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 #undef __FUNCT__ 6 #define __FUNCT__ "PCBDDCGraphASCIIView" 7 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 8 { 9 PetscInt i,j,tabs; 10 PetscInt* queue_in_global_numbering; 11 PetscErrorCode ierr; 12 13 PetscFunctionBegin; 14 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 15 ierr = PetscViewerASCIIGetTab(viewer,&tabs);CHKERRQ(ierr); 16 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 17 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 18 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 19 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 20 if (verbosity_level > 1) { 21 for (i=0;i<graph->nvtxs;i++) { 22 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 23 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 24 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 25 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 26 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 27 if (graph->count[i]) { 28 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 29 for (j=0;j<graph->count[i];j++) { 30 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 31 } 32 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 33 } 34 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 35 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 36 if (graph->mirrors) { 37 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 38 if (graph->mirrors[i]) { 39 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 40 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 41 for (j=0;j<graph->mirrors[i];j++) { 42 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 43 } 44 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 45 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 46 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 47 } 48 } 49 if (verbosity_level > 2) { 50 if (graph->xadj && graph->adjncy) { 51 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 52 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 53 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 54 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 55 } 56 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 57 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 58 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 59 } 60 } 61 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 62 if (graph->subset[i] && graph->subset_ncc) { 63 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 64 } 65 } 66 } 67 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 68 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 69 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 70 for (i=0;i<graph->ncc;i++) { 71 PetscInt node_num=graph->queue[graph->cptr[i]]; 72 PetscBool printcc = PETSC_FALSE; 73 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (neighs:",i);CHKERRQ(ierr); 74 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 75 for (j=0;j<graph->count[node_num];j++) { 76 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 77 } 78 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 79 if (verbosity_level > 1) { 80 printcc = PETSC_TRUE; 81 } else if (graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 82 printcc = PETSC_TRUE; 83 } 84 if (printcc) { 85 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 86 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 87 } 88 } 89 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 90 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 91 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 92 } 93 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 94 if (graph->custom_minimal_size > 1 && verbosity_level > 1) { 95 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 96 } 97 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 98 PetscFunctionReturn(0); 99 } 100 101 #undef __FUNCT__ 102 #define __FUNCT__ "PCBDDCGraphGetCandidatesIS" 103 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 104 { 105 IS *ISForFaces,*ISForEdges,ISForVertices; 106 PetscInt i,j,nfc,nec,nvc,*idx; 107 PetscBool twodim_flag; 108 PetscErrorCode ierr; 109 110 PetscFunctionBegin; 111 /* loop on ccs to evalute number of faces, edges and vertices */ 112 nfc = 0; 113 nec = 0; 114 nvc = 0; 115 twodim_flag = PETSC_FALSE; 116 for (i=0;i<graph->ncc;i++) { 117 PetscInt repdof = graph->queue[graph->cptr[i]]; 118 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 119 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 120 nfc++; 121 } else { /* note that nec will be zero in 2d */ 122 nec++; 123 } 124 } else { 125 nvc += graph->cptr[i+1]-graph->cptr[i]; 126 } 127 } 128 j=0; 129 ierr = MPI_Allreduce(&nec,&j,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 130 if (!j) { /* we are in a 2D case -> no faces, only edges */ 131 nec = nfc; 132 nfc = 0; 133 twodim_flag = PETSC_TRUE; 134 } 135 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 136 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 137 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 138 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 139 /* loop on ccs to compute index sets for faces and edges */ 140 nfc = 0; 141 nec = 0; 142 for (i=0;i<graph->ncc;i++) { 143 PetscInt repdof = graph->queue[graph->cptr[i]]; 144 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size) { 145 if (graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 146 if (twodim_flag) { 147 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 148 nec++; 149 } else { 150 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForFaces[nfc]);CHKERRQ(ierr); 151 nfc++; 152 } 153 } else { 154 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_COPY_VALUES,&ISForEdges[nec]);CHKERRQ(ierr); 155 nec++; 156 } 157 } 158 } 159 /* index set for vertices */ 160 if (nvc) { 161 nvc = 0; 162 for (i=0;i<graph->ncc;i++) { 163 if (graph->cptr[i+1]-graph->cptr[i] <= graph->custom_minimal_size) { 164 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 165 idx[nvc]=graph->queue[j]; 166 nvc++; 167 } 168 } 169 } 170 /* sort vertex set (by local ordering) */ 171 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 172 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 173 } 174 /* get back info */ 175 if (n_faces) *n_faces = nfc; 176 if (FacesIS) { 177 *FacesIS = ISForFaces; 178 } else { 179 for (i=0;i<nfc;i++) { 180 ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 181 } 182 ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 183 } 184 if (n_edges) *n_edges = nec; 185 if (EdgesIS) { 186 *EdgesIS = ISForEdges; 187 } else { 188 for (i=0;i<nec;i++) { 189 ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 190 } 191 ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 192 } 193 if (VerticesIS) { 194 *VerticesIS = ISForVertices; 195 } else { 196 ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 197 } 198 PetscFunctionReturn(0); 199 } 200 201 #undef __FUNCT__ 202 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponents" 203 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 204 { 205 PetscInt i,adapt_interface,adapt_interface_reduced; 206 MPI_Comm interface_comm; 207 PetscErrorCode ierr; 208 209 PetscFunctionBegin; 210 /* compute connected components locally */ 211 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 212 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 213 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 214 adapt_interface = 0; 215 adapt_interface_reduced = 0; 216 for (i=0;i<graph->n_subsets;i++) { 217 /* We are not sure that on a given subset of the local interface, 218 with two connected components, the latters be the same among sharing subdomains */ 219 if (graph->subset_ncc[i] > 1) { 220 adapt_interface=1; 221 break; 222 } 223 } 224 ierr = MPI_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_INT,MPI_LOR,interface_comm);CHKERRQ(ierr); 225 226 if (graph->n_subsets && adapt_interface_reduced) { 227 MPI_Request *send_requests; 228 MPI_Request *recv_requests; 229 PetscInt *aux_new_xadj,*new_xadj,*new_adjncy,**temp_buffer; 230 PetscInt *old_xadj,*old_adjncy; 231 PetscInt j,k,s,sum_requests,buffer_size,size_of_recv,temp_buffer_size; 232 PetscMPIInt rank,neigh,tag,mpi_buffer_size; 233 PetscInt *cum_recv_counts,*subset_to_nodes_indices,*recv_buffer_subset,*nodes_to_temp_buffer_indices; 234 PetscInt *send_buffer,*recv_buffer,*queue_in_global_numbering,*sizes_of_sends,*add_to_subset; 235 PetscInt start_of_recv,start_of_send,size_of_send,global_subset_counter,ins_val; 236 PetscBool *subset_cc_adapt,same_set; 237 238 /* Retrict adjacency graph using information from previously computed connected components */ 239 ierr = PetscMalloc1(graph->nvtxs,&aux_new_xadj);CHKERRQ(ierr); 240 for (i=0;i<graph->nvtxs;i++) { 241 aux_new_xadj[i]=1; 242 } 243 for (i=0;i<graph->ncc;i++) { 244 k = graph->cptr[i+1]-graph->cptr[i]; 245 for (j=0;j<k;j++) { 246 aux_new_xadj[graph->queue[graph->cptr[i]+j]]=k; 247 } 248 } 249 j = 0; 250 for (i=0;i<graph->nvtxs;i++) { 251 j += aux_new_xadj[i]; 252 } 253 ierr = PetscMalloc1((graph->nvtxs+1),&new_xadj);CHKERRQ(ierr); 254 ierr = PetscMalloc1(j,&new_adjncy);CHKERRQ(ierr); 255 new_xadj[0]=0; 256 for (i=0;i<graph->nvtxs;i++) { 257 new_xadj[i+1]=new_xadj[i]+aux_new_xadj[i]; 258 if (aux_new_xadj[i]==1) { 259 new_adjncy[new_xadj[i]]=i; 260 } 261 } 262 ierr = PetscFree(aux_new_xadj);CHKERRQ(ierr); 263 for (i=0;i<graph->ncc;i++) { 264 k = graph->cptr[i+1]-graph->cptr[i]; 265 for (j=0;j<k;j++) { 266 ierr = PetscMemcpy(&new_adjncy[new_xadj[graph->queue[graph->cptr[i]+j]]],&graph->queue[graph->cptr[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 267 } 268 } 269 /* set temporarly new CSR into graph */ 270 old_xadj = graph->xadj; 271 old_adjncy = graph->adjncy; 272 graph->xadj = new_xadj; 273 graph->adjncy = new_adjncy; 274 /* allocate some space */ 275 ierr = MPI_Comm_rank(interface_comm,&rank);CHKERRQ(ierr); 276 ierr = PetscMalloc1((graph->n_subsets+1),&cum_recv_counts);CHKERRQ(ierr); 277 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 278 ierr = PetscMalloc1(graph->n_subsets,&subset_to_nodes_indices);CHKERRQ(ierr); 279 /* first count how many neighbours per connected component I will receive from */ 280 cum_recv_counts[0]=0; 281 for (i=1;i<graph->n_subsets+1;i++) { 282 j = 0; 283 while (graph->subset[j] != i) { 284 j++; 285 } 286 subset_to_nodes_indices[i-1]=j; 287 /* We don't want sends/recvs_to/from_self -> here I don't count myself */ 288 cum_recv_counts[i]=cum_recv_counts[i-1]+graph->count[j]; 289 } 290 ierr = PetscMalloc1(2*cum_recv_counts[graph->n_subsets],&recv_buffer_subset);CHKERRQ(ierr); 291 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&send_requests);CHKERRQ(ierr); 292 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 293 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 294 send_requests[i]=MPI_REQUEST_NULL; 295 recv_requests[i]=MPI_REQUEST_NULL; 296 } 297 /* exchange with my neighbours the number of my connected components on the shared interface */ 298 sum_requests = 0; 299 for (i=0;i<graph->n_subsets;i++) { 300 j = subset_to_nodes_indices[i]; 301 ierr = PetscMPIIntCast(graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 302 for (k=0;k<graph->count[j];k++) { 303 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 304 ierr = MPI_Isend(&graph->subset_ncc[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 305 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 306 sum_requests++; 307 } 308 } 309 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 310 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 311 /* determine the connected component I need to adapt */ 312 ierr = PetscMalloc1(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 313 ierr = PetscMemzero(subset_cc_adapt,graph->n_subsets*sizeof(*subset_cc_adapt));CHKERRQ(ierr); 314 for (i=0;i<graph->n_subsets;i++) { 315 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 316 /* The first condition is natural (someone has a different number of ccs than me), the second one is just to be safe */ 317 if (graph->subset_ncc[i] != recv_buffer_subset[j] || graph->subset_ncc[i] > 1) { 318 subset_cc_adapt[i] = PETSC_TRUE; 319 break; 320 } 321 } 322 } 323 buffer_size = 0; 324 for (i=0;i<graph->n_subsets;i++) { 325 if (subset_cc_adapt[i]) { 326 for (j=i;j<graph->ncc;j++) { 327 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 328 buffer_size += 1 + graph->cptr[j+1]-graph->cptr[j]; 329 } 330 } 331 } 332 } 333 ierr = PetscMalloc1(buffer_size,&send_buffer);CHKERRQ(ierr); 334 /* now get from neighbours their ccs (in global numbering) and adapt them (in case it is needed) */ 335 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 336 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 337 /* determine how much data to send (size of each queue plus the global indices) and communicate it to neighbours */ 338 ierr = PetscMalloc1(graph->n_subsets,&sizes_of_sends);CHKERRQ(ierr); 339 ierr = PetscMemzero(sizes_of_sends,graph->n_subsets*sizeof(*sizes_of_sends));CHKERRQ(ierr); 340 sum_requests = 0; 341 start_of_send = 0; 342 start_of_recv = cum_recv_counts[graph->n_subsets]; 343 for (i=0;i<graph->n_subsets;i++) { 344 if (subset_cc_adapt[i]) { 345 size_of_send = 0; 346 for (j=i;j<graph->ncc;j++) { 347 if (graph->subset[graph->queue[graph->cptr[j]]] == i+1) { /* WARNING -> subset values goes from 1 to graph->n_subsets included */ 348 send_buffer[start_of_send+size_of_send]=graph->cptr[j+1]-graph->cptr[j]; 349 size_of_send += 1; 350 ierr = PetscMemcpy(&send_buffer[start_of_send+size_of_send], 351 &queue_in_global_numbering[graph->cptr[j]], 352 (graph->cptr[j+1]-graph->cptr[j])*sizeof(*send_buffer));CHKERRQ(ierr); 353 size_of_send = size_of_send+graph->cptr[j+1]-graph->cptr[j]; 354 } 355 } 356 j = subset_to_nodes_indices[i]; 357 sizes_of_sends[i] = size_of_send; 358 ierr = PetscMPIIntCast(graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 359 for (k=0;k<graph->count[j];k++) { 360 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 361 ierr = MPI_Isend(&sizes_of_sends[i],1,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 362 ierr = MPI_Irecv(&recv_buffer_subset[sum_requests+start_of_recv],1,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 363 sum_requests++; 364 } 365 start_of_send += size_of_send; 366 } 367 } 368 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 369 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 370 buffer_size = 0; 371 for (k=0;k<sum_requests;k++) { 372 buffer_size += recv_buffer_subset[start_of_recv+k]; 373 } 374 ierr = PetscMalloc1(buffer_size,&recv_buffer);CHKERRQ(ierr); 375 /* now exchange the data */ 376 start_of_recv = 0; 377 start_of_send = 0; 378 sum_requests = 0; 379 for (i=0;i<graph->n_subsets;i++) { 380 if (subset_cc_adapt[i]) { 381 size_of_send = sizes_of_sends[i]; 382 j = subset_to_nodes_indices[i]; 383 ierr = PetscMPIIntCast(graph->subset_ref_node[i]+2,&tag);CHKERRQ(ierr); 384 for (k=0;k<graph->count[j];k++) { 385 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 386 ierr = PetscMPIIntCast(size_of_send,&mpi_buffer_size);CHKERRQ(ierr); 387 ierr = MPI_Isend(&send_buffer[start_of_send],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 388 size_of_recv = recv_buffer_subset[cum_recv_counts[graph->n_subsets]+sum_requests]; 389 ierr = PetscMPIIntCast(size_of_recv,&mpi_buffer_size);CHKERRQ(ierr); 390 ierr = MPI_Irecv(&recv_buffer[start_of_recv],mpi_buffer_size,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 391 start_of_recv += size_of_recv; 392 sum_requests++; 393 } 394 start_of_send += size_of_send; 395 } 396 } 397 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 398 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 399 for (j=0;j<buffer_size;) { 400 ierr = ISGlobalToLocalMappingApply(graph->l2gmap,IS_GTOLM_MASK,recv_buffer[j],&recv_buffer[j+1],&recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 401 /* we need to adapt the output of GlobalToLocal mapping if there are mirrored nodes */ 402 if (graph->mirrors) { 403 PetscBool mirrored_found=PETSC_FALSE; 404 for (k=0;k<recv_buffer[j];k++) { 405 if (graph->mirrors[recv_buffer[j+k+1]]) { 406 mirrored_found=PETSC_TRUE; 407 recv_buffer[j+k+1]=graph->mirrors_set[recv_buffer[j+k+1]][0]; 408 } 409 } 410 if (mirrored_found) { 411 ierr = PetscSortInt(recv_buffer[j],&recv_buffer[j+1]);CHKERRQ(ierr); 412 k=0; 413 while (k<recv_buffer[j]) { 414 for (s=1;s<graph->mirrors[recv_buffer[j+1+k]];s++) { 415 recv_buffer[j+1+k+s] = graph->mirrors_set[recv_buffer[j+1+k]][s]; 416 } 417 k+=graph->mirrors[recv_buffer[j+1+k]]+s; 418 } 419 } 420 } 421 k = recv_buffer[j]+1; 422 j += k; 423 } 424 sum_requests = cum_recv_counts[graph->n_subsets]; 425 start_of_recv = 0; 426 ierr = PetscMalloc1(graph->nvtxs,&nodes_to_temp_buffer_indices);CHKERRQ(ierr); 427 global_subset_counter = 0; 428 for (i=0;i<graph->n_subsets;i++) { 429 if (subset_cc_adapt[i]) { 430 temp_buffer_size = 0; 431 /* find nodes on the shared interface we need to adapt */ 432 for (j=0;j<graph->nvtxs;j++) { 433 if (graph->subset[j]==i+1) { 434 nodes_to_temp_buffer_indices[j] = temp_buffer_size; 435 temp_buffer_size++; 436 } else { 437 nodes_to_temp_buffer_indices[j] = -1; 438 } 439 } 440 /* allocate some temporary space */ 441 ierr = PetscMalloc1(temp_buffer_size,&temp_buffer);CHKERRQ(ierr); 442 ierr = PetscMalloc1(temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i]),&temp_buffer[0]);CHKERRQ(ierr); 443 ierr = PetscMemzero(temp_buffer[0],temp_buffer_size*(cum_recv_counts[i+1]-cum_recv_counts[i])*sizeof(PetscInt));CHKERRQ(ierr); 444 for (j=1;j<temp_buffer_size;j++) { 445 temp_buffer[j] = temp_buffer[j-1]+cum_recv_counts[i+1]-cum_recv_counts[i]; 446 } 447 /* analyze contributions from neighbouring subdomains for i-th conn comp 448 temp buffer structure: 449 supposing part of the interface has dimension 5 (for example with global dofs 0,1,2,3,4) 450 3 neighs procs with structured connected components: 451 neigh 0: [0 1 4], [2 3]; (2 connected components) 452 neigh 1: [0 1], [2 3 4]; (2 connected components) 453 neigh 2: [0 4], [1], [2 3]; (3 connected components) 454 tempbuffer (row-oriented) will be filled as: 455 [ 0, 0, 0; 456 0, 0, 1; 457 1, 1, 2; 458 1, 1, 2; 459 0, 1, 0; ]; 460 This way we can simply find intersections of ccs among neighs. 461 For the example above, the graph->subset array will be modified to reproduce the following 4 connected components [0], [1], [2 3], [4]; 462 */ 463 for (j=0;j<cum_recv_counts[i+1]-cum_recv_counts[i];j++) { 464 ins_val = 0; 465 size_of_recv = recv_buffer_subset[sum_requests]; /* total size of recv from neighs */ 466 for (buffer_size=0;buffer_size<size_of_recv;) { /* loop until all data from neighs has been taken into account */ 467 for (k=1;k<recv_buffer[buffer_size+start_of_recv]+1;k++) { /* filling properly temp_buffer using data from a single recv */ 468 temp_buffer[nodes_to_temp_buffer_indices[recv_buffer[start_of_recv+buffer_size+k]]][j] = ins_val; 469 } 470 buffer_size += k; 471 ins_val++; 472 } 473 start_of_recv += size_of_recv; 474 sum_requests++; 475 } 476 ierr = PetscMalloc1(temp_buffer_size,&add_to_subset);CHKERRQ(ierr); 477 ierr = PetscMemzero(add_to_subset,temp_buffer_size*sizeof(*add_to_subset));CHKERRQ(ierr); 478 for (j=0;j<temp_buffer_size;j++) { 479 if (!add_to_subset[j]) { /* found a new cc */ 480 global_subset_counter++; 481 add_to_subset[j] = global_subset_counter; 482 for (k=j+1;k<temp_buffer_size;k++) { /* check for other nodes in new cc */ 483 same_set = PETSC_TRUE; 484 for (s=0;s<cum_recv_counts[i+1]-cum_recv_counts[i];s++) { 485 if (temp_buffer[j][s]!=temp_buffer[k][s]) { 486 same_set = PETSC_FALSE; 487 break; 488 } 489 } 490 if (same_set) { 491 add_to_subset[k] = global_subset_counter; 492 } 493 } 494 } 495 } 496 /* insert new data in subset array */ 497 temp_buffer_size = 0; 498 for (j=0;j<graph->nvtxs;j++) { 499 if (graph->subset[j]==i+1) { 500 graph->subset[j] = graph->n_subsets+add_to_subset[temp_buffer_size]; 501 temp_buffer_size++; 502 } 503 } 504 ierr = PetscFree(temp_buffer[0]);CHKERRQ(ierr); 505 ierr = PetscFree(temp_buffer);CHKERRQ(ierr); 506 ierr = PetscFree(add_to_subset);CHKERRQ(ierr); 507 } 508 } 509 ierr = PetscFree(nodes_to_temp_buffer_indices);CHKERRQ(ierr); 510 ierr = PetscFree(sizes_of_sends);CHKERRQ(ierr); 511 ierr = PetscFree(send_requests);CHKERRQ(ierr); 512 ierr = PetscFree(recv_requests);CHKERRQ(ierr); 513 ierr = PetscFree(recv_buffer);CHKERRQ(ierr); 514 ierr = PetscFree(recv_buffer_subset);CHKERRQ(ierr); 515 ierr = PetscFree(send_buffer);CHKERRQ(ierr); 516 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 517 ierr = PetscFree(subset_to_nodes_indices);CHKERRQ(ierr); 518 ierr = PetscFree(subset_cc_adapt);CHKERRQ(ierr); 519 /* We are ready to find for connected components consistent among neighbouring subdomains */ 520 if (global_subset_counter) { 521 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 522 global_subset_counter = 0; 523 for (i=0;i<graph->nvtxs;i++) { 524 if (graph->subset[i] && !PetscBTLookup(graph->touched,i)) { 525 global_subset_counter++; 526 for (j=i+1;j<graph->nvtxs;j++) { 527 if (!PetscBTLookup(graph->touched,j) && graph->subset[j]==graph->subset[i]) { 528 graph->subset[j] = global_subset_counter; 529 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 530 } 531 } 532 graph->subset[i] = global_subset_counter; 533 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 534 } 535 } 536 /* refine connected components locally */ 537 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 538 } 539 /* restore original CSR graph of dofs */ 540 ierr = PetscFree(new_xadj);CHKERRQ(ierr); 541 ierr = PetscFree(new_adjncy);CHKERRQ(ierr); 542 graph->xadj = old_xadj; 543 graph->adjncy = old_adjncy; 544 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 545 } 546 PetscFunctionReturn(0); 547 } 548 549 /* The following code has been adapted from function IsConnectedSubdomain contained 550 in source file contig.c of METIS library (version 5.0.1) 551 It finds connected components for each subset */ 552 #undef __FUNCT__ 553 #define __FUNCT__ "PCBDDCGraphComputeConnectedComponentsLocal" 554 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 555 { 556 PetscInt i,j,k,first,last,nleft,ncc,pid,cum_queue,n,ncc_pid; 557 PetscInt *queue_global; 558 PetscErrorCode ierr; 559 560 PetscFunctionBegin; 561 /* quiet return if no csr info is available */ 562 if (!graph->xadj || !graph->adjncy) { 563 PetscFunctionReturn(0); 564 } 565 566 /* reset any previous search of connected components */ 567 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 568 graph->n_subsets = 0; 569 for (i=0;i<graph->nvtxs;i++) { 570 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 571 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 572 graph->subset[i] = 0; 573 } 574 graph->n_subsets = PetscMax(graph->n_subsets,graph->subset[i]); 575 } 576 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 577 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 578 ierr = PetscMemzero(graph->subset_ncc,graph->n_subsets*sizeof(*graph->subset_ncc));CHKERRQ(ierr); 579 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(*graph->cptr));CHKERRQ(ierr); 580 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(*graph->queue));CHKERRQ(ierr); 581 582 /* begin search for connected components */ 583 cum_queue = 0; 584 ncc = 0; 585 for (n=0;n<graph->n_subsets;n++) { 586 pid = n+1; /* partition labeled by 0 is discarded */ 587 nleft = 0; 588 for (i=0;i<graph->nvtxs;i++) { 589 if (graph->subset[i] == pid) { 590 nleft++; 591 } 592 } 593 for (i=0; i<graph->nvtxs; i++) { 594 if (graph->subset[i] == pid) { 595 break; 596 } 597 } 598 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 599 graph->queue[cum_queue] = i; 600 first = 0; 601 last = 1; 602 graph->cptr[ncc] = cum_queue; 603 ncc_pid = 0; 604 while (first != nleft) { 605 if (first == last) { 606 graph->cptr[++ncc] = first+cum_queue; 607 ncc_pid++; 608 for (i=0; i<graph->nvtxs; i++) { /* TODO-> use a while! */ 609 if (graph->subset[i] == pid && !PetscBTLookup(graph->touched,i)) { 610 break; 611 } 612 } 613 graph->queue[cum_queue+last] = i; 614 last++; 615 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 616 } 617 i = graph->queue[cum_queue+first]; 618 first++; 619 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 620 k = graph->adjncy[j]; 621 if (graph->subset[k] == pid && !PetscBTLookup(graph->touched,k)) { 622 graph->queue[cum_queue+last] = k; 623 last++; 624 ierr = PetscBTSet(graph->touched,k);CHKERRQ(ierr); 625 } 626 } 627 } 628 graph->cptr[++ncc] = first+cum_queue; 629 ncc_pid++; 630 cum_queue = graph->cptr[ncc]; 631 graph->subset_ncc[n] = ncc_pid; 632 } 633 graph->ncc = ncc; 634 /* For consistency among neighbours, I need to sort (by global ordering) each connected component */ 635 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 636 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 637 for (i=0;i<graph->ncc;i++) { 638 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 639 } 640 ierr = PetscFree(queue_global);CHKERRQ(ierr); 641 PetscFunctionReturn(0); 642 } 643 644 #undef __FUNCT__ 645 #define __FUNCT__ "PCBDDCGraphSetUp" 646 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 647 { 648 VecScatter scatter_ctx; 649 Vec local_vec,local_vec2,global_vec; 650 IS to,from; 651 MPI_Comm comm; 652 PetscScalar *array,*array2; 653 const PetscInt *is_indices; 654 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global; 655 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 656 PetscErrorCode ierr; 657 PetscBool same_set,mirrors_found; 658 659 PetscFunctionBegin; 660 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 661 /* custom_minimal_size */ 662 graph->custom_minimal_size = PetscMax(graph->custom_minimal_size,custom_minimal_size); 663 /* get info l2gmap and allocate work vectors */ 664 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 665 ierr = ISLocalToGlobalMappingGetIndices(graph->l2gmap,&is_indices);CHKERRQ(ierr); 666 j = 0; 667 for (i=0;i<graph->nvtxs;i++) { 668 j = PetscMax(j,is_indices[i]); 669 } 670 ierr = MPI_Allreduce(&j,&i,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 671 i++; 672 ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 673 ierr = VecSetSizes(local_vec,PETSC_DECIDE,graph->nvtxs);CHKERRQ(ierr); 674 ierr = VecSetType(local_vec,VECSTANDARD);CHKERRQ(ierr); 675 ierr = VecDuplicate(local_vec,&local_vec2);CHKERRQ(ierr); 676 ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 677 ierr = VecSetSizes(global_vec,PETSC_DECIDE,i);CHKERRQ(ierr); 678 ierr = VecSetType(global_vec,VECSTANDARD);CHKERRQ(ierr); 679 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 680 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 681 ierr = VecScatterCreate(global_vec,from,local_vec,to,&scatter_ctx);CHKERRQ(ierr); 682 683 /* get local periodic nodes */ 684 mirrors_found = PETSC_FALSE; 685 if (graph->nvtxs && n_neigh) { 686 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 687 for (i=0; i<n_shared[0]; i++) { 688 if (graph->count[shared[0][i]] > 1) { 689 mirrors_found = PETSC_TRUE; 690 break; 691 } 692 } 693 } 694 /* compute local mirrors (if any) */ 695 if (mirrors_found) { 696 PetscInt *local_indices,*global_indices; 697 /* get arrays of local and global indices */ 698 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 699 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 700 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 701 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 702 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 703 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 704 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 705 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 706 /* allocate space for mirrors */ 707 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 708 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 709 graph->mirrors_set[0] = 0; 710 711 k=0; 712 for (i=0;i<n_shared[0];i++) { 713 j=shared[0][i]; 714 if (graph->count[j] > 1) { 715 graph->mirrors[j]++; 716 k++; 717 } 718 } 719 /* allocate space for set of mirrors */ 720 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 721 for (i=1;i<graph->nvtxs;i++) 722 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 723 724 /* fill arrays */ 725 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 726 for (j=0;j<n_shared[0];j++) { 727 i=shared[0][j]; 728 if (graph->count[i] > 1) 729 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 730 } 731 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 732 for (i=0;i<graph->nvtxs;i++) { 733 if (graph->mirrors[i] > 0) { 734 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 735 j = global_indices[k]; 736 while ( k > 0 && global_indices[k-1] == j) k--; 737 for (j=0;j<graph->mirrors[i];j++) { 738 graph->mirrors_set[i][j]=local_indices[k+j]; 739 } 740 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 741 } 742 } 743 ierr = PetscFree(local_indices);CHKERRQ(ierr); 744 ierr = PetscFree(global_indices);CHKERRQ(ierr); 745 } 746 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 747 ierr = ISDestroy(&to);CHKERRQ(ierr); 748 ierr = ISDestroy(&from);CHKERRQ(ierr); 749 750 /* Count total number of neigh per node */ 751 k=0; 752 for (i=1;i<n_neigh;i++) { 753 k += n_shared[i]; 754 for (j=0;j<n_shared[i];j++) { 755 graph->count[shared[i][j]] += 1; 756 } 757 } 758 /* Allocate space for storing the set of neighbours for each node */ 759 if (graph->nvtxs) { 760 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 761 } 762 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 763 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 764 } 765 /* Get information for sharing subdomains */ 766 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 767 for (i=1;i<n_neigh;i++) { /* dont count myself */ 768 s = n_shared[i]; 769 for (j=0;j<s;j++) { 770 k = shared[i][j]; 771 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 772 graph->count[k] += 1; 773 } 774 } 775 /* sort set of sharing subdomains */ 776 for (i=0;i<graph->nvtxs;i++) { 777 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 778 } 779 /* 780 Get info for dofs splitting 781 User can specify only a subset; an additional field is considered as a complementary set 782 */ 783 for (i=0;i<graph->nvtxs;i++) { 784 graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 785 } 786 for (i=0;i<n_ISForDofs;i++) { 787 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 788 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 789 for (j=0;j<is_size;j++) { 790 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 791 graph->which_dof[is_indices[j]] = i; 792 } 793 } 794 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 795 } 796 /* Take into account Neumann nodes */ 797 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 798 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 799 if (neumann_is) { 800 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 801 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 802 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 803 for (i=0;i<is_size;i++) { 804 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 805 array[is_indices[i]] = 1.0; 806 } 807 } 808 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 809 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 810 } 811 /* Neumann nodes: impose consistency among neighbours */ 812 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 813 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 814 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 815 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 816 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 817 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 818 for (i=0;i<graph->nvtxs;i++) { 819 if (PetscRealPart(array[i]) > 0.1) { 820 graph->special_dof[i] = PCBDDCGRAPH_NEUMANN_MARK; 821 } 822 } 823 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 824 /* Take into account Dirichlet nodes */ 825 ierr = VecSet(local_vec2,0.0);CHKERRQ(ierr); 826 if (dirichlet_is) { 827 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 828 ierr = VecGetArray(local_vec2,&array2);CHKERRQ(ierr); 829 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 830 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 831 for (i=0;i<is_size;i++){ 832 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 833 k = is_indices[i]; 834 if (graph->count[k] && !PetscBTLookup(graph->touched,k)) { 835 if (PetscRealPart(array[k]) > 0.1) { 836 SETERRQ1(comm,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); 837 } 838 array2[k] = 1.0; 839 } 840 } 841 } 842 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 843 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 844 ierr = VecRestoreArray(local_vec2,&array2);CHKERRQ(ierr); 845 } 846 /* Dirichlet nodes: impose consistency among neighbours */ 847 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 848 ierr = VecScatterBegin(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 849 ierr = VecScatterEnd(scatter_ctx,local_vec2,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 850 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 851 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 852 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 853 for (i=0;i<graph->nvtxs;i++) { 854 if (PetscRealPart(array[i]) > 0.1) { 855 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 856 graph->subset[i] = 0; /* dirichlet nodes treated as internal -> is it ok? */ 857 graph->special_dof[i] = PCBDDCGRAPH_DIRICHLET_MARK; 858 } 859 } 860 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 861 862 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 863 if (graph->mirrors) { 864 for (i=0;i<graph->nvtxs;i++) 865 if (graph->mirrors[i]) 866 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 867 868 if (graph->xadj && graph->adjncy) { 869 PetscInt *new_xadj,*new_adjncy; 870 /* sort CSR graph */ 871 for (i=0;i<graph->nvtxs;i++) 872 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 873 874 /* adapt local CSR graph in case of local periodicity */ 875 k=0; 876 for (i=0;i<graph->nvtxs;i++) 877 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 878 k += graph->mirrors[graph->adjncy[j]]; 879 880 ierr = PetscMalloc1((graph->nvtxs+1),&new_xadj);CHKERRQ(ierr); 881 ierr = PetscMalloc1((k+graph->xadj[graph->nvtxs]),&new_adjncy);CHKERRQ(ierr); 882 new_xadj[0]=0; 883 for (i=0;i<graph->nvtxs;i++) { 884 k = graph->xadj[i+1]-graph->xadj[i]; 885 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 886 new_xadj[i+1]=new_xadj[i]+k; 887 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 888 k = graph->mirrors[graph->adjncy[j]]; 889 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 890 new_xadj[i+1]+=k; 891 } 892 k = new_xadj[i+1]-new_xadj[i]; 893 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 894 new_xadj[i+1]=new_xadj[i]+k; 895 } 896 /* set new CSR into graph */ 897 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 898 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 899 graph->xadj = new_xadj; 900 graph->adjncy = new_adjncy; 901 } 902 } 903 904 /* mark special nodes -> each will become a single node equivalence class */ 905 ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 906 if (custom_primal_vertices) { 907 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 908 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 909 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 910 for (i=0;i<is_size;i++){ 911 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 912 array[is_indices[i]] = 1.0; 913 } 914 } 915 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 916 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 917 } 918 /* special nodes: impose consistency among neighbours */ 919 ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 920 ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 921 ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 922 ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 923 ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 924 ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 925 j = 0; 926 for (i=0;i<graph->nvtxs;i++) { 927 if (PetscRealPart(array[i]) > 0.1 && graph->special_dof[i] != PCBDDCGRAPH_DIRICHLET_MARK) { 928 graph->special_dof[i] = PCBDDCGRAPH_SPECIAL_MARK-j; 929 j++; 930 } 931 } 932 ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 933 934 /* mark interior nodes as touched and belonging to partition number 0 */ 935 for (i=0;i<graph->nvtxs;i++) { 936 if (!graph->count[i]) { 937 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 938 graph->subset[i] = 0; 939 } 940 } 941 /* init graph structure and compute default subsets */ 942 nodes_touched=0; 943 for (i=0;i<graph->nvtxs;i++) { 944 if (PetscBTLookup(graph->touched,i)) { 945 nodes_touched++; 946 } 947 } 948 i = 0; 949 graph->ncc = 0; 950 total_counts = 0; 951 while (nodes_touched<graph->nvtxs) { 952 /* find first untouched node in local ordering */ 953 while (PetscBTLookup(graph->touched,i)) { 954 i++; 955 } 956 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 957 graph->subset[i] = graph->ncc+1; 958 graph->cptr[graph->ncc] = total_counts; 959 graph->queue[total_counts] = i; 960 total_counts++; 961 nodes_touched++; 962 /* now find all other nodes having the same set of sharing subdomains */ 963 for (j=i+1;j<graph->nvtxs;j++) { 964 /* check for same number of sharing subdomains, dof number and same special mark */ 965 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]) { 966 /* check for same set of sharing subdomains */ 967 same_set=PETSC_TRUE; 968 for (k=0;k<graph->count[j];k++){ 969 if (graph->neighbours_set[i][k]!=graph->neighbours_set[j][k]) { 970 same_set=PETSC_FALSE; 971 } 972 } 973 /* I found a friend of mine */ 974 if (same_set) { 975 graph->subset[j]=graph->ncc+1; 976 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 977 nodes_touched++; 978 graph->queue[total_counts] = j; 979 total_counts++; 980 } 981 } 982 } 983 graph->ncc++; 984 } 985 /* set default number of subsets (at this point no info on csr graph has been taken into account, so n_subsets = ncc */ 986 graph->n_subsets = graph->ncc; 987 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 988 for (i=0;i<graph->n_subsets;i++) { 989 graph->subset_ncc[i] = 1; 990 } 991 /* final pointer */ 992 graph->cptr[graph->ncc] = total_counts; 993 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 994 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 995 /* get a reference node (min index in global ordering) for each subset */ 996 ierr = PetscMalloc1(graph->ncc,&graph->subset_ref_node);CHKERRQ(ierr); 997 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 998 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 999 for (i=0;i<graph->ncc;i++) { 1000 PetscInt minval = queue_global[graph->cptr[i]]; 1001 for (j=graph->cptr[i]+1;j<graph->cptr[i+1];j++) { 1002 minval = PetscMin(minval,queue_global[j]); 1003 } 1004 graph->subset_ref_node[i] = minval; 1005 } 1006 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1007 /* free objects */ 1008 ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 1009 ierr = VecDestroy(&local_vec2);CHKERRQ(ierr); 1010 ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 1011 ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 1012 PetscFunctionReturn(0); 1013 } 1014 1015 #undef __FUNCT__ 1016 #define __FUNCT__ "PCBDDCGraphResetCSR" 1017 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1018 { 1019 PetscErrorCode ierr; 1020 1021 PetscFunctionBegin; 1022 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1023 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1024 graph->nvtxs_csr = 0; 1025 PetscFunctionReturn(0); 1026 } 1027 1028 #undef __FUNCT__ 1029 #define __FUNCT__ "PCBDDCGraphReset" 1030 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1031 { 1032 PetscErrorCode ierr; 1033 1034 PetscFunctionBegin; 1035 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1036 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1037 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1038 if (graph->nvtxs) { 1039 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1040 } 1041 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1042 ierr = PetscFree7(graph->count, 1043 graph->neighbours_set, 1044 graph->subset, 1045 graph->which_dof, 1046 graph->cptr, 1047 graph->queue, 1048 graph->special_dof);CHKERRQ(ierr); 1049 if (graph->mirrors) { 1050 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1051 } 1052 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1053 graph->nvtxs = 0; 1054 graph->n_subsets = 0; 1055 graph->custom_minimal_size = 1; 1056 PetscFunctionReturn(0); 1057 } 1058 1059 #undef __FUNCT__ 1060 #define __FUNCT__ "PCBDDCGraphInit" 1061 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap) 1062 { 1063 PetscInt n; 1064 PetscErrorCode ierr; 1065 1066 PetscFunctionBegin; 1067 PetscValidPointer(graph,1); 1068 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1069 /* raise an error if already allocated */ 1070 if (graph->nvtxs) { 1071 SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1072 } 1073 /* set number of vertices */ 1074 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1075 graph->l2gmap = l2gmap; 1076 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1077 graph->nvtxs = n; 1078 /* allocate used space */ 1079 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1080 ierr = PetscMalloc7(graph->nvtxs,&graph->count, 1081 graph->nvtxs,&graph->neighbours_set, 1082 graph->nvtxs,&graph->subset, 1083 graph->nvtxs,&graph->which_dof, 1084 graph->nvtxs+1,&graph->cptr, 1085 graph->nvtxs,&graph->queue, 1086 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1087 /* zeroes memory */ 1088 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1089 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1090 /* use -1 as a default value for which_dof array */ 1091 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1092 ierr = PetscMemzero(graph->cptr,(graph->nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr); 1093 ierr = PetscMemzero(graph->queue,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1094 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1095 /* zeroes first pointer to neighbour set */ 1096 if (graph->nvtxs) { 1097 graph->neighbours_set[0] = 0; 1098 } 1099 /* zeroes workspace for values of ncc */ 1100 graph->subset_ncc = 0; 1101 graph->subset_ref_node = 0; 1102 PetscFunctionReturn(0); 1103 } 1104 1105 #undef __FUNCT__ 1106 #define __FUNCT__ "PCBDDCGraphDestroy" 1107 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1108 { 1109 PetscErrorCode ierr; 1110 1111 PetscFunctionBegin; 1112 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1113 ierr = PetscFree(*graph);CHKERRQ(ierr); 1114 PetscFunctionReturn(0); 1115 } 1116 1117 #undef __FUNCT__ 1118 #define __FUNCT__ "PCBDDCGraphCreate" 1119 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1120 { 1121 PCBDDCGraph new_graph; 1122 PetscErrorCode ierr; 1123 1124 PetscFunctionBegin; 1125 ierr = PetscMalloc(sizeof(*new_graph),&new_graph);CHKERRQ(ierr); 1126 /* local to global mapping of dofs */ 1127 new_graph->l2gmap = 0; 1128 /* vertex size */ 1129 new_graph->nvtxs = 0; 1130 new_graph->n_subsets = 0; 1131 new_graph->custom_minimal_size = 1; 1132 /* zeroes ponters */ 1133 new_graph->mirrors = 0; 1134 new_graph->mirrors_set = 0; 1135 new_graph->neighbours_set = 0; 1136 new_graph->subset = 0; 1137 new_graph->which_dof = 0; 1138 new_graph->special_dof = 0; 1139 new_graph->cptr = 0; 1140 new_graph->queue = 0; 1141 new_graph->count = 0; 1142 new_graph->subset_ncc = 0; 1143 new_graph->subset_ref_node = 0; 1144 new_graph->touched = 0; 1145 /* zeroes pointers to csr graph of local nodes connectivity (optional data) */ 1146 new_graph->nvtxs_csr = 0; 1147 new_graph->xadj = 0; 1148 new_graph->adjncy = 0; 1149 *graph = new_graph; 1150 PetscFunctionReturn(0); 1151 } 1152