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 PetscInt node_num=graph->queue[graph->cptr[i]]; 78 PetscBool printcc = PETSC_FALSE; 79 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 80 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 81 for (j=0;j<graph->count[node_num];j++) { 82 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 83 } 84 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 85 if (verbosity_level > 1) { 86 printcc = PETSC_TRUE; 87 } else if (graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == NEUMANN_MARK)) { 88 printcc = PETSC_TRUE; 89 } 90 if (printcc) { 91 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 92 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 93 } 94 } 95 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 96 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 97 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 98 } 99 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 100 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 101 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 102 } 103 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 104 PetscFunctionReturn(0); 105 } 106 107 #undef __FUNCT__ 108 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 109 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) 110 { 111 IS *ISForFaces,*ISForEdges,ISForVertices; 112 PetscInt i,j,nfc,nec,nvc,*idx; 113 PetscBool twodim_flag; 114 PetscErrorCode ierr; 115 116 PetscFunctionBegin; 117 /* loop on ccs to evalute number of faces, edges and vertices */ 118 nfc = 0; 119 nec = 0; 120 nvc = 0; 121 twodim_flag = PETSC_FALSE; 122 for (i=0;i<graph->ncc;i++) { 123 PetscInt repdof = graph->queue[graph->cptr[i]]; 124 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 125 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != NEUMANN_MARK) { 126 nfc++; 127 } else { /* note that nec will be zero in 2d */ 128 nec++; 129 } 130 } else { 131 if (graph->count[repdof] > 1 || (graph->count[repdof] == 1 && graph->special_dof[repdof] == NEUMANN_MARK)) { 132 nvc += graph->cptr[i+1]-graph->cptr[i]; 133 } 134 } 135 } 136 if (!nec) { /* we are in a 2d case -> no faces, only edges */ 137 nec = nfc; 138 nfc = 0; 139 twodim_flag = PETSC_TRUE; 140 } 141 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 142 ISForFaces = 0; 143 ISForEdges = 0; 144 ISForVertices = 0; 145 if (use_faces && nfc) { 146 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 147 } 148 if (use_edges && nec) { 149 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 150 } 151 if (use_vertices && nvc) { 152 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 153 } 154 /* loop on ccs to compute index sets for faces and edges */ 155 nfc = 0; 156 nec = 0; 157 for (i=0;i<graph->ncc;i++) { 158 PetscInt repdof = graph->queue[graph->cptr[i]]; 159 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 160 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != NEUMANN_MARK) { 161 if (twodim_flag) { 162 if (use_edges) { 163 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 164 nec++; 165 } 166 } else { 167 if (use_faces) { 168 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 169 nfc++; 170 } 171 } 172 } else { 173 if (use_edges) { 174 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 175 nec++; 176 } 177 } 178 } 179 } 180 /* index set for vertices */ 181 if (use_vertices && nvc) { 182 nvc = 0; 183 for (i=0;i<graph->ncc;i++) { 184 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 185 PetscInt repdof = graph->queue[graph->cptr[i]]; 186 if (graph->count[repdof] > 1 || (graph->count[repdof] == 1 && graph->special_dof[repdof] == NEUMANN_MARK)) { 187 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 188 idx[nvc]=graph->queue[j]; 189 nvc++; 190 } 191 } 192 } 193 } 194 /* sort vertex set (by local ordering) */ 195 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 196 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 197 } 198 /* get back info */ 199 *n_faces = nfc; 200 *FacesIS = ISForFaces; 201 *n_edges = nec; 202 *EdgesIS = ISForEdges; 203 *VerticesIS = ISForVertices; 204 PetscFunctionReturn(0); 205 } 206 207 #undef __FUNCT__ 208 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 209 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 210 { 211 PetscInt adapt_interface,adapt_interface_reduced; 212 MPI_Comm interface_comm; 213 MPI_Request *send_requests; 214 MPI_Request *recv_requests; 215 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 216 PetscInt i,j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 217 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 218 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 219 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 220 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 221 PetscBool *subset_cc_adapt,same_set; 222 PetscErrorCode ierr; 223 224 PetscFunctionBegin; 225 /* compute connected components locally */ 226 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 227 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 228 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 229 adapt_interface = 0; 230 adapt_interface_reduced = 0; 231 for (i=0;i<graph->n_subsets;i++) { 232 /* We are not sure that on a given subset of the local interface, 233 with two connected components, the latters be the same among sharing subdomains */ 234 if (graph->subset_ncc[i] > 1) { 235 adapt_interface=1; 236 break; 237 } 238 } 239 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_INT,MPI_LOR,interface_comm);CHKERRQ(ierr); 240 241 if (graph->n_subsets && adapt_interface_reduced) { 242 /* Retrict adjacency graph using information from previously computed connected components */ 243 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 244 for (i=0;i<graph->nvtxs;i++) { 245 aux_new_xadj[i]=1; 246 } 247 for (i=0;i<graph->ncc;i++) { 248 k = graph->cptr[i+1]-graph->cptr[i]; 249 for (j=0;j<k;j++) { 250 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 251 } 252 } 253 j = 0; 254 for (i=0;i<graph->nvtxs;i++) { 255 j += aux_new_xadj[i]; 256 } 257 ierr = PetscMalloc1((graph->nvtxs+1),&new_xadj);CHKERRQ(ierr); 258 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 259 new_xadj[0]=0; 260 for (i=0;i<graph->nvtxs;i++) { 261 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 262 if (aux_new_xadj[i]==1) { 263 new_adjncy[new_xadj[i]]=i; 264 } 265 } 266 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 267 for (i=0;i<graph->ncc;i++) { 268 k = graph->cptr[i+1]-graph->cptr[i]; 269 for (j=0;j<k;j++) { 270 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 271 } 272 } 273 /* set new CSR into graph */ 274 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 275 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 276 graph->xadj = new_xadj; 277 graph->adjncy = new_adjncy; 278 /* allocate some space */ 279 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 280 ierr = PetscMalloc1((graph->n_subsets+1),&cum_recv_counts);CHKERRQ(ierr); 281 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 282 ierr = PetscMalloc1(graph->n_subsets,&subset_to_nodes_indices);CHKERRQ(ierr); 283 /* first count how many neighbours per connected component I will receive from */ 284 cum_recv_counts[0]=0; 285 for (i=1;i<graph->n_subsets+1;i++) { 286 j = 0; 287 while (graph->subset[j] != i) { 288 j++; 289 } 290 subset_to_nodes_indices[i-1]=j; 291 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 292 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 293 } 294 ierr = PetscMalloc1(2*cum_recv_counts[graph->n_subsets],&recv_buffer_subset);CHKERRQ(ierr); 295 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&send_requests);CHKERRQ(ierr); 296 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 297 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 298 send_requests[i]=MPI_REQUEST_NULL; 299 recv_requests[i]=MPI_REQUEST_NULL; 300 } 301 /* exchange with my neighbours the number of my connected components on the shared interface */ 302 sum_requests = 0; 303 for (i=0;i<graph->n_subsets;i++) { 304 j = subset_to_nodes_indices[i]; 305 for (k=0;k<graph->count[j];k++) { 306 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 307 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 308 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 309 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 310 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 311 sum_requests++; 312 } 313 } 314 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 315 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 316 /* determine the connected component I need to adapt */ 317 ierr = PetscMalloc1(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 318 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 319 for (i=0;i<graph->n_subsets;i++) { 320 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 321 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 322 if ( graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1 ) { 323 subset_cc_adapt[i] = PETSC_TRUE; 324 break; 325 } 326 } 327 } 328 buffer_size = 0; 329 for (i=0;i<graph->n_subsets;i++) { 330 if (subset_cc_adapt[i]) { 331 for (j=i;j<graph->ncc;j++) { 332 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 333 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 334 } 335 } 336 } 337 } 338 ierr = PetscMalloc1(buffer_size,&send_buffer);CHKERRQ(ierr); 339 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 340 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 341 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 342 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 343 ierr = PetscMalloc1(graph->n_subsets,&sizes_of_sends);CHKERRQ(ierr); 344 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 345 sum_requests = 0; 346 start_of_send = 0; 347 start_of_recv = cum_recv_counts[graph->n_subsets]; 348 for (i=0;i<graph->n_subsets;i++) { 349 if (subset_cc_adapt[i]) { 350 size_of_send = 0; 351 for (j=i;j<graph->ncc;j++) { 352 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 353 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 354 size_of_send += 1; 355 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 356 &queue_in_global_numbering[graph->cptr[j]], 357 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 358 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 359 } 360 } 361 j = subset_to_nodes_indices[i]; 362 sizes_of_sends[i] = size_of_send; 363 for (k=0;k<graph->count[j];k++) { 364 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 365 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 366 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 367 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 368 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 369 sum_requests++; 370 } 371 start_of_send += size_of_send; 372 } 373 } 374 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 375 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 376 buffer_size = 0; 377 for (k=0;k<sum_requests;k++) { 378 buffer_size += recv_buffer_subset[start_of_recv+k]; 379 } 380 ierr = PetscMalloc1(buffer_size,&recv_buffer);CHKERRQ(ierr); 381 /* now exchange the data */ 382 start_of_recv = 0; 383 start_of_send = 0; 384 sum_requests = 0; 385 for (i=0;i<graph->n_subsets;i++) { 386 if (subset_cc_adapt[i]) { 387 size_of_send = sizes_of_sends[i]; 388 j = subset_to_nodes_indices[i]; 389 for (k=0;k<graph->count[j];k++) { 390 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 391 ierr = PetscMPIIntCast((PetscInt)(rank+1)*graph->count[j],&tag);CHKERRQ(ierr); 392 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 393 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 394 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 395 ierr = PetscMPIIntCast((PetscInt)(neigh+1)*graph->count[j],&tag);CHKERRQ(ierr); 396 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 397 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 398 start_of_recv += size_of_recv; 399 sum_requests++; 400 } 401 start_of_send += size_of_send; 402 } 403 } 404 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 405 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 406 for (j=0;j<buffer_size;) { 407 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_MASK,recv_buffer[j],&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 408 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 409 if (graph->mirrors) { 410 PetscBool mirrored_found=PETSC_FALSE; 411 for (k=0;k<recv_buffer[j];k++) { 412 if (graph->mirrors[recv_buffer[j+k+1]]) { 413 mirrored_found=PETSC_TRUE; 414 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 415 } 416 } 417 if (mirrored_found) { 418 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 419 k=0; 420 while (k<recv_buffer[j]) { 421 for (s=1;s<graph->mirrors[recv_buffer[j+1+k]];s++) { 422 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 423 } 424 k+=graph->mirrors[recv_buffer[j+1+k]]+s; 425 } 426 } 427 } 428 k = recv_buffer[j]+1; 429 j += k; 430 } 431 sum_requests = cum_recv_counts[graph->n_subsets]; 432 start_of_recv = 0; 433 ierr = PetscMalloc1(graph->nvtxs,&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 434 global_subset_counter = 0; 435 for (i=0;i<graph->n_subsets;i++) { 436 if (subset_cc_adapt[i]) { 437 temp_buffer_size = 0; 438 /* find nodes on the shared interface we need to adapt */ 439 for (j=0;j<graph->nvtxs;j++) { 440 if (graph->subset[j]==i+1) { 441 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 442 temp_buffer_size++; 443 } else { 444 nodes_to_temp_buffer_indices[j] = -1; 445 } 446 } 447 /* allocate some temporary space */ 448 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 449 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 450 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 451 for (j=1;j<temp_buffer_size;j++) { 452 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 453 } 454 /* analyze contributions from neighbouring subdomains for i-th conn comp 455 temp buffer structure: 456 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 457 3 neighs procs with structured connected components: 458 neigh 0: [0 1 4], [2 3]; (2 connected components) 459 neigh 1: [0 1], [2 3 4]; (2 connected components) 460 neigh 2: [0 4], [1], [2 3]; (3 connected components) 461 tempbuffer (row-oriented) will be filled as: 462 [ 0, 0, 0; 463 0, 0, 1; 464 1, 1, 2; 465 1, 1, 2; 466 0, 1, 0; ]; 467 This way we can simply find intersections of ccs among neighs. 468 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 469 */ 470 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 471 ins_val = 0; 472 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 473 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 474 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 475 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 476 } 477 buffer_size += k; 478 ins_val++; 479 } 480 start_of_recv += size_of_recv; 481 sum_requests++; 482 } 483 ierr = PetscMalloc1(temp_buffer_size,&add_to_subset);CHKERRQ(ierr); 484 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 485 for (j=0;j<temp_buffer_size;j++) { 486 if (!add_to_subset[j]) { /* found a new cc */ 487 global_subset_counter++; 488 add_to_subset[j] = global_subset_counter; 489 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 490 same_set = PETSC_TRUE; 491 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 492 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 493 same_set = PETSC_FALSE; 494 break; 495 } 496 } 497 if (same_set) { 498 add_to_subset[k] = global_subset_counter; 499 } 500 } 501 } 502 } 503 /* insert new data in subset array */ 504 temp_buffer_size = 0; 505 for (j=0;j<graph->nvtxs;j++) { 506 if (graph->subset[j]==i+1) { 507 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 508 temp_buffer_size++; 509 } 510 } 511 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 512 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 513 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 514 } 515 } 516 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 517 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 518 ierr = PetscFree(send_requests);CHKERRQ(ierr); 519 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 520 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 521 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 522 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 523 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 524 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 525 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 526 /* We are ready to find for connected components consistent among neighbouring subdomains */ 527 if (global_subset_counter) { 528 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 529 global_subset_counter = 0; 530 for (i=0;i<graph->nvtxs;i++) { 531 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 532 global_subset_counter++; 533 for (j=i+1;j<graph->nvtxs;j++) { 534 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 535 graph->subset[j] = global_subset_counter; 536 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 537 } 538 } 539 graph->subset[i] = global_subset_counter; 540 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 541 } 542 } 543 /* refine connected components locally */ 544 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 545 } 546 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 547 } 548 PetscFunctionReturn(0); 549 } 550 551 /* The following code has been adapted from function IsConnectedSubdomain contained 552 in source file contig.c of METIS library (version 5.0.1) 553 It finds connected components for each subset */ 554 #undef __FUNCT__ 555 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 556 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 557 { 558 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 559 PetscInt *queue_global; 560 PetscErrorCode ierr; 561 562 PetscFunctionBegin; 563 /* quiet return if no csr info is available */ 564 if (!graph->xadj || !graph->adjncy) { 565 PetscFunctionReturn(0); 566 } 567 568 /* reset any previous search of connected components */ 569 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 570 graph->n_subsets = 0; 571 for (i=0;i<graph->nvtxs;i++) { 572 if (graph->special_dof[i] == DIRICHLET_MARK || !graph->count[i]) { 573 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 574 graph->subset[i] = 0; 575 } 576 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 577 } 578 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 579 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 580 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 581 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 582 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 583 584 /* begin search for connected components */ 585 cum_queue = 0; 586 ncc = 0; 587 for (n=0;n<graph->n_subsets;n++) { 588 pid = n+1; /* partition labeled by 0 is discarded */ 589 nleft = 0; 590 for (i=0;i<graph->nvtxs;i++) { 591 if (graph->subset[i] == pid) { 592 nleft++; 593 } 594 } 595 for (i=0; i<graph->nvtxs; i++) { 596 if (graph->subset[i] == pid) { 597 break; 598 } 599 } 600 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 601 graph->queue[cum_queue] = i; 602 first = 0; 603 last = 1; 604 graph->cptr[ncc] = cum_queue; 605 ncc_pid = 0; 606 while (first != nleft) { 607 if (first == last) { 608 graph->cptr[++ncc] = first+cum_queue; 609 ncc_pid++; 610 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 611 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 612 break; 613 } 614 } 615 graph->queue[cum_queue+last] = i; 616 last++; 617 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 618 } 619 i = graph->queue[cum_queue+first]; 620 first++; 621 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 622 k = graph->adjncy[j]; 623 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 624 graph->queue[cum_queue+last] = k; 625 last++; 626 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 627 } 628 } 629 } 630 graph->cptr[++ncc] = first+cum_queue; 631 ncc_pid++; 632 cum_queue = graph->cptr[ncc]; 633 graph->subset_ncc[n] = ncc_pid; 634 } 635 graph->ncc = ncc; 636 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 637 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 638 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 639 for (i=0;i<graph->ncc;i++) { 640 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 641 } 642 ierr = PetscFree(queue_global);CHKERRQ(ierr); 643 PetscFunctionReturn(0); 644 } 645 646 #undef __FUNCT__ 647 #define __FUNCT__ "PCBDDCGraphSetUp" 648 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 649 { 650 VecScatter scatter_ctx; 651 Vec local_vec,local_vec2,global_vec; 652 IS to,from; 653 MPI_Comm comm; 654 PetscScalar *array,*array2; 655 const PetscInt *is_indices; 656 PetscInt n_neigh,*neigh,*n_shared,**shared; 657 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 658 PetscErrorCode ierr; 659 PetscBool same_set,mirrors_found; 660 661 PetscFunctionBegin; 662 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 663 /* custom_minimal_size */ 664 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 665 /* get info l2gmap and allocate work vectors */ 666 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 667 ierr = ISLocalToGlobalMappingGetIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 668 j = 0; 669 for (i=0;i<graph->nvtxs;i++) { 670 j = PetscMax(j,is_indices[i]); 671 } 672 ierr = MPI_Allreduce(&j,&i,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 673 i++; 674 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 675 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 676 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 677 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 678 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 679 ierr = VecSetSizes(global_vec,PETSC_DECIDE,i);CHKERRQ(ierr); 680 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 681 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 682 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 683 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 684 685 /* get local periodic nodes */ 686 mirrors_found = PETSC_FALSE; 687 if (graph->nvtxs && n_neigh) { 688 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 689 for (i=0; i<n_shared[0]; i++) { 690 if (graph->count[shared[0][i]] > 1) { 691 mirrors_found = PETSC_TRUE; 692 break; 693 } 694 } 695 } 696 /* compute local mirrors (if any) */ 697 if (mirrors_found) { 698 PetscInt *local_indices,*global_indices; 699 /* get arrays of local and global indices */ 700 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 701 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 702 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 703 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 704 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 705 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 706 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 707 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 708 /* allocate space for mirrors */ 709 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 710 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 711 graph->mirrors_set[0] = 0; 712 713 k=0; 714 for (i=0;i<n_shared[0];i++) { 715 j=shared[0][i]; 716 if (graph->count[j] > 1) { 717 graph->mirrors[j]++; 718 k++; 719 } 720 } 721 /* allocate space for set of mirrors */ 722 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 723 for (i=1;i<graph->nvtxs;i++) 724 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 725 726 /* fill arrays */ 727 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 728 for (j=0;j<n_shared[0];j++) { 729 i=shared[0][j]; 730 if (graph->count[i] > 1) 731 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 732 } 733 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 734 for (i=0;i<graph->nvtxs;i++) { 735 if (graph->mirrors[i] > 0) { 736 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 737 j = global_indices[k]; 738 while ( k > 0 && global_indices[k-1] == j) k--; 739 for (j=0;j<graph->mirrors[i];j++) { 740 graph->mirrors_set[i][j]=local_indices[k+j]; 741 } 742 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 743 } 744 } 745 ierr = PetscFree(local_indices);CHKERRQ(ierr); 746 ierr = PetscFree(global_indices);CHKERRQ(ierr); 747 } 748 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 749 ierr = ISDestroy(&to);CHKERRQ(ierr); 750 ierr = ISDestroy(&from);CHKERRQ(ierr); 751 752 /* Count total number of neigh per node */ 753 k=0; 754 for (i=1;i<n_neigh;i++) { 755 k += n_shared[i]; 756 for (j=0;j<n_shared[i];j++) { 757 graph->count[shared[i][j]] += 1; 758 } 759 } 760 /* Allocate space for storing the set of neighbours for each node */ 761 if (graph->nvtxs) { 762 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 763 } 764 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 765 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 766 } 767 /* Get information for sharing subdomains */ 768 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 769 for (i=1;i<n_neigh;i++) { /* dont count myself */ 770 s = n_shared[i]; 771 for (j=0;j<s;j++) { 772 k = shared[i][j]; 773 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 774 graph->count[k] += 1; 775 } 776 } 777 /* sort set of sharing subdomains */ 778 for (i=0;i<graph->nvtxs;i++) { 779 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 780 } 781 /* Get info for dofs splitting */ 782 for (i=0;i<n_ISForDofs;i++) { 783 ierr = ISGetSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 784 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 785 for (j=0;j<is_size;j++) { 786 graph->which_dof[is_indices[j]]=i; 787 } 788 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 789 } 790 /* Take into account Neumann nodes */ 791 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 792 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 793 if (neumann_is) { 794 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 795 ierr = ISGetSize(neumann_is,&is_size);CHKERRQ(ierr); 796 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 797 for (i=0;i<is_size;i++) { 798 array[is_indices[i]] = 1.0; 799 } 800 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 801 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 802 } 803 /* Neumann nodes: impose consistency among neighbours */ 804 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 805 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 806 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 807 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 808 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 809 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 810 for (i=0;i<graph->nvtxs;i++) { 811 if (PetscRealPart(array[i]) > 0.0) { 812 graph->special_dof[i] = NEUMANN_MARK; 813 } 814 } 815 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 816 /* Take into account Dirichlet nodes */ 817 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 818 if (dirichlet_is) { 819 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 820 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 821 ierr = ISGetSize(dirichlet_is,&is_size);CHKERRQ(ierr); 822 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 823 for (i=0;i<is_size;i++){ 824 k = is_indices[i]; 825 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 826 if (PetscRealPart(array[k]) > 0.0) { 827 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); 828 } 829 array2[k] = 1.0; 830 } 831 } 832 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 833 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 834 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 835 } 836 /* Dirichlet nodes: impose consistency among neighbours */ 837 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 838 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 839 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 840 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 841 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 842 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 843 for (i=0;i<graph->nvtxs;i++) { 844 if (PetscRealPart(array[i]) > 0.0) { 845 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 846 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 847 graph->special_dof[i] = DIRICHLET_MARK; 848 } 849 } 850 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 851 852 /* mark local periodic nodes (if any) and adapt CSR graph */ 853 if (graph->mirrors) { 854 PetscInt *new_xadj,*new_adjncy; 855 856 for (i=0;i<graph->nvtxs;i++) 857 if (graph->mirrors[i]) 858 graph->special_dof[i] = LOCAL_PERIODIC_MARK; 859 860 /* sort CSR graph */ 861 for (i=0;i<graph->nvtxs;i++) 862 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 863 864 /* adapt local CSR graph in case of local periodicity */ 865 k=0; 866 for (i=0;i<graph->nvtxs;i++) 867 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 868 k += graph->mirrors[graph->adjncy[j]]; 869 870 ierr = PetscMalloc1((graph->nvtxs+1),&new_xadj);CHKERRQ(ierr); 871 ierr = PetscMalloc1((k+graph->xadj[graph->nvtxs]),&new_adjncy);CHKERRQ(ierr); 872 new_xadj[0]=0; 873 for (i=0;i<graph->nvtxs;i++) { 874 k = graph->xadj[i+1]-graph->xadj[i]; 875 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 876 new_xadj[i+1]=new_xadj[i]+k; 877 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 878 k = graph->mirrors[graph->adjncy[j]]; 879 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 880 new_xadj[i+1]+=k; 881 } 882 k = new_xadj[i+1]-new_xadj[i]; 883 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 884 new_xadj[i+1]=new_xadj[i]+k; 885 } 886 /* set new CSR into graph */ 887 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 888 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 889 graph->xadj = new_xadj; 890 graph->adjncy = new_adjncy; 891 } 892 893 /* mark special nodes -> each will become a single node equivalence class */ 894 if (custom_primal_vertices) { 895 ierr = ISGetSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 896 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 897 for (i=0;i<is_size;i++) { 898 graph->special_dof[is_indices[i]] = SPECIAL_MARK-i; 899 } 900 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 901 } 902 /* mark interior nodes as touched and belonging to partition number 0 */ 903 for (i=0;i<graph->nvtxs;i++) { 904 if (!graph->count[i]) { 905 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 906 graph->subset[i] = 0; 907 } 908 } 909 /* init graph structure and compute default subsets */ 910 nodes_touched=0; 911 for (i=0;i<graph->nvtxs;i++) { 912 if (PetscBTLookup(graph->touched,i)) { 913 nodes_touched++; 914 } 915 } 916 i = 0; 917 graph->ncc = 0; 918 total_counts = 0; 919 while (nodes_touched<graph->nvtxs) { 920 /* find first untouched node in local ordering */ 921 while (PetscBTLookup(graph->touched,i)) { 922 i++; 923 } 924 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 925 graph->subset[i] = graph->ncc+1; 926 graph->cptr[graph->ncc] = total_counts; 927 graph->queue[total_counts] = i; 928 total_counts++; 929 nodes_touched++; 930 /* now find all other nodes having the same set of sharing subdomains */ 931 for (j=i+1;j<graph->nvtxs;j++) { 932 /* check for same number of sharing subdomains, dof number and same special mark */ 933 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]) { 934 /* check for same set of sharing subdomains */ 935 same_set=PETSC_TRUE; 936 for (k=0;k<graph->count[j];k++){ 937 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 938 same_set=PETSC_FALSE; 939 } 940 } 941 /* I found a friend of mine */ 942 if (same_set) { 943 graph->subset[j]=graph->ncc+1; 944 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 945 nodes_touched++; 946 graph->queue[total_counts] = j; 947 total_counts++; 948 } 949 } 950 } 951 graph->ncc++; 952 } 953 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 954 graph->n_subsets = graph->ncc; 955 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 956 for (i=0;i<graph->n_subsets;i++) { 957 graph->subset_ncc[i] = 1; 958 } 959 /* final pointer */ 960 graph->cptr[graph->ncc] = total_counts; 961 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 962 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 963 /* free objects */ 964 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 965 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 966 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 967 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 968 PetscFunctionReturn(0); 969 } 970 971 #undef __FUNCT__ 972 #define __FUNCT__ "PCBDDCGraphResetCSR" 973 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 974 { 975 PetscErrorCode ierr; 976 977 PetscFunctionBegin; 978 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 979 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 980 graph->nvtxs_csr = 0; 981 PetscFunctionReturn(0); 982 } 983 984 #undef __FUNCT__ 985 #define __FUNCT__ "PCBDDCGraphReset" 986 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 987 { 988 PetscErrorCode ierr; 989 990 PetscFunctionBegin; 991 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 992 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 993 if (graph->nvtxs) { 994 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 995 } 996 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 997 ierr = PetscFree7(graph->count, 998 graph->neighbours_set, 999 graph->subset, 1000 graph->which_dof, 1001 graph->cptr, 1002 graph->queue, 1003 graph->special_dof);CHKERRQ(ierr); 1004 if (graph->mirrors) { 1005 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1006 } 1007 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1008 graph->nvtxs = 0; 1009 graph->n_subsets = 0; 1010 graph->custom_minimal_size = 1; 1011 PetscFunctionReturn(0); 1012 } 1013 1014 #undef __FUNCT__ 1015 #define __FUNCT__ "PCBDDCGraphInit" 1016 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap) 1017 { 1018 PetscInt n; 1019 PetscErrorCode ierr; 1020 1021 PetscFunctionBegin; 1022 PetscValidPointer(graph,1); 1023 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1024 /* raise an error if already allocated */ 1025 if (graph->nvtxs) { 1026 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1027 } 1028 /* set number of vertices */ 1029 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1030 graph->l2gmap = l2gmap; 1031 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1032 graph->nvtxs = n; 1033 /* allocate used space */ 1034 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1035 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1036 graph->nvtxs,&graph->neighbours_set, 1037 graph->nvtxs,&graph->subset, 1038 graph->nvtxs,&graph->which_dof, 1039 graph->nvtxs+1,&graph->cptr, 1040 graph->nvtxs,&graph->queue, 1041 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1042 /* zeroes memory */ 1043 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1044 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1045 ierr = PetscMemzero(graph->which_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1046 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1047 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1048 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1049 /* zeroes first pointer to neighbour set */ 1050 if (graph->nvtxs) { 1051 graph->neighbours_set[0] = 0; 1052 } 1053 /* zeroes workspace for values of ncc */ 1054 graph->subset_ncc = 0; 1055 PetscFunctionReturn(0); 1056 } 1057 1058 #undef __FUNCT__ 1059 #define __FUNCT__ "PCBDDCGraphDestroy" 1060 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1061 { 1062 PetscErrorCode ierr; 1063 1064 PetscFunctionBegin; 1065 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1066 ierr = PetscFree(*graph);CHKERRQ(ierr); 1067 PetscFunctionReturn(0); 1068 } 1069 1070 #undef __FUNCT__ 1071 #define __FUNCT__ "PCBDDCGraphCreate" 1072 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1073 { 1074 PCBDDCGraph new_graph; 1075 PetscErrorCode ierr; 1076 1077 PetscFunctionBegin; 1078 ierr = PetscMalloc(sizeof(*new_graph),&new_graph);CHKERRQ(ierr); 1079 /* local to global mapping of dofs */ 1080 new_graph->l2gmap = 0; 1081 /* vertex size */ 1082 new_graph->nvtxs = 0; 1083 new_graph->n_subsets = 0; 1084 new_graph->custom_minimal_size = 1; 1085 /* zeroes ponters */ 1086 new_graph->mirrors = 0; 1087 new_graph->mirrors_set = 0; 1088 new_graph->neighbours_set = 0; 1089 new_graph->subset = 0; 1090 new_graph->which_dof = 0; 1091 new_graph->special_dof = 0; 1092 new_graph->cptr = 0; 1093 new_graph->queue = 0; 1094 new_graph->count = 0; 1095 new_graph->subset_ncc = 0; 1096 new_graph->touched = 0; 1097 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1098 new_graph->nvtxs_csr = 0; 1099 new_graph->xadj = 0; 1100 new_graph->adjncy = 0; 1101 *graph = new_graph; 1102 PetscFunctionReturn(0); 1103 } 1104