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