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