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