1 #include <petsc-private/petscimpl.h> 2 #include "bddcprivate.h" 3 #include "bddcstructs.h" 4 5 /* special marks */ 6 #define NEUMANN_MARK -1 7 #define DIRICHLET_MARK -2 8 #define LOCAL_PERIODIC_MARK -3 9 #define SPECIAL_MARK -4 10 11 #undef __FUNCT__ 12 #define __FUNCT__ "PCBDDCGraphASCIIView" 13 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 14 { 15 PetscInt i,j; 16 PetscInt* queue_in_global_numbering; 17 PetscErrorCode ierr; 18 19 PetscFunctionBegin; 20 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 21 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 23 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 24 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 25 if (verbosity_level > 1) { 26 for (i=0;i<graph->nvtxs;i++) { 27 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 28 ierr = PetscViewerASCIISynchronizedPrintf(viewer," touched: %d\n",graph->touched[i]);CHKERRQ(ierr); 29 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 30 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 31 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 32 if (graph->count[i]) { 33 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 34 for (j=0;j<graph->count[i];j++) { 35 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 36 } 37 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 38 } 39 if (graph->mirrors) { 40 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 41 if (graph->mirrors[i]) { 42 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 43 for (j=0;j<graph->mirrors[i];j++) { 44 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 45 } 46 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 47 } 48 } 49 if (verbosity_level > 2) { 50 if (graph->xadj && graph->adjncy) { 51 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 52 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 53 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 54 } 55 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 56 } 57 } 58 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 59 if (graph->subset[i] && graph->subset_ncc) { 60 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 61 } 62 } 63 } 64 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 65 ierr = PetscMalloc(graph->cptr[graph->ncc]*sizeof(*queue_in_global_numbering),&queue_in_global_numbering);CHKERRQ(ierr); 66 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 67 for (i=0;i<graph->ncc;i++) { 68 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 69 PetscInt node_num=graph->queue[graph->cptr[i]]; 70 for (j=0;j<graph->count[node_num];j++) { 71 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 72 } 73 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 74 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 75 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 76 } 77 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 78 } 79 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 80 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 81 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 82 } 83 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 84 PetscFunctionReturn(0); 85 } 86 87 #undef __FUNCT__ 88 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 89 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscBool use_faces, PetscBool use_edges, PetscBool use_vertices, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 90 { 91 IS *ISForFaces,*ISForEdges,ISForVertices; 92 PetscInt i,j,nfc,nec,nvc,*idx; 93 PetscBool twodim_flag; 94 PetscErrorCode ierr; 95 96 PetscFunctionBegin; 97 /* loop on ccs to evalute number of faces, edges and vertices */ 98 nfc = 0; 99 nec = 0; 100 nvc = 0; 101 twodim_flag = PETSC_FALSE; 102 for (i=0;i<graph->ncc;i++) { 103 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 104 if (graph->count[graph->queue[graph->cptr[i]]] == 1 && graph->special_dof[graph->queue[graph->cptr[i]]] != NEUMANN_MARK) { 105 nfc++; 106 } else { /* note that nec will be zero in 2d */ 107 nec++; 108 } 109 } else { 110 nvc += graph->cptr[i+1]-graph->cptr[i]; 111 } 112 } 113 if (!nec) { /* we are in a 2d case -> no faces, only edges */ 114 nec = nfc; 115 nfc = 0; 116 twodim_flag = PETSC_TRUE; 117 } 118 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 119 ISForFaces = 0; 120 ISForEdges = 0; 121 ISForVertices = 0; 122 if (use_faces) { 123 ierr = PetscMalloc(nfc*sizeof(IS),&ISForFaces);CHKERRQ(ierr); 124 } 125 if (use_edges) { 126 ierr = PetscMalloc(nec*sizeof(IS),&ISForEdges);CHKERRQ(ierr); 127 } 128 if (use_vertices) { 129 ierr = PetscMalloc(nvc*sizeof(PetscInt),&idx);CHKERRQ(ierr); 130 } 131 /* loop on ccs to compute index sets for faces and edges */ 132 nfc = 0; 133 nec = 0; 134 nvc = 0; 135 for (i=0;i<graph->ncc;i++) { 136 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 137 if (graph->count[graph->queue[graph->cptr[i]]] == 1 && graph->special_dof[graph->queue[graph->cptr[i]]] != NEUMANN_MARK) { 138 if (twodim_flag) { 139 if (use_edges) { 140 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 141 nec++; 142 } 143 } else { 144 if (use_faces) { 145 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 146 nfc++; 147 } 148 } 149 } else { 150 if (use_edges) { 151 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 152 nec++; 153 } 154 } 155 } 156 } 157 /* index set for vertices */ 158 if (use_vertices) { 159 for (i=0;i<graph->ncc;i++) { 160 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 161 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 162 idx[nvc]=graph->queue[j]; 163 nvc++; 164 } 165 } 166 } 167 /* sort vertex set (by local ordering) */ 168 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 169 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 170 } 171 /* get back info */ 172 *n_faces = nfc; 173 *FacesIS = ISForFaces; 174 *n_edges = nec; 175 *EdgesIS = ISForEdges; 176 *VerticesIS = ISForVertices; 177 PetscFunctionReturn(0); 178 } 179 180 #undef __FUNCT__ 181 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 182 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 183 { 184 PetscInt adapt_interface,adapt_interface_reduced; 185 MPI_Comm interface_comm; 186 MPI_Request *send_requests; 187 MPI_Request *recv_requests; 188 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 189 PetscInt i,j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 190 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 191 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 192 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 193 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 194 PetscBool *subset_cc_adapt,same_set; 195 PetscErrorCode ierr; 196 197 PetscFunctionBegin; 198 /* compute connected components locally */ 199 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 200 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 201 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 202 adapt_interface = 0; 203 adapt_interface_reduced = 0; 204 for (i=0;i<graph->n_subsets;i++) { 205 /* We are not sure that on a given subset of the local interface, 206 with two connected components, the latters be the same among sharing subdomains */ 207 if (graph->subset_ncc[i] > 1) { 208 adapt_interface=1; 209 break; 210 } 211 } 212 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_INT,MPI_LOR,interface_comm);CHKERRQ(ierr); 213 214 if (graph->n_subsets && adapt_interface_reduced) { 215 /* Retrict adjacency graph using information from previously computed connected components */ 216 ierr = PetscMalloc(graph->nvtxs*sizeof(PetscInt),&aux_new_xadj);CHKERRQ(ierr); 217 for (i=0;i<graph->nvtxs;i++) { 218 aux_new_xadj[i]=1; 219 } 220 for (i=0;i<graph->ncc;i++) { 221 k = graph->cptr[i+1]-graph->cptr[i]; 222 for (j=0;j<k;j++) { 223 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 224 } 225 } 226 j = 0; 227 for (i=0;i<graph->nvtxs;i++) { 228 j += aux_new_xadj[i]; 229 } 230 ierr = PetscMalloc((graph->nvtxs+1)*sizeof(PetscInt),&new_xadj);CHKERRQ(ierr); 231 ierr = PetscMalloc(j*sizeof(PetscInt),&new_adjncy);CHKERRQ(ierr); 232 new_xadj[0]=0; 233 for (i=0;i<graph->nvtxs;i++) { 234 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 235 if (aux_new_xadj[i]==1) { 236 new_adjncy[new_xadj[i]]=i; 237 } 238 } 239 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 240 for (i=0;i<graph->ncc;i++) { 241 k = graph->cptr[i+1]-graph->cptr[i]; 242 for (j=0;j<k;j++) { 243 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 244 } 245 } 246 /* set new CSR into graph */ 247 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 248 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 249 graph->xadj = new_xadj; 250 graph->adjncy = new_adjncy; 251 /* allocate some space */ 252 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 253 ierr = PetscMalloc((graph->n_subsets+1)*sizeof(*cum_recv_counts),&cum_recv_counts);CHKERRQ(ierr); 254 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 255 ierr = PetscMalloc(graph->n_subsets*sizeof(*subset_to_nodes_indices),&subset_to_nodes_indices);CHKERRQ(ierr); 256 /* first count how many neighbours per connected component I will receive from */ 257 cum_recv_counts[0]=0; 258 for (i=1;i<graph->n_subsets+1;i++) { 259 j = 0; 260 while (graph->subset[j] != i) { 261 j++; 262 } 263 subset_to_nodes_indices[i-1]=j; 264 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 265 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 266 } 267 ierr = PetscMalloc(2*cum_recv_counts[graph->n_subsets]*sizeof(*recv_buffer_subset),&recv_buffer_subset);CHKERRQ(ierr); 268 ierr = PetscMalloc(cum_recv_counts[graph->n_subsets]*sizeof(MPI_Request),&send_requests);CHKERRQ(ierr); 269 ierr = PetscMalloc(cum_recv_counts[graph->n_subsets]*sizeof(MPI_Request),&recv_requests);CHKERRQ(ierr); 270 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 271 send_requests[i]=MPI_REQUEST_NULL; 272 recv_requests[i]=MPI_REQUEST_NULL; 273 } 274 /* exchange with my neighbours the number of my connected components on the shared interface */ 275 sum_requests = 0; 276 for (i=0;i<graph->n_subsets;i++) { 277 j = subset_to_nodes_indices[i]; 278 for (k=0;k<graph->count[j];k++) { 279 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 280 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 281 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 282 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 283 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 284 sum_requests++; 285 } 286 } 287 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 288 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 289 /* determine the connected component I need to adapt */ 290 ierr = PetscMalloc(graph->n_subsets*sizeof(*subset_cc_adapt),&subset_cc_adapt);CHKERRQ(ierr); 291 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 292 for (i=0;i<graph->n_subsets;i++) { 293 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 294 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 295 if ( graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1 ) { 296 subset_cc_adapt[i] = PETSC_TRUE; 297 break; 298 } 299 } 300 } 301 buffer_size = 0; 302 for (i=0;i<graph->n_subsets;i++) { 303 if (subset_cc_adapt[i]) { 304 for (j=i;j<graph->ncc;j++) { 305 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 306 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 307 } 308 } 309 } 310 } 311 ierr = PetscMalloc(buffer_size*sizeof(*send_buffer),&send_buffer);CHKERRQ(ierr); 312 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 313 ierr = PetscMalloc(graph->cptr[graph->ncc]*sizeof(*queue_in_global_numbering),&queue_in_global_numbering);CHKERRQ(ierr); 314 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 315 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 316 ierr = PetscMalloc(graph->n_subsets*sizeof(*sizes_of_sends),&sizes_of_sends);CHKERRQ(ierr); 317 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 318 sum_requests = 0; 319 start_of_send = 0; 320 start_of_recv = cum_recv_counts[graph->n_subsets]; 321 for (i=0;i<graph->n_subsets;i++) { 322 if (subset_cc_adapt[i]) { 323 size_of_send = 0; 324 for (j=i;j<graph->ncc;j++) { 325 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 326 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 327 size_of_send += 1; 328 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 329 &queue_in_global_numbering[graph->cptr[j]], 330 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 331 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 332 } 333 } 334 j = subset_to_nodes_indices[i]; 335 sizes_of_sends[i] = size_of_send; 336 for (k=0;k<graph->count[j];k++) { 337 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 338 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 339 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 340 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 341 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 342 sum_requests++; 343 } 344 start_of_send += size_of_send; 345 } 346 } 347 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 348 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 349 buffer_size = 0; 350 for (k=0;k<sum_requests;k++) { 351 buffer_size += recv_buffer_subset[start_of_recv+k]; 352 } 353 ierr = PetscMalloc(buffer_size*sizeof(*recv_buffer),&recv_buffer);CHKERRQ(ierr); 354 /* now exchange the data */ 355 start_of_recv = 0; 356 start_of_send = 0; 357 sum_requests = 0; 358 for (i=0;i<graph->n_subsets;i++) { 359 if (subset_cc_adapt[i]) { 360 size_of_send = sizes_of_sends[i]; 361 j = subset_to_nodes_indices[i]; 362 for (k=0;k<graph->count[j];k++) { 363 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 364 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 365 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 366 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 367 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 368 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 369 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 370 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 371 start_of_recv += size_of_recv; 372 sum_requests++; 373 } 374 start_of_send += size_of_send; 375 } 376 } 377 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 378 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 379 for (j=0;j<buffer_size;) { 380 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_MASK,recv_buffer[j],&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 381 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 382 if (graph->mirrors) { 383 PetscBool mirrored_found=PETSC_FALSE; 384 for (k=0;k<recv_buffer[j];k++) { 385 if (graph->mirrors[recv_buffer[j+k+1]]) { 386 mirrored_found=PETSC_TRUE; 387 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 388 } 389 } 390 if (mirrored_found) { 391 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 392 k=0; 393 while (k<recv_buffer[j]) { 394 for (s=1;s<graph->mirrors[recv_buffer[j+1+k]];s++) { 395 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 396 } 397 k+=graph->mirrors[recv_buffer[j+1+k]]+s; 398 } 399 } 400 } 401 k = recv_buffer[j]+1; 402 j += k; 403 } 404 sum_requests = cum_recv_counts[graph->n_subsets]; 405 start_of_recv = 0; 406 ierr = PetscMalloc(graph->nvtxs*sizeof(*nodes_to_temp_buffer_indices),&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 407 global_subset_counter = 0; 408 for (i=0;i<graph->n_subsets;i++) { 409 if (subset_cc_adapt[i]) { 410 temp_buffer_size = 0; 411 /* find nodes on the shared interface we need to adapt */ 412 for (j=0;j<graph->nvtxs;j++) { 413 if (graph->subset[j]==i+1) { 414 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 415 temp_buffer_size++; 416 } else { 417 nodes_to_temp_buffer_indices[j] = -1; 418 } 419 } 420 /* allocate some temporary space */ 421 ierr = PetscMalloc(temp_buffer_size*sizeof(PetscInt*),&temp_buffer);CHKERRQ(ierr); 422 ierr = PetscMalloc(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt),&temp_buffer[0]);CHKERRQ(ierr); 423 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 424 for (j=1;j<temp_buffer_size;j++) { 425 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 426 } 427 /* analyze contributions from neighbouring subdomains for i-th conn comp 428 temp buffer structure: 429 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 430 3 neighs procs with structured connected components: 431 neigh 0: [0 1 4], [2 3]; (2 connected components) 432 neigh 1: [0 1], [2 3 4]; (2 connected components) 433 neigh 2: [0 4], [1], [2 3]; (3 connected components) 434 tempbuffer (row-oriented) will be filled as: 435 [ 0, 0, 0; 436 0, 0, 1; 437 1, 1, 2; 438 1, 1, 2; 439 0, 1, 0; ]; 440 This way we can simply find intersections of ccs among neighs. 441 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 442 */ 443 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 444 ins_val = 0; 445 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 446 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 447 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 448 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 449 } 450 buffer_size += k; 451 ins_val++; 452 } 453 start_of_recv += size_of_recv; 454 sum_requests++; 455 } 456 ierr = PetscMalloc(temp_buffer_size*sizeof(*add_to_subset),&add_to_subset);CHKERRQ(ierr); 457 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 458 for (j=0;j<temp_buffer_size;j++) { 459 if (!add_to_subset[j]) { /* found a new cc */ 460 global_subset_counter++; 461 add_to_subset[j] = global_subset_counter; 462 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 463 same_set = PETSC_TRUE; 464 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 465 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 466 same_set = PETSC_FALSE; 467 break; 468 } 469 } 470 if (same_set) { 471 add_to_subset[k] = global_subset_counter; 472 } 473 } 474 } 475 } 476 /* insert new data in subset array */ 477 temp_buffer_size = 0; 478 for (j=0;j<graph->nvtxs;j++) { 479 if (graph->subset[j]==i+1) { 480 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 481 temp_buffer_size++; 482 } 483 } 484 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 485 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 486 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 487 } 488 } 489 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 490 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 491 ierr = PetscFree(send_requests);CHKERRQ(ierr); 492 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 493 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 494 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 495 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 496 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 497 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 498 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 499 /* We are ready to find for connected components consistent among neighbouring subdomains */ 500 if (global_subset_counter) { 501 ierr = PetscMemzero(graph->touched,graph->nvtxs*sizeof(*graph->touched));CHKERRQ(ierr); 502 global_subset_counter = 0; 503 for (i=0;i<graph->nvtxs;i++) { 504 if (graph->subset[i] && !graph->touched[i]) { 505 global_subset_counter++; 506 for (j=i+1;j<graph->nvtxs;j++) { 507 if (!graph->touched[j] && graph->subset[j]==graph->subset[i]) { 508 graph->subset[j] = global_subset_counter; 509 graph->touched[j] = PETSC_TRUE; 510 } 511 } 512 graph->subset[i] = global_subset_counter; 513 graph->touched[i] = PETSC_TRUE; 514 } 515 } 516 /* refine connected components locally */ 517 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 518 } 519 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 520 } 521 PetscFunctionReturn(0); 522 } 523 524 /* The following code has been adapted from function IsConnectedSubdomain contained 525 in source file contig.c of METIS library (version 5.0.1) 526 It finds connected components for each subset */ 527 #undef __FUNCT__ 528 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 529 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 530 { 531 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 532 PetscInt *queue_global; 533 PetscErrorCode ierr; 534 535 PetscFunctionBegin; 536 /* quiet return if no csr info is available */ 537 if (!graph->xadj || !graph->adjncy) { 538 PetscFunctionReturn(0); 539 } 540 541 /* reset any previous search of connected components */ 542 ierr = PetscMemzero(graph->touched,graph->nvtxs*sizeof(*graph->touched));CHKERRQ(ierr); 543 graph->n_subsets = 0; 544 for (i=0;i<graph->nvtxs;i++) { 545 if (graph->special_dof[i] == DIRICHLET_MARK || !graph->count[i]) { 546 graph->touched[i] = PETSC_TRUE; 547 graph->subset[i] = 0; 548 } 549 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 550 } 551 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 552 ierr = PetscMalloc(graph->n_subsets*sizeof(*graph->subset_ncc),&graph->subset_ncc);CHKERRQ(ierr); 553 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 554 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 555 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 556 557 /* begin search for connected components */ 558 cum_queue = 0; 559 ncc = 0; 560 for (n=0;n<graph->n_subsets;n++) { 561 pid = n+1; /* partition labeled by 0 is discarded */ 562 nleft = 0; 563 for (i=0;i<graph->nvtxs;i++) { 564 if (graph->subset[i] == pid) { 565 nleft++; 566 } 567 } 568 for (i=0; i<graph->nvtxs; i++) { 569 if (graph->subset[i] == pid) { 570 break; 571 } 572 } 573 graph->touched[i] = PETSC_TRUE; 574 graph->queue[cum_queue] = i; 575 first = 0; 576 last = 1; 577 graph->cptr[ncc] = cum_queue; 578 ncc_pid = 0; 579 while (first != nleft) { 580 if (first == last) { 581 graph->cptr[++ncc] = first+cum_queue; 582 ncc_pid++; 583 for (i=0; i<graph->nvtxs; i++) { 584 if (graph->subset[i] == pid && !graph->touched[i]) { 585 break; 586 } 587 } 588 graph->queue[cum_queue+last] = i; 589 last++; 590 graph->touched[i] = PETSC_TRUE; 591 } 592 i = graph->queue[cum_queue+first]; 593 first++; 594 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 595 k = graph->adjncy[j]; 596 if (graph->subset[k] == pid && !graph->touched[k]) { 597 graph->queue[cum_queue+last] = k; 598 last++; 599 graph->touched[k] = PETSC_TRUE; 600 } 601 } 602 } 603 graph->cptr[++ncc] = first+cum_queue; 604 ncc_pid++; 605 cum_queue = graph->cptr[ncc]; 606 graph->subset_ncc[n] = ncc_pid; 607 } 608 graph->ncc = ncc; 609 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 610 ierr = PetscMalloc(graph->cptr[graph->ncc]*sizeof(*queue_global),&queue_global);CHKERRQ(ierr); 611 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 612 for (i=0;i<graph->ncc;i++) { 613 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 614 } 615 ierr = PetscFree(queue_global);CHKERRQ(ierr); 616 PetscFunctionReturn(0); 617 } 618 619 #undef __FUNCT__ 620 #define __FUNCT__ "PCBDDCGraphSetUp" 621 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 622 { 623 VecScatter scatter_ctx; 624 Vec local_vec,local_vec2,global_vec; 625 IS to,from; 626 MPI_Comm comm; 627 PetscScalar *array,*array2; 628 const PetscInt *is_indices; 629 PetscInt n_neigh,*neigh,*n_shared,**shared; 630 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 631 PetscErrorCode ierr; 632 PetscBool same_set,mirrors_found; 633 634 PetscFunctionBegin; 635 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 636 /* custom_minimal_size */ 637 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 638 /* get info l2gmap and allocate work vectors */ 639 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 640 ierr = ISLocalToGlobalMappingGetIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 641 j = 0; 642 for (i=0;i<graph->nvtxs;i++) { 643 j = PetscMax(j,is_indices[i]); 644 } 645 ierr = MPI_Allreduce(&j,&i,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 646 i++; 647 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 648 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 649 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 650 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 651 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 652 ierr = VecSetSizes(global_vec,PETSC_DECIDE,i);CHKERRQ(ierr); 653 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 654 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 655 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 656 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 657 658 /* get local periodic nodes */ 659 mirrors_found=PETSC_FALSE; 660 for (i=0;i<n_shared[0];i++) 661 graph->count[shared[0][i]] += 1; 662 for (i=0;i<n_shared[0];i++) { 663 if (graph->count[shared[0][i]] > 1) { 664 mirrors_found=PETSC_TRUE; 665 break; 666 } 667 } 668 /* compute local mirrors (if any) */ 669 if (mirrors_found) { 670 PetscInt *local_indices,*global_indices; 671 /* get arrays of local and global indices */ 672 ierr = PetscMalloc(graph->nvtxs*sizeof(PetscInt),&local_indices);CHKERRQ(ierr); 673 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 674 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 675 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 676 ierr = PetscMalloc(graph->nvtxs*sizeof(PetscInt),&global_indices);CHKERRQ(ierr); 677 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 678 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 679 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 680 /* allocate space for mirrors */ 681 ierr = PetscMalloc2(graph->nvtxs,PetscInt,&graph->mirrors, 682 graph->nvtxs,PetscInt*,&graph->mirrors_set);CHKERRQ(ierr); 683 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 684 graph->mirrors_set[0] = 0; 685 686 k=0; 687 for (i=0;i<n_shared[0];i++) { 688 j=shared[0][i]; 689 if (graph->count[j] > 1) { 690 graph->mirrors[j]++; 691 k++; 692 } 693 } 694 /* allocate space for set of mirrors */ 695 ierr = PetscMalloc(k*sizeof(PetscInt*),&graph->mirrors_set[0]);CHKERRQ(ierr); 696 for (i=1;i<graph->nvtxs;i++) 697 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 698 699 /* fill arrays */ 700 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 701 for (j=0;j<n_shared[0];j++) { 702 i=shared[0][j]; 703 if (graph->count[i] > 1) 704 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 705 } 706 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 707 for (i=0;i<graph->nvtxs;i++) { 708 if (graph->mirrors[i] > 0) { 709 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 710 j = global_indices[k]; 711 while ( k > 0 && global_indices[k-1] == j) k--; 712 for (j=0;j<graph->mirrors[i];j++) { 713 graph->mirrors_set[i][j]=local_indices[k+j]; 714 } 715 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 716 } 717 } 718 ierr = PetscFree(local_indices);CHKERRQ(ierr); 719 ierr = PetscFree(global_indices);CHKERRQ(ierr); 720 } 721 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 722 ierr = ISDestroy(&to);CHKERRQ(ierr); 723 ierr = ISDestroy(&from);CHKERRQ(ierr); 724 725 /* Count total number of neigh per node */ 726 k=0; 727 for (i=1;i<n_neigh;i++) { 728 k += n_shared[i]; 729 for (j=0;j<n_shared[i];j++) { 730 graph->count[shared[i][j]] += 1; 731 } 732 } 733 /* Allocate space for storing the set of neighbours for each node */ 734 if (graph->nvtxs) { 735 ierr = PetscMalloc(k*sizeof(PetscInt),&graph->neighbours_set[0]);CHKERRQ(ierr); 736 } 737 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 738 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 739 } 740 /* Get information for sharing subdomains */ 741 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 742 for (i=1;i<n_neigh;i++) { /* dont count myself */ 743 s = n_shared[i]; 744 for (j=0;j<s;j++) { 745 k = shared[i][j]; 746 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 747 graph->count[k] += 1; 748 } 749 } 750 /* sort set of sharing subdomains */ 751 for (i=0;i<graph->nvtxs;i++) { 752 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 753 } 754 /* Get info for dofs splitting */ 755 for (i=0;i<n_ISForDofs;i++) { 756 ierr = ISGetSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 757 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 758 for (j=0;j<is_size;j++) { 759 graph->which_dof[is_indices[j]]=i; 760 } 761 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 762 } 763 /* Take into account Neumann nodes */ 764 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 765 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 766 if (neumann_is) { 767 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 768 ierr = ISGetSize(neumann_is,&is_size);CHKERRQ(ierr); 769 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 770 for (i=0;i<is_size;i++) { 771 array[is_indices[i]] = 1.0; 772 } 773 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 774 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 775 } 776 /* Neumann nodes: impose consistency among neighbours */ 777 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 778 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 779 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 780 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 781 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 782 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 783 for (i=0;i<graph->nvtxs;i++) { 784 if (PetscRealPart(array[i]) > 0.0) { 785 graph->special_dof[i] = NEUMANN_MARK; 786 } 787 } 788 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 789 /* Take into account Dirichlet nodes */ 790 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 791 if (dirichlet_is) { 792 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 793 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 794 ierr = ISGetSize(dirichlet_is,&is_size);CHKERRQ(ierr); 795 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 796 for (i=0;i<is_size;i++){ 797 k = is_indices[i]; 798 if (graph->count[k] && !graph->touched[k]) { 799 if (PetscRealPart(array[k]) > 0.0) { 800 SETERRQ1(PETSC_COMM_SELF,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); 801 } 802 array2[k] = 1.0; 803 } 804 } 805 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 806 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 807 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 808 } 809 /* Dirichlet nodes: impose consistency among neighbours */ 810 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 811 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 812 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 813 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 814 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 815 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 816 for (i=0;i<graph->nvtxs;i++) { 817 if (PetscRealPart(array[i]) > 0.0) { 818 graph->touched[i] = PETSC_TRUE; 819 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 820 graph->special_dof[i] = DIRICHLET_MARK; 821 } 822 } 823 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 824 825 /* mark local periodic nodes (if any) and adapt CSR graph */ 826 if (graph->mirrors) { 827 PetscInt *new_xadj,*new_adjncy; 828 829 for (i=0;i<graph->nvtxs;i++) 830 if (graph->mirrors[i]) 831 graph->special_dof[i] = LOCAL_PERIODIC_MARK; 832 833 /* sort CSR graph */ 834 for (i=0;i<graph->nvtxs;i++) 835 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 836 837 /* adapt local CSR graph in case of local periodicity */ 838 k=0; 839 for (i=0;i<graph->nvtxs;i++) 840 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 841 k += graph->mirrors[graph->adjncy[j]]; 842 843 ierr = PetscMalloc((graph->nvtxs+1)*sizeof(PetscInt),&new_xadj);CHKERRQ(ierr); 844 ierr = PetscMalloc((k+graph->xadj[graph->nvtxs])*sizeof(PetscInt),&new_adjncy);CHKERRQ(ierr); 845 new_xadj[0]=0; 846 for (i=0;i<graph->nvtxs;i++) { 847 k = graph->xadj[i+1]-graph->xadj[i]; 848 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 849 new_xadj[i+1]=new_xadj[i]+k; 850 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 851 k = graph->mirrors[graph->adjncy[j]]; 852 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 853 new_xadj[i+1]+=k; 854 } 855 k = new_xadj[i+1]-new_xadj[i]; 856 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 857 new_xadj[i+1]=new_xadj[i]+k; 858 } 859 /* set new CSR into graph */ 860 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 861 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 862 graph->xadj = new_xadj; 863 graph->adjncy = new_adjncy; 864 } 865 866 /* mark special nodes -> each will become a single node equivalence class */ 867 if (custom_primal_vertices) { 868 ierr = ISGetSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 869 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 870 for (i=0;i<is_size;i++) { 871 graph->special_dof[is_indices[i]] = SPECIAL_MARK-i; 872 } 873 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 874 } 875 /* mark interior nodes as touched and belonging to partition number 0 */ 876 for (i=0;i<graph->nvtxs;i++) { 877 if (!graph->count[i]) { 878 graph->touched[i] = PETSC_TRUE; 879 graph->subset[i] = 0; 880 } 881 } 882 /* init graph structure and compute default subsets */ 883 nodes_touched=0; 884 for (i=0;i<graph->nvtxs;i++) { 885 if (graph->touched[i]) { 886 nodes_touched++; 887 } 888 } 889 i = 0; 890 graph->ncc = 0; 891 total_counts = 0; 892 while (nodes_touched<graph->nvtxs) { 893 /* find first untouched node in local ordering */ 894 while (graph->touched[i]) { 895 i++; 896 } 897 graph->touched[i] = PETSC_TRUE; 898 graph->subset[i] = graph->ncc+1; 899 graph->cptr[graph->ncc] = total_counts; 900 graph->queue[total_counts] = i; 901 total_counts++; 902 nodes_touched++; 903 /* now find all other nodes having the same set of sharing subdomains */ 904 for (j=i+1;j<graph->nvtxs;j++){ 905 /* check for same number of sharing subdomains, dof number and same special mark */ 906 if (!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]){ 907 /* check for same set of sharing subdomains */ 908 same_set=PETSC_TRUE; 909 for (k=0;k<graph->count[j];k++){ 910 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 911 same_set=PETSC_FALSE; 912 } 913 } 914 /* I found a friend of mine */ 915 if (same_set) { 916 graph->subset[j]=graph->ncc+1; 917 graph->touched[j]=PETSC_TRUE; 918 nodes_touched++; 919 graph->queue[total_counts] = j; 920 total_counts++; 921 } 922 } 923 } 924 graph->ncc++; 925 } 926 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 927 graph->n_subsets = graph->ncc; 928 ierr = PetscMalloc(graph->n_subsets*sizeof(*graph->subset_ncc),&graph->subset_ncc);CHKERRQ(ierr); 929 for (i=0;i<graph->n_subsets;i++) { 930 graph->subset_ncc[i] = 1; 931 } 932 /* final pointer */ 933 graph->cptr[graph->ncc] = total_counts; 934 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 935 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 936 /* free objects */ 937 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 938 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 939 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 940 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 941 PetscFunctionReturn(0); 942 } 943 944 #undef __FUNCT__ 945 #define __FUNCT__ "PCBDDCGraphResetCSR" 946 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 947 { 948 PetscErrorCode ierr; 949 950 PetscFunctionBegin; 951 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 952 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 953 graph->nvtxs_csr = 0; 954 PetscFunctionReturn(0); 955 } 956 957 #undef __FUNCT__ 958 #define __FUNCT__ "PCBDDCGraphReset" 959 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 960 { 961 PetscErrorCode ierr; 962 963 PetscFunctionBegin; 964 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 965 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 966 if (graph->nvtxs) { 967 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 968 } 969 ierr = PetscFree7(graph->touched, 970 graph->count, 971 graph->neighbours_set, 972 graph->subset, 973 graph->which_dof, 974 graph->cptr, 975 graph->queue);CHKERRQ(ierr); 976 ierr = PetscFree(graph->special_dof);CHKERRQ(ierr); 977 if (graph->mirrors) { 978 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 979 } 980 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 981 graph->nvtxs = 0; 982 graph->n_subsets = 0; 983 graph->custom_minimal_size = 1; 984 PetscFunctionReturn(0); 985 } 986 987 #undef __FUNCT__ 988 #define __FUNCT__ "PCBDDCGraphInit" 989 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap) 990 { 991 PetscInt n; 992 PetscErrorCode ierr; 993 994 PetscFunctionBegin; 995 PetscValidPointer(graph,1); 996 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 997 /* reset graph if already allocated */ 998 if (graph->nvtxs) { 999 ierr = PCBDDCGraphReset(graph);CHKERRQ(ierr); 1000 } 1001 /* set number of vertices */ 1002 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1003 graph->l2gmap = l2gmap; 1004 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1005 graph->nvtxs = n; 1006 /* allocate used space */ 1007 ierr = PetscMalloc7(graph->nvtxs,PetscBool,&graph->touched, 1008 graph->nvtxs,PetscInt,&graph->count, 1009 graph->nvtxs,PetscInt*,&graph->neighbours_set, 1010 graph->nvtxs,PetscInt,&graph->subset, 1011 graph->nvtxs,PetscInt,&graph->which_dof, 1012 graph->nvtxs+1,PetscInt,&graph->cptr, 1013 graph->nvtxs,PetscInt,&graph->queue);CHKERRQ(ierr); 1014 ierr = PetscMalloc(graph->nvtxs*sizeof(PetscInt),&graph->special_dof);CHKERRQ(ierr); 1015 /* zeroes memory */ 1016 ierr = PetscMemzero(graph->touched,graph->nvtxs*sizeof(PetscBool));CHKERRQ(ierr); 1017 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1018 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1019 ierr = PetscMemzero(graph->which_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1020 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1021 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1022 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1023 /* zeroes first pointer to neighbour set */ 1024 if (graph->nvtxs) { 1025 graph->neighbours_set[0] = 0; 1026 } 1027 /* zeroes workspace for values of ncc */ 1028 graph->subset_ncc = 0; 1029 PetscFunctionReturn(0); 1030 } 1031 1032 #undef __FUNCT__ 1033 #define __FUNCT__ "PCBDDCGraphDestroy" 1034 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1035 { 1036 PetscErrorCode ierr; 1037 1038 PetscFunctionBegin; 1039 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1040 ierr = PetscFree(*graph);CHKERRQ(ierr); 1041 PetscFunctionReturn(0); 1042 } 1043 1044 #undef __FUNCT__ 1045 #define __FUNCT__ "PCBDDCGraphCreate" 1046 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1047 { 1048 PCBDDCGraph new_graph; 1049 PetscErrorCode ierr; 1050 1051 PetscFunctionBegin; 1052 ierr = PetscMalloc(sizeof(*new_graph),&new_graph);CHKERRQ(ierr); 1053 /* local to global mapping of dofs */ 1054 new_graph->l2gmap = 0; 1055 /* vertex size */ 1056 new_graph->nvtxs = 0; 1057 new_graph->n_subsets = 0; 1058 new_graph->custom_minimal_size = 1; 1059 /* zeroes ponters */ 1060 new_graph->mirrors = 0; 1061 new_graph->mirrors_set = 0; 1062 new_graph->neighbours_set = 0; 1063 new_graph->subset = 0; 1064 new_graph->which_dof = 0; 1065 new_graph->special_dof = 0; 1066 new_graph->cptr = 0; 1067 new_graph->queue = 0; 1068 new_graph->count = 0; 1069 new_graph->subset_ncc = 0; 1070 new_graph->touched = 0; 1071 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1072 new_graph->nvtxs_csr = 0; 1073 new_graph->xadj = 0; 1074 new_graph->adjncy = 0; 1075 *graph = new_graph; 1076 PetscFunctionReturn(0); 1077 } 1078