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