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