierr; 714 PetscBool same_set,mirrors_found; 715 716 PetscFunctionBegin; 717 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 718 /* custom_minimal_size */ 719 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 720 /* get info l2gmap and allocate work vectors */ 721 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 722 ierr = ISLocalToGlobalMappingGetIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 723 j = 0; 724 for (i=0;i<graph->nvtxs;i++) { 725 j = PetscMax(j,is_indices[i]); 726 } 727 ierr = ISLocalToGlobalMappingRestoreIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 728 ierr = MPI_Allreduce(&j,&i,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 729 i++; 730 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 731 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 732 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 733 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 734 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 735 ierr = VecSetSizes(global_vec,PETSC_DECIDE,i);CHKERRQ(ierr); 736 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 737 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 738 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 739 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 740 741 /* get local periodic nodes */ 742 mirrors_found = PETSC_FALSE; 743 if (graph->nvtxs && n_neigh) { 744 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 745 for (i=0; i<n_shared[0]; i++) { 746 if (graph->count[shared[0][i]] > 1) { 747 mirrors_found = PETSC_TRUE; 748 break; 749 } 750 } 751 } 752 /* compute local mirrors (if any) */ 753 if (mirrors_found) { 754 PetscInt *local_indices,*global_indices; 755 /* get arrays of local and global indices */ 756 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 757 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 758 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 759 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 760 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 761 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 762 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 763 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 764 /* allocate space for mirrors */ 765 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 766 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 767 graph->mirrors_set[0] = 0; 768 769 k=0; 770 for (i=0;i<n_shared[0];i++) { 771 j=shared[0][i]; 772 if (graph->count[j] > 1) { 773 graph->mirrors[j]++; 774 k++; 775 } 776 } 777 /* allocate space for set of mirrors */ 778 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 779 for (i=1;i<graph->nvtxs;i++) 780 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 781 782 /* fill arrays */ 783 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 784 for (j=0;j<n_shared[0];j++) { 785 i=shared[0][j]; 786 if (graph->count[i] > 1) 787 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 788 } 789 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 790 for (i=0;i<graph->nvtxs;i++) { 791 if (graph->mirrors[i] > 0) { 792 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 793 j = global_indices[k]; 794 while ( k > 0 && global_indices[k-1] == j) k--; 795 for (j=0;j<graph->mirrors[i];j++) { 796 graph->mirrors_set[i][j]=local_indices[k+j]; 797 } 798 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 799 } 800 } 801 ierr = PetscFree(local_indices);CHKERRQ(ierr); 802 ierr = PetscFree(global_indices);CHKERRQ(ierr); 803 } 804 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 805 ierr = ISDestroy(&to);CHKERRQ(ierr); 806 ierr = ISDestroy(&from);CHKERRQ(ierr); 807 808 /* Count total number of neigh per node */ 809 k=0; 810 for (i=1;i<n_neigh;i++) { 811 k += n_shared[i]; 812 for (j=0;j<n_shared[i];j++) { 813 graph->count[shared[i][j]] += 1; 814 } 815 } 816 /* Allocate space for storing the set of neighbours for each node */ 817 if (graph->nvtxs) { 818 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 819 } 820 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 821 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 822 } 823 /* Get information for sharing subdomains */ 824 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 825 for (i=1;i<n_neigh;i++) { /* dont count myself */ 826 s = n_shared[i]; 827 for (j=0;j<s;j++) { 828 k = shared[i][j]; 829 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 830 graph->count[k] += 1; 831 } 832 } 833 /* sort set of sharing subdomains */ 834 for (i=0;i<graph->nvtxs;i++) { 835 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 836 } 837 /* 838 Get info for dofs splitting 839 User can specify only a subset; an additional field is considered as a complementary set 840 */ 841 for (i=0;i<graph->nvtxs;i++) { 842 graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 843 } 844 for (i=0;i<n_ISForDofs;i++) { 845 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 846 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 847 for (j=0;j<is_size;j++) { 848 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 849 graph->which_dof[is_indices[j]] = i; 850 } 851 } 852 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 853 } 854 /* Take into account Neumann nodes */ 855 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 856 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 857 if (neumann_is) { 858 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 859 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 860 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 861 for (i=0;i<is_size;i++) { 862 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 863 array[is_indices[i]] = 1.0; 864 } 865 } 866 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 867 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 868 } 869 /* Neumann nodes: impose consistency among neighbours */ 870 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 871 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 872 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 873 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 874 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 875 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 876 for (i=0;i<graph->nvtxs;i++) { 877 if (PetscRealPart(array[i]) > 0.1) { 878 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 879 } 880 } 881 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 882 /* Take into account Dirichlet nodes */ 883 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 884 if (dirichlet_is) { 885 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 886 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 887 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 888 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 889 for (i=0;i<is_size;i++){ 890 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 891 k = is_indices[i]; 892 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 893 if (PetscRealPart(array[k]) > 0.1) { 894 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); 895 } 896 array2[k] = 1.0; 897 } 898 } 899 } 900 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 901 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 902 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 903 } 904 /* Dirichlet nodes: impose consistency among neighbours */ 905 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 906 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 907 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 908 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 909 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 910 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 911 for (i=0;i<graph->nvtxs;i++) { 912 if (PetscRealPart(array[i]) > 0.1) { 913 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 914 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 915 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 916 } 917 } 918 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 919 920 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 921 if (graph->mirrors) { 922 for (i=0;i<graph->nvtxs;i++) 923 if (graph->mirrors[i]) 924 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 925 926 if (graph->xadj && graph->adjncy) { 927 PetscInt *new_xadj,*new_adjncy; 928 /* sort CSR graph */ 929 for (i=0;i<graph->nvtxs;i++) 930 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 931 932 /* adapt local CSR graph in case of local periodicity */ 933 k=0; 934 for (i=0;i<graph->nvtxs;i++) 935 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 936 k += graph->mirrors[graph->adjncy[j]]; 937 938 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 939 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 940 new_xadj[0]=0; 941 for (i=0;i<graph->nvtxs;i++) { 942 k = graph->xadj[i+1]-graph->xadj[i]; 943 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 944 new_xadj[i+1]=new_xadj[i]+k; 945 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 946 k = graph->mirrors[graph->adjncy[j]]; 947 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 948 new_xadj[i+1]+=k; 949 } 950 k = new_xadj[i+1]-new_xadj[i]; 951 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 952 new_xadj[i+1]=new_xadj[i]+k; 953 } 954 /* set new CSR into graph */ 955 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 956 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 957 graph->xadj = new_xadj; 958 graph->adjncy = new_adjncy; 959 } 960 } 961 962 /* mark special nodes -> each will become a single node equivalence class */ 963 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 964 if (custom_primal_vertices) { 965 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 966 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 967 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 968 for (i=0;i<is_size;i++){ 969 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 970 array[is_indices[i]] = 1.0; 971 } 972 } 973 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 974 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 975 } 976 /* special nodes: impose consistency among neighbours */ 977 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 978 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 979 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 980 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 981 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 982 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 983 j = 0; 984 for (i=0;i<graph->nvtxs;i++) { 985 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 986 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 987 j++; 988 } 989 } 990 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 991 992 /* mark interior nodes as touched and belonging to partition number 0 */ 993 for (i=0;i<graph->nvtxs;i++) { 994 if (!graph->count[i]) { 995 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 996 graph->subset[i] = 0; 997 } 998 } 999 /* init graph structure and compute default subsets */ 1000 nodes_touched=0; 1001 for (i=0;i<graph->nvtxs;i++) { 1002 if (PetscBTLookup(graph->touched,i)) { 1003 nodes_touched++; 1004 } 1005 } 1006 i = 0; 1007 graph->ncc = 0; 1008 total_counts = 0; 1009 while (nodes_touched<graph->nvtxs) { 1010 /* find first untouched node in local ordering */ 1011 while (PetscBTLookup(graph->touched,i)) { 1012 i++; 1013 } 1014 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1015 graph->subset[i] = graph->ncc+1; 1016 graph->cptr[graph->ncc] = total_counts; 1017 graph->queue[total_counts] = i; 1018 total_counts++; 1019 nodes_touched++; 1020 /* now find all other nodes having the same set of sharing subdomains */ 1021 for (j=i+1;j<graph->nvtxs;j++) { 1022 /* check for same number of sharing subdomains, dof number and same special mark */ 1023 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]) { 1024 /* check for same set of sharing subdomains */ 1025 same_set=PETSC_TRUE; 1026 for (k=0;k<graph->count[j];k++){ 1027 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 1028 same_set=PETSC_FALSE; 1029 } 1030 } 1031 /* I found a friend of mine */ 1032 if (same_set) { 1033 graph->subset[j]=graph->ncc+1; 1034 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1035 nodes_touched++; 1036 graph->queue[total_counts] = j; 1037 total_counts++; 1038 } 1039 } 1040 } 1041 graph->ncc++; 1042 } 1043 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 1044 graph->n_subsets = graph->ncc; 1045 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1046 for (i=0;i<graph->n_subsets;i++) { 1047 graph->subset_ncc[i] = 1; 1048 } 1049 /* final pointer */ 1050 graph->cptr[graph->ncc] = total_counts; 1051 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 1052 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 1053 /* get a reference node (min index in global ordering) for each subset */ 1054 ierr = PetscMalloc1(graph->ncc,&subset_ref_node_global);CHKERRQ(ierr); 1055 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1056 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1057 for (i=0;i<graph->ncc;i++) { 1058 PetscInt minval = queue_global[graph->cptr[i]]; 1059 PetscInt minloc = graph->queue[graph->cptr[i]]; 1060 for (j=graph->cptr[i]+1;j<graph->cptr[i+1];j++) { 1061 if (minval > queue_global[j]) { 1062 minval = queue_global[j]; 1063 minloc = graph->queue[j]; 1064 } 1065 } 1066 subset_ref_node_global[i] = minloc; 1067 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 1068 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 1069 } 1070 /* renumber reference nodes */ 1071 ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->l2gmap,graph->ncc,subset_ref_node_global,NULL,&k,&graph->subset_ref_node);CHKERRQ(ierr); 1072 ierr = PetscFree(subset_ref_node_global);CHKERRQ(ierr); 1073 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1074 /* free objects */ 1075 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1076 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1077 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1078 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1079 PetscFunctionReturn(0); 1080 } 1081 1082 #undef __FUNCT__ 1083 #define __FUNCT__ "PCBDDCGraphResetCSR" 1084 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1085 { 1086 PetscErrorCode ierr; 1087 1088 PetscFunctionBegin; 1089 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1090 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1091 graph->nvtxs_csr = 0; 1092 PetscFunctionReturn(0); 1093 } 1094 1095 #undef __FUNCT__ 1096 #define __FUNCT__ "PCBDDCGraphReset" 1097 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1098 { 1099 PetscErrorCode ierr; 1100 1101 PetscFunctionBegin; 1102 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1103 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1104 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1105 if (graph->nvtxs) { 1106 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1107 } 1108 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1109 ierr = PetscFree7(graph->count, 1110 graph->neighbours_set, 1111 graph->subset, 1112 graph->which_dof, 1113 graph->cptr, 1114 graph->queue, 1115 graph->special_dof);CHKERRQ(ierr); 1116 if (graph->mirrors) { 1117 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1118 } 1119 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1120 graph->nvtxs = 0; 1121 graph->n_subsets = 0; 1122 graph->custom_minimal_size = 1; 1123 PetscFunctionReturn(0); 1124 } 1125 1126 #undef __FUNCT__ 1127 #define __FUNCT__ "PCBDDCGraphInit" 1128 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap) 1129 { 1130 PetscInt n; 1131 PetscErrorCode ierr; 1132 1133 PetscFunctionBegin; 1134 PetscValidPointer(graph,1); 1135 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1136 /* raise an error if already allocated */ 1137 if (graph->nvtxs) { 1138 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1139 } 1140 /* set number of vertices */ 1141 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1142 graph->l2gmap = l2gmap; 1143 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1144 graph->nvtxs = n; 1145 /* allocate used space */ 1146 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1147 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1148 graph->nvtxs,&graph->neighbours_set, 1149 graph->nvtxs,&graph->subset, 1150 graph->nvtxs,&graph->which_dof, 1151 graph->nvtxs+1,&graph->cptr, 1152 graph->nvtxs,&graph->queue, 1153 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1154 /* zeroes memory */ 1155 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1156 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1157 /* use -1 as a default value for which_dof array */ 1158 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1159 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1160 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1161 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1162 /* zeroes first pointer to neighbour set */ 1163 if (graph->nvtxs) { 1164 graph->neighbours_set[0] = 0; 1165 } 1166 /* zeroes workspace for values of ncc */ 1167 graph->subset_ncc = 0; 1168 graph->subset_ref_node = 0; 1169 PetscFunctionReturn(0); 1170 } 1171 1172 #undef __FUNCT__ 1173 #define __FUNCT__ "PCBDDCGraphDestroy" 1174 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1175 { 1176 PetscErrorCode ierr; 1177 1178 PetscFunctionBegin; 1179 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1180 ierr = PetscFree(*graph);CHKERRQ(ierr); 1181 PetscFunctionReturn(0); 1182 } 1183 1184 #undef __FUNCT__ 1185 #define __FUNCT__ "PCBDDCGraphCreate" 1186 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1187 { 1188 PCBDDCGraph new_graph; 1189 PetscErrorCode ierr; 1190 1191 PetscFunctionBegin; 1192 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1193 /* local to global mapping of dofs */ 1194 new_graph->l2gmap = 0; 1195 /* vertex size */ 1196 new_graph->nvtxs = 0; 1197 new_graph->n_subsets = 0; 1198 new_graph->custom_minimal_size = 1; 1199 /* zeroes ponters */ 1200 new_graph->mirrors = 0; 1201 new_graph->mirrors_set = 0; 1202 new_graph->neighbours_set = 0; 1203 new_graph->subset = 0; 1204 new_graph->which_dof = 0; 1205 new_graph->special_dof = 0; 1206 new_graph->cptr = 0; 1207 new_graph->queue = 0; 1208 new_graph->count = 0; 1209 new_graph->subset_ncc = 0; 1210 new_graph->subset_ref_node = 0; 1211 new_graph->touched = 0; 1212 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1213 new_graph->nvtxs_csr = 0; 1214 new_graph->xadj = 0; 1215 new_graph->adjncy = 0; 1216 *graph = new_graph; 1217 PetscFunctionReturn(0); 1218 } 1219