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 PetscErrorCode PCBDDCGraphGetDirichletDofsB(PCBDDCGraph graph, IS* dirdofs) 6 { 7 PetscErrorCode ierr; 8 9 PetscFunctionBegin; 10 if (graph->dirdofsB) { 11 ierr = PetscObjectReference((PetscObject)graph->dirdofsB);CHKERRQ(ierr); 12 } else if (graph->has_dirichlet) { 13 PetscInt i,size; 14 PetscInt *dirdofs_idxs; 15 16 size = 0; 17 for (i=0;i<graph->nvtxs;i++) { 18 if (graph->count[i] && graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 19 } 20 21 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 22 size = 0; 23 for (i=0;i<graph->nvtxs;i++) { 24 if (graph->count[i] && graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 25 } 26 ierr = ISCreateGeneral(PETSC_COMM_SELF,size,dirdofs_idxs,PETSC_OWN_POINTER,&graph->dirdofsB);CHKERRQ(ierr); 27 ierr = PetscObjectReference((PetscObject)graph->dirdofsB);CHKERRQ(ierr); 28 } 29 *dirdofs = graph->dirdofsB; 30 PetscFunctionReturn(0); 31 } 32 33 PetscErrorCode PCBDDCGraphGetDirichletDofs(PCBDDCGraph graph, IS* dirdofs) 34 { 35 PetscErrorCode ierr; 36 37 PetscFunctionBegin; 38 if (graph->dirdofs) { 39 ierr = PetscObjectReference((PetscObject)graph->dirdofs);CHKERRQ(ierr); 40 } else if (graph->has_dirichlet) { 41 PetscInt i,size; 42 PetscInt *dirdofs_idxs; 43 44 size = 0; 45 for (i=0;i<graph->nvtxs;i++) { 46 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) size++; 47 } 48 49 ierr = PetscMalloc1(size,&dirdofs_idxs);CHKERRQ(ierr); 50 size = 0; 51 for (i=0;i<graph->nvtxs;i++) { 52 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK) dirdofs_idxs[size++] = i; 53 } 54 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)graph->l2gmap),size,dirdofs_idxs,PETSC_OWN_POINTER,&graph->dirdofs);CHKERRQ(ierr); 55 ierr = PetscObjectReference((PetscObject)graph->dirdofs);CHKERRQ(ierr); 56 } 57 *dirdofs = graph->dirdofs; 58 PetscFunctionReturn(0); 59 } 60 61 PetscErrorCode PCBDDCGraphASCIIView(PCBDDCGraph graph, PetscInt verbosity_level, PetscViewer viewer) 62 { 63 PetscInt i,j,tabs; 64 PetscInt* queue_in_global_numbering; 65 PetscErrorCode ierr; 66 67 PetscFunctionBegin; 68 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 69 ierr = PetscViewerASCIIGetTab(viewer,&tabs);CHKERRQ(ierr); 70 ierr = PetscViewerASCIIPrintf(viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 71 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 72 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Local BDDC graph for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 73 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Number of vertices %d\n",graph->nvtxs);CHKERRQ(ierr); 74 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Custom minimal size %d\n",graph->custom_minimal_size);CHKERRQ(ierr); 75 if (graph->maxcount != PETSC_MAX_INT) { 76 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Max count %d\n",graph->maxcount);CHKERRQ(ierr); 77 } 78 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Topological two dim? %d (set %d)\n",graph->twodim,graph->twodimset);CHKERRQ(ierr); 79 if (verbosity_level > 2) { 80 for (i=0;i<graph->nvtxs;i++) { 81 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"%d:\n",i);CHKERRQ(ierr); 82 ierr = PetscViewerASCIISynchronizedPrintf(viewer," which_dof: %d\n",graph->which_dof[i]);CHKERRQ(ierr); 83 ierr = PetscViewerASCIISynchronizedPrintf(viewer," special_dof: %d\n",graph->special_dof[i]);CHKERRQ(ierr); 84 ierr = PetscViewerASCIISynchronizedPrintf(viewer," neighbours: %d\n",graph->count[i]);CHKERRQ(ierr); 85 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 86 if (graph->count[i]) { 87 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of neighbours:");CHKERRQ(ierr); 88 for (j=0;j<graph->count[i];j++) { 89 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[i][j]);CHKERRQ(ierr); 90 } 91 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 92 } 93 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 94 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 95 if (graph->mirrors) { 96 ierr = PetscViewerASCIISynchronizedPrintf(viewer," mirrors: %d\n",graph->mirrors[i]);CHKERRQ(ierr); 97 if (graph->mirrors[i]) { 98 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 99 ierr = PetscViewerASCIISynchronizedPrintf(viewer," set of mirrors:");CHKERRQ(ierr); 100 for (j=0;j<graph->mirrors[i];j++) { 101 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->mirrors_set[i][j]);CHKERRQ(ierr); 102 } 103 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 104 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 105 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 106 } 107 } 108 if (verbosity_level > 3) { 109 if (graph->xadj) { 110 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local adj list:");CHKERRQ(ierr); 111 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 112 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 113 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->adjncy[j]);CHKERRQ(ierr); 114 } 115 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 116 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 117 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 118 } else { 119 ierr = PetscViewerASCIISynchronizedPrintf(viewer," no adj info\n");CHKERRQ(ierr); 120 } 121 } 122 if (graph->n_local_subs) { 123 ierr = PetscViewerASCIISynchronizedPrintf(viewer," local sub id: %d\n",graph->local_subs[i]);CHKERRQ(ierr); 124 } 125 ierr = PetscViewerASCIISynchronizedPrintf(viewer," interface subset id: %d\n",graph->subset[i]);CHKERRQ(ierr); 126 if (graph->subset[i] && graph->subset_ncc) { 127 ierr = PetscViewerASCIISynchronizedPrintf(viewer," ncc for subset: %d\n",graph->subset_ncc[graph->subset[i]-1]);CHKERRQ(ierr); 128 } 129 } 130 } 131 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"Total number of connected components %d\n",graph->ncc);CHKERRQ(ierr); 132 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_in_global_numbering);CHKERRQ(ierr); 133 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_in_global_numbering);CHKERRQ(ierr); 134 for (i=0;i<graph->ncc;i++) { 135 PetscInt node_num=graph->queue[graph->cptr[i]]; 136 PetscBool printcc = PETSC_FALSE; 137 ierr = PetscViewerASCIISynchronizedPrintf(viewer," cc %d (size %d, fid %d, neighs:",i,graph->cptr[i+1]-graph->cptr[i],graph->which_dof[node_num]);CHKERRQ(ierr); 138 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 139 for (j=0;j<graph->count[node_num];j++) { 140 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d",graph->neighbours_set[node_num][j]);CHKERRQ(ierr); 141 } 142 if (verbosity_level > 1) { 143 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"):");CHKERRQ(ierr); 144 if (verbosity_level > 2 || graph->twodim || graph->count[node_num] > 1 || (graph->count[node_num] == 1 && graph->special_dof[node_num] == PCBDDCGRAPH_NEUMANN_MARK)) { 145 printcc = PETSC_TRUE; 146 } 147 if (printcc) { 148 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 149 ierr = PetscViewerASCIISynchronizedPrintf(viewer," %d (%d)",graph->queue[j],queue_in_global_numbering[j]);CHKERRQ(ierr); 150 } 151 } 152 } else { 153 ierr = PetscViewerASCIISynchronizedPrintf(viewer,")");CHKERRQ(ierr); 154 } 155 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"\n");CHKERRQ(ierr); 156 ierr = PetscViewerASCIISetTab(viewer,tabs);CHKERRQ(ierr); 157 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 158 } 159 ierr = PetscFree(queue_in_global_numbering);CHKERRQ(ierr); 160 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 161 PetscFunctionReturn(0); 162 } 163 164 PetscErrorCode PCBDDCGraphRestoreCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 165 { 166 PetscInt i; 167 PetscErrorCode ierr; 168 169 PetscFunctionBegin; 170 if (n_faces) { 171 if (FacesIS) { 172 for (i=0;i<*n_faces;i++) { 173 ierr = ISDestroy(&((*FacesIS)[i]));CHKERRQ(ierr); 174 } 175 ierr = PetscFree(*FacesIS);CHKERRQ(ierr); 176 } 177 *n_faces = 0; 178 } 179 if (n_edges) { 180 if (EdgesIS) { 181 for (i=0;i<*n_edges;i++) { 182 ierr = ISDestroy(&((*EdgesIS)[i]));CHKERRQ(ierr); 183 } 184 ierr = PetscFree(*EdgesIS);CHKERRQ(ierr); 185 } 186 *n_edges = 0; 187 } 188 if (VerticesIS) { 189 ierr = ISDestroy(VerticesIS);CHKERRQ(ierr); 190 } 191 PetscFunctionReturn(0); 192 } 193 194 PetscErrorCode PCBDDCGraphGetCandidatesIS(PCBDDCGraph graph, PetscInt *n_faces, IS *FacesIS[], PetscInt *n_edges, IS *EdgesIS[], IS *VerticesIS) 195 { 196 IS *ISForFaces,*ISForEdges,ISForVertices; 197 PetscInt i,nfc,nec,nvc,*idx,*mark; 198 PetscErrorCode ierr; 199 200 PetscFunctionBegin; 201 ierr = PetscCalloc1(graph->ncc,&mark);CHKERRQ(ierr); 202 /* loop on ccs to evalute number of faces, edges and vertices */ 203 nfc = 0; 204 nec = 0; 205 nvc = 0; 206 for (i=0;i<graph->ncc;i++) { 207 PetscInt repdof = graph->queue[graph->cptr[i]]; 208 if (graph->cptr[i+1]-graph->cptr[i] > graph->custom_minimal_size && graph->count[repdof] < graph->maxcount) { 209 if (!graph->twodim && graph->count[repdof] == 1 && graph->special_dof[repdof] != PCBDDCGRAPH_NEUMANN_MARK) { 210 nfc++; 211 mark[i] = 2; 212 } else { 213 nec++; 214 mark[i] = 1; 215 } 216 } else { 217 nvc += graph->cptr[i+1]-graph->cptr[i]; 218 } 219 } 220 221 /* allocate IS arrays for faces, edges. Vertices need a single index set. */ 222 if (FacesIS) { 223 ierr = PetscMalloc1(nfc,&ISForFaces);CHKERRQ(ierr); 224 } 225 if (EdgesIS) { 226 ierr = PetscMalloc1(nec,&ISForEdges);CHKERRQ(ierr); 227 } 228 if (VerticesIS) { 229 ierr = PetscMalloc1(nvc,&idx);CHKERRQ(ierr); 230 } 231 232 /* loop on ccs to compute index sets for faces and edges */ 233 if (!graph->queue_sorted) { 234 PetscInt *queue_global; 235 236 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 237 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 238 for (i=0;i<graph->ncc;i++) { 239 ierr = PetscSortIntWithArray(graph->cptr[i+1]-graph->cptr[i],&queue_global[graph->cptr[i]],&graph->queue[graph->cptr[i]]);CHKERRQ(ierr); 240 } 241 ierr = PetscFree(queue_global);CHKERRQ(ierr); 242 graph->queue_sorted = PETSC_TRUE; 243 } 244 nfc = 0; 245 nec = 0; 246 for (i=0;i<graph->ncc;i++) { 247 if (mark[i] == 2) { 248 if (FacesIS) { 249 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_USE_POINTER,&ISForFaces[nfc]);CHKERRQ(ierr); 250 } 251 nfc++; 252 } else if (mark[i] == 1) { 253 if (EdgesIS) { 254 ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],&graph->queue[graph->cptr[i]],PETSC_USE_POINTER,&ISForEdges[nec]);CHKERRQ(ierr); 255 } 256 nec++; 257 } 258 } 259 260 /* index set for vertices */ 261 if (VerticesIS) { 262 nvc = 0; 263 for (i=0;i<graph->ncc;i++) { 264 if (!mark[i]) { 265 PetscInt j; 266 267 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 268 idx[nvc]=graph->queue[j]; 269 nvc++; 270 } 271 } 272 } 273 /* sort vertex set (by local ordering) */ 274 ierr = PetscSortInt(nvc,idx);CHKERRQ(ierr); 275 ierr = ISCreateGeneral(PETSC_COMM_SELF,nvc,idx,PETSC_OWN_POINTER,&ISForVertices);CHKERRQ(ierr); 276 } 277 ierr = PetscFree(mark);CHKERRQ(ierr); 278 279 /* get back info */ 280 if (n_faces) *n_faces = nfc; 281 if (FacesIS) *FacesIS = ISForFaces; 282 if (n_edges) *n_edges = nec; 283 if (EdgesIS) *EdgesIS = ISForEdges; 284 if (VerticesIS) *VerticesIS = ISForVertices; 285 PetscFunctionReturn(0); 286 } 287 288 PetscErrorCode PCBDDCGraphComputeConnectedComponents(PCBDDCGraph graph) 289 { 290 PetscBool adapt_interface_reduced; 291 MPI_Comm interface_comm; 292 PetscMPIInt size; 293 PetscInt i; 294 PetscBT cornerp; 295 PetscErrorCode ierr; 296 297 PetscFunctionBegin; 298 /* compute connected components locally */ 299 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&interface_comm);CHKERRQ(ierr); 300 ierr = PCBDDCGraphComputeConnectedComponentsLocal(graph);CHKERRQ(ierr); 301 302 cornerp = NULL; 303 if (graph->active_coords) { /* face based corner selection */ 304 PetscReal *wdist; 305 PetscInt n_neigh,*neigh,*n_shared,**shared; 306 PetscInt maxc, ns; 307 308 ierr = PetscBTCreate(graph->nvtxs,&cornerp);CHKERRQ(ierr); 309 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 310 for (ns = 1, maxc = 0; ns < n_neigh; ns++) maxc = PetscMax(maxc,n_shared[ns]); 311 ierr = PetscMalloc1(maxc*graph->cdim,&wdist);CHKERRQ(ierr); 312 313 for (ns = 1; ns < n_neigh; ns++) { /* first proc is self */ 314 PetscReal *anchor,mdist; 315 PetscInt j,k,d,cdim = graph->cdim; 316 PetscInt point1,point2,point3; 317 /* 318 PetscBool isface = PETSC_FALSE; 319 320 for (i=0;i<n_shared[ns];i++) { 321 if (graph->count[shared[ns][i]] == 1) { 322 isface = PETSC_TRUE; 323 break; 324 } 325 } 326 if (!isface) continue; 327 */ 328 /* import coordinates on shared interface */ 329 for (j=0,k=0;j<n_shared[ns];j++) 330 for (d=0;d<cdim;d++) 331 wdist[k++] = graph->coords[shared[ns][j]*cdim+d]; 332 333 /* the dofs are sorted by global numbering, so each rank start from the same id and will detect the same corners */ 334 anchor = wdist; 335 336 /* find the farthest point from the starting one */ 337 mdist = -1.0; 338 for (j=0,point1=0;j<n_shared[ns];j++) { 339 PetscReal dist = 0.0; 340 341 for (d=0;d<cdim;d++) dist += (wdist[j*cdim+d]-anchor[d])*(wdist[j*cdim+d]-anchor[d]); 342 if (dist > mdist) { mdist = dist; point1 = j; } 343 } 344 345 /* find the farthest point from point1 */ 346 anchor = wdist + point1*cdim; 347 mdist = -1.0; 348 for (j=0,point2=0;j<n_shared[ns];j++) { 349 PetscReal dist = 0.0; 350 351 for (d=0;d<cdim;d++) dist += (wdist[j*cdim+d]-anchor[d])*(wdist[j*cdim+d]-anchor[d]); 352 if (dist > mdist) { mdist = dist; point2 = j; } 353 } 354 355 /* find the third point maximizing the triangle area */ 356 point3 = point2; 357 if (cdim > 2) { 358 PetscReal a = 0.0; 359 360 for (d=0;d<cdim;d++) a += (wdist[point1*cdim+d]-wdist[point2*cdim+d])*(wdist[point1*cdim+d]-wdist[point2*cdim+d]); 361 mdist = -1.0; 362 for (j=0,point3=0;j<n_shared[ns];j++) { 363 PetscReal area,b = 0.0, c = 0.0; 364 365 for (d=0;d<cdim;d++) { 366 b += (wdist[point1*cdim+d]-wdist[j*cdim+d])*(wdist[point1*cdim+d]-wdist[j*cdim+d]); 367 c += (wdist[point2*cdim+d]-wdist[j*cdim+d])*(wdist[point2*cdim+d]-wdist[j*cdim+d]); 368 } 369 area = (a+b+c)*(-a+b+c)*(a-b+c)*(a+b-c); /* Heron's formula without divisions by 2 */ 370 if (area > mdist) { mdist = area; point3 = j; } 371 } 372 } 373 374 /* all dofs having the same coordinates will be primal */ 375 for (j=0;j<n_shared[ns];j++) { 376 PetscBool same[3] = {PETSC_TRUE,PETSC_TRUE,PETSC_TRUE}; 377 378 for (d=0;d<cdim;d++) { 379 same[0] = (PetscBool)(same[0] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point1*cdim+d]) < PETSC_SMALL)); 380 same[1] = (PetscBool)(same[1] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point2*cdim+d]) < PETSC_SMALL)); 381 same[2] = (PetscBool)(same[2] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point3*cdim+d]) < PETSC_SMALL)); 382 } 383 if (same[0] || same[1] || same[2]) { 384 ierr = PetscBTSet(cornerp,shared[ns][j]);CHKERRQ(ierr); 385 } 386 } 387 } 388 ierr = PetscFree(wdist);CHKERRQ(ierr); 389 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 390 } 391 392 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 393 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 394 adapt_interface_reduced = PETSC_FALSE; 395 if (size > 1) { 396 PetscInt i; 397 PetscBool adapt_interface = cornerp ? PETSC_TRUE : PETSC_FALSE; 398 for (i=0;i<graph->n_subsets && !adapt_interface;i++) { 399 /* We are not sure that on a given subset of the local interface, 400 with two connected components, the latters be the same among sharing subdomains */ 401 if (graph->subset_ncc[i] > 1) adapt_interface = PETSC_TRUE; 402 } 403 ierr = MPIU_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 404 } 405 406 if (graph->n_subsets && adapt_interface_reduced) { 407 PetscBT subset_cc_adapt; 408 MPI_Request *send_requests,*recv_requests; 409 PetscInt *send_buffer,*recv_buffer; 410 PetscInt sum_requests,start_of_recv,start_of_send; 411 PetscInt *cum_recv_counts; 412 PetscInt *labels; 413 PetscInt ncc,cum_queue,mss,mns,j,k,s; 414 PetscInt **refine_buffer=NULL,*private_labels = NULL; 415 PetscBool *subset_has_corn,*recv_buffer_bool; 416 417 ierr = PetscCalloc1(graph->n_subsets,&subset_has_corn);CHKERRQ(ierr); 418 if (cornerp) { 419 for (i=0;i<graph->n_subsets;i++) { 420 for (j=0;j<graph->subset_size[i];j++) { 421 if (PetscBTLookup(cornerp,graph->subset_idxs[i][j])) { 422 subset_has_corn[i] = PETSC_TRUE; 423 break; 424 } 425 } 426 } 427 } 428 ierr = PetscMalloc1(graph->nvtxs,&labels);CHKERRQ(ierr); 429 ierr = PetscMemzero(labels,graph->nvtxs*sizeof(*labels));CHKERRQ(ierr); 430 for (i=0,k=0;i<graph->ncc;i++) { 431 PetscInt s = 1; 432 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 433 if (cornerp && PetscBTLookup(cornerp,graph->queue[j])) { 434 labels[graph->queue[j]] = k+s; 435 s += 1; 436 } else { 437 labels[graph->queue[j]] = k; 438 } 439 } 440 k += s; 441 } 442 443 /* allocate some space */ 444 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 445 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 446 447 /* first count how many neighbours per connected component I will receive from */ 448 cum_recv_counts[0] = 0; 449 for (i=0;i<graph->n_subsets;i++) cum_recv_counts[i+1] = cum_recv_counts[i]+graph->count[graph->subset_idxs[i][0]]; 450 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_buffer_bool);CHKERRQ(ierr); 451 ierr = PetscMalloc2(cum_recv_counts[graph->n_subsets],&send_requests,cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 452 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 453 send_requests[i] = MPI_REQUEST_NULL; 454 recv_requests[i] = MPI_REQUEST_NULL; 455 } 456 457 /* exchange with my neighbours the number of my connected components on the subset of interface */ 458 sum_requests = 0; 459 for (i=0;i<graph->n_subsets;i++) { 460 PetscMPIInt neigh,tag; 461 PetscInt count,*neighs; 462 463 count = graph->count[graph->subset_idxs[i][0]]; 464 neighs = graph->neighbours_set[graph->subset_idxs[i][0]]; 465 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 466 for (k=0;k<count;k++) { 467 PetscBool ss = (PetscBool)(graph->subset_ncc[i] > 1 || subset_has_corn[i]); 468 469 ierr = PetscMPIIntCast(neighs[k],&neigh);CHKERRQ(ierr); 470 ierr = MPI_Isend(&ss,1,MPIU_BOOL,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 471 ierr = MPI_Irecv(&recv_buffer_bool[sum_requests],1,MPIU_BOOL,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 472 sum_requests++; 473 } 474 } 475 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 476 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 477 478 /* determine the subsets I have to adapt (those having more than 1 cc) */ 479 ierr = PetscBTCreate(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 480 ierr = PetscBTMemzero(graph->n_subsets,subset_cc_adapt);CHKERRQ(ierr); 481 for (i=0;i<graph->n_subsets;i++) { 482 if (graph->subset_ncc[i] > 1 || subset_has_corn[i]) { 483 ierr = PetscBTSet(subset_cc_adapt,i);CHKERRQ(ierr); 484 continue; 485 } 486 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 487 if (recv_buffer_bool[j]) { 488 ierr = PetscBTSet(subset_cc_adapt,i);CHKERRQ(ierr); 489 break; 490 } 491 } 492 } 493 ierr = PetscFree(recv_buffer_bool);CHKERRQ(ierr); 494 ierr = PetscFree(subset_has_corn);CHKERRQ(ierr); 495 496 /* determine send/recv buffers sizes */ 497 j = 0; 498 mss = 0; 499 for (i=0;i<graph->n_subsets;i++) { 500 if (PetscBTLookup(subset_cc_adapt,i)) { 501 j += graph->subset_size[i]; 502 mss = PetscMax(graph->subset_size[i],mss); 503 } 504 } 505 k = 0; 506 mns = 0; 507 for (i=0;i<graph->n_subsets;i++) { 508 if (PetscBTLookup(subset_cc_adapt,i)) { 509 k += (cum_recv_counts[i+1]-cum_recv_counts[i])*graph->subset_size[i]; 510 mns = PetscMax(cum_recv_counts[i+1]-cum_recv_counts[i],mns); 511 } 512 } 513 ierr = PetscMalloc2(j,&send_buffer,k,&recv_buffer);CHKERRQ(ierr); 514 515 /* fill send buffer (order matters: subset_idxs ordered by global ordering) */ 516 j = 0; 517 for (i=0;i<graph->n_subsets;i++) 518 if (PetscBTLookup(subset_cc_adapt,i)) 519 for (k=0;k<graph->subset_size[i];k++) 520 send_buffer[j++] = labels[graph->subset_idxs[i][k]]; 521 522 /* now exchange the data */ 523 start_of_recv = 0; 524 start_of_send = 0; 525 sum_requests = 0; 526 for (i=0;i<graph->n_subsets;i++) { 527 if (PetscBTLookup(subset_cc_adapt,i)) { 528 PetscMPIInt neigh,tag; 529 PetscInt size_of_send = graph->subset_size[i]; 530 531 j = graph->subset_idxs[i][0]; 532 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 533 for (k=0;k<graph->count[j];k++) { 534 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 535 ierr = MPI_Isend(&send_buffer[start_of_send],size_of_send,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 536 ierr = MPI_Irecv(&recv_buffer[start_of_recv],size_of_send,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 537 start_of_recv += size_of_send; 538 sum_requests++; 539 } 540 start_of_send += size_of_send; 541 } 542 } 543 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 544 545 /* refine connected components */ 546 start_of_recv = 0; 547 /* allocate some temporary space */ 548 if (mss) { 549 ierr = PetscMalloc1(mss,&refine_buffer);CHKERRQ(ierr); 550 ierr = PetscMalloc2(mss*(mns+1),&refine_buffer[0],mss,&private_labels);CHKERRQ(ierr); 551 } 552 ncc = 0; 553 cum_queue = 0; 554 graph->cptr[0] = 0; 555 for (i=0;i<graph->n_subsets;i++) { 556 if (PetscBTLookup(subset_cc_adapt,i)) { 557 PetscInt subset_counter = 0; 558 PetscInt sharingprocs = cum_recv_counts[i+1]-cum_recv_counts[i]+1; /* count myself */ 559 PetscInt buffer_size = graph->subset_size[i]; 560 561 /* compute pointers */ 562 for (j=1;j<buffer_size;j++) refine_buffer[j] = refine_buffer[j-1] + sharingprocs; 563 /* analyze contributions from subdomains that share the i-th subset 564 The structure of refine_buffer is suitable to find intersections of ccs among sharingprocs. 565 supposing the current subset is shared by 3 processes and has dimension 5 with global dofs 0,1,2,3,4 (local 0,4,3,1,2) 566 sharing procs connected components: 567 neigh 0: [0 1 4], [2 3], labels [4,7] (2 connected components) 568 neigh 1: [0 1], [2 3 4], labels [3 2] (2 connected components) 569 neigh 2: [0 4], [1], [2 3], labels [1 5 6] (3 connected components) 570 refine_buffer will be filled as: 571 [ 4, 3, 1; 572 4, 2, 1; 573 7, 2, 6; 574 4, 3, 5; 575 7, 2, 6; ]; 576 The connected components in local ordering are [0], [1], [2 3], [4] */ 577 /* fill temp_buffer */ 578 for (k=0;k<buffer_size;k++) refine_buffer[k][0] = labels[graph->subset_idxs[i][k]]; 579 for (j=0;j<sharingprocs-1;j++) { 580 for (k=0;k<buffer_size;k++) refine_buffer[k][j+1] = recv_buffer[start_of_recv+k]; 581 start_of_recv += buffer_size; 582 } 583 ierr = PetscMemzero(private_labels,buffer_size*sizeof(PetscInt));CHKERRQ(ierr); 584 for (j=0;j<buffer_size;j++) { 585 if (!private_labels[j]) { /* found a new cc */ 586 PetscBool same_set; 587 588 graph->cptr[ncc] = cum_queue; 589 ncc++; 590 subset_counter++; 591 private_labels[j] = subset_counter; 592 graph->queue[cum_queue++] = graph->subset_idxs[i][j]; 593 for (k=j+1;k<buffer_size;k++) { /* check for other nodes in new cc */ 594 same_set = PETSC_TRUE; 595 for (s=0;s<sharingprocs;s++) { 596 if (refine_buffer[j][s] != refine_buffer[k][s]) { 597 same_set = PETSC_FALSE; 598 break; 599 } 600 } 601 if (same_set) { 602 private_labels[k] = subset_counter; 603 graph->queue[cum_queue++] = graph->subset_idxs[i][k]; 604 } 605 } 606 } 607 } 608 graph->cptr[ncc] = cum_queue; 609 graph->subset_ncc[i] = subset_counter; 610 graph->queue_sorted = PETSC_FALSE; 611 } else { /* this subset does not need to be adapted */ 612 ierr = PetscMemcpy(graph->queue+cum_queue,graph->subset_idxs[i],graph->subset_size[i]*sizeof(PetscInt));CHKERRQ(ierr); 613 ncc++; 614 cum_queue += graph->subset_size[i]; 615 graph->cptr[ncc] = cum_queue; 616 } 617 } 618 graph->cptr[ncc] = cum_queue; 619 graph->ncc = ncc; 620 if (mss) { 621 ierr = PetscFree2(refine_buffer[0],private_labels);CHKERRQ(ierr); 622 ierr = PetscFree(refine_buffer);CHKERRQ(ierr); 623 } 624 ierr = PetscFree(labels);CHKERRQ(ierr); 625 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 626 ierr = PetscFree2(send_requests,recv_requests);CHKERRQ(ierr); 627 ierr = PetscFree2(send_buffer,recv_buffer);CHKERRQ(ierr); 628 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 629 ierr = PetscBTDestroy(&subset_cc_adapt);CHKERRQ(ierr); 630 } 631 ierr = PetscBTDestroy(&cornerp);CHKERRQ(ierr); 632 633 /* Determine if we are in 2D or 3D */ 634 if (!graph->twodimset) { 635 PetscBool twodim = PETSC_TRUE; 636 for (i=0;i<graph->ncc;i++) { 637 PetscInt repdof = graph->queue[graph->cptr[i]]; 638 PetscInt ccsize = graph->cptr[i+1]-graph->cptr[i]; 639 if (graph->count[repdof] > 1 && ccsize > graph->custom_minimal_size) { 640 twodim = PETSC_FALSE; 641 break; 642 } 643 } 644 ierr = MPIU_Allreduce(&twodim,&graph->twodim,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 645 graph->twodimset = PETSC_TRUE; 646 } 647 PetscFunctionReturn(0); 648 } 649 650 651 PETSC_STATIC_INLINE PetscErrorCode PCBDDCGraphComputeCC_Private(PCBDDCGraph graph,PetscInt pid,PetscInt* queue_tip,PetscInt n_prev,PetscInt* n_added) 652 { 653 PetscInt i,j,n; 654 PetscInt *xadj = graph->xadj,*adjncy = graph->adjncy; 655 PetscBT touched = graph->touched; 656 PetscBool havecsr = (PetscBool)(!!xadj); 657 PetscBool havesubs = (PetscBool)(!!graph->n_local_subs); 658 PetscErrorCode ierr; 659 660 PetscFunctionBegin; 661 n = 0; 662 if (havecsr && !havesubs) { 663 for (i=-n_prev;i<0;i++) { 664 PetscInt start_dof = queue_tip[i]; 665 /* we assume that if a dof has a size 1 adjacency list and the corresponding entry is negative, it is connected to all dofs */ 666 if (xadj[start_dof+1]-xadj[start_dof] == 1 && adjncy[xadj[start_dof]] < 0) { 667 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 668 PetscInt dof = graph->subset_idxs[pid-1][j]; 669 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 670 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 671 queue_tip[n] = dof; 672 n++; 673 } 674 } 675 } else { 676 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 677 PetscInt dof = adjncy[j]; 678 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 679 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 680 queue_tip[n] = dof; 681 n++; 682 } 683 } 684 } 685 } 686 } else if (havecsr && havesubs) { 687 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 688 for (i=-n_prev;i<0;i++) { 689 PetscInt start_dof = queue_tip[i]; 690 /* we assume that if a dof has a size 1 adjacency list and the corresponding entry is negative, it is connected to all dofs belonging to the local sub */ 691 if (xadj[start_dof+1]-xadj[start_dof] == 1 && adjncy[xadj[start_dof]] < 0) { 692 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 693 PetscInt dof = graph->subset_idxs[pid-1][j]; 694 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 695 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 696 queue_tip[n] = dof; 697 n++; 698 } 699 } 700 } else { 701 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 702 PetscInt dof = adjncy[j]; 703 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 704 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 705 queue_tip[n] = dof; 706 n++; 707 } 708 } 709 } 710 } 711 } else if (havesubs) { /* sub info only */ 712 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 713 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 714 PetscInt dof = graph->subset_idxs[pid-1][j]; 715 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 716 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 717 queue_tip[n] = dof; 718 n++; 719 } 720 } 721 } else { 722 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 723 PetscInt dof = graph->subset_idxs[pid-1][j]; 724 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 725 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 726 queue_tip[n] = dof; 727 n++; 728 } 729 } 730 } 731 *n_added = n; 732 PetscFunctionReturn(0); 733 } 734 735 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 736 { 737 PetscInt ncc,cum_queue,n; 738 PetscMPIInt commsize; 739 PetscErrorCode ierr; 740 741 PetscFunctionBegin; 742 if (!graph->setupcalled) SETERRQ(PetscObjectComm((PetscObject)graph->l2gmap),PETSC_ERR_ORDER,"PCBDDCGraphSetUp should be called first"); 743 /* quiet return if there isn't any local info */ 744 if (!graph->xadj && !graph->n_local_subs) { 745 PetscFunctionReturn(0); 746 } 747 748 /* reset any previous search of connected components */ 749 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 750 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&commsize);CHKERRQ(ierr); 751 if (commsize > graph->commsizelimit) { 752 PetscInt i; 753 for (i=0;i<graph->nvtxs;i++) { 754 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 755 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 756 } 757 } 758 } 759 760 /* begin search for connected components */ 761 cum_queue = 0; 762 ncc = 0; 763 for (n=0;n<graph->n_subsets;n++) { 764 PetscInt pid = n+1; /* partition labeled by 0 is discarded */ 765 PetscInt found = 0,prev = 0,first = 0,ncc_pid = 0; 766 while (found != graph->subset_size[n]) { 767 PetscInt added = 0; 768 if (!prev) { /* search for new starting dof */ 769 while (PetscBTLookup(graph->touched,graph->subset_idxs[n][first])) first++; 770 ierr = PetscBTSet(graph->touched,graph->subset_idxs[n][first]);CHKERRQ(ierr); 771 graph->queue[cum_queue] = graph->subset_idxs[n][first]; 772 graph->cptr[ncc] = cum_queue; 773 prev = 1; 774 cum_queue++; 775 found++; 776 ncc_pid++; 777 ncc++; 778 } 779 ierr = PCBDDCGraphComputeCC_Private(graph,pid,graph->queue + cum_queue,prev,&added);CHKERRQ(ierr); 780 if (!added) { 781 graph->subset_ncc[n] = ncc_pid; 782 graph->cptr[ncc] = cum_queue; 783 } 784 prev = added; 785 found += added; 786 cum_queue += added; 787 if (added && found == graph->subset_size[n]) { 788 graph->subset_ncc[n] = ncc_pid; 789 graph->cptr[ncc] = cum_queue; 790 } 791 } 792 } 793 graph->ncc = ncc; 794 graph->queue_sorted = PETSC_FALSE; 795 PetscFunctionReturn(0); 796 } 797 798 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 799 { 800 IS subset,subset_n; 801 MPI_Comm comm; 802 const PetscInt *is_indices; 803 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global; 804 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 805 PetscMPIInt commsize; 806 PetscBool same_set,mirrors_found; 807 PetscErrorCode ierr; 808 809 PetscFunctionBegin; 810 PetscValidLogicalCollectiveInt(graph->l2gmap,custom_minimal_size,2); 811 if (neumann_is) { 812 PetscValidHeaderSpecific(neumann_is,IS_CLASSID,3); 813 PetscCheckSameComm(graph->l2gmap,1,neumann_is,3); 814 } 815 graph->has_dirichlet = PETSC_FALSE; 816 if (dirichlet_is) { 817 PetscValidHeaderSpecific(dirichlet_is,IS_CLASSID,4); 818 PetscCheckSameComm(graph->l2gmap,1,dirichlet_is,4); 819 graph->has_dirichlet = PETSC_TRUE; 820 } 821 PetscValidLogicalCollectiveInt(graph->l2gmap,n_ISForDofs,5); 822 for (i=0;i<n_ISForDofs;i++) { 823 PetscValidHeaderSpecific(ISForDofs[i],IS_CLASSID,6); 824 PetscCheckSameComm(graph->l2gmap,1,ISForDofs[i],6); 825 } 826 if (custom_primal_vertices) { 827 PetscValidHeaderSpecific(custom_primal_vertices,IS_CLASSID,7); 828 PetscCheckSameComm(graph->l2gmap,1,custom_primal_vertices,7); 829 } 830 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 831 ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 832 833 /* custom_minimal_size */ 834 graph->custom_minimal_size = custom_minimal_size; 835 /* get info l2gmap and allocate work vectors */ 836 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 837 /* check if we have any local periodic nodes (periodic BCs) */ 838 mirrors_found = PETSC_FALSE; 839 if (graph->nvtxs && n_neigh) { 840 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 841 for (i=0; i<n_shared[0]; i++) { 842 if (graph->count[shared[0][i]] > 1) { 843 mirrors_found = PETSC_TRUE; 844 break; 845 } 846 } 847 } 848 /* compute local mirrors (if any) */ 849 if (mirrors_found) { 850 IS to,from; 851 PetscInt *local_indices,*global_indices; 852 853 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 854 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 855 /* get arrays of local and global indices */ 856 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 857 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 858 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 859 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 860 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 861 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 862 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 863 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 864 /* allocate space for mirrors */ 865 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 866 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 867 graph->mirrors_set[0] = 0; 868 869 k=0; 870 for (i=0;i<n_shared[0];i++) { 871 j=shared[0][i]; 872 if (graph->count[j] > 1) { 873 graph->mirrors[j]++; 874 k++; 875 } 876 } 877 /* allocate space for set of mirrors */ 878 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 879 for (i=1;i<graph->nvtxs;i++) 880 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 881 882 /* fill arrays */ 883 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 884 for (j=0;j<n_shared[0];j++) { 885 i=shared[0][j]; 886 if (graph->count[i] > 1) 887 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 888 } 889 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 890 for (i=0;i<graph->nvtxs;i++) { 891 if (graph->mirrors[i] > 0) { 892 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 893 j = global_indices[k]; 894 while ( k > 0 && global_indices[k-1] == j) k--; 895 for (j=0;j<graph->mirrors[i];j++) { 896 graph->mirrors_set[i][j]=local_indices[k+j]; 897 } 898 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 899 } 900 } 901 ierr = PetscFree(local_indices);CHKERRQ(ierr); 902 ierr = PetscFree(global_indices);CHKERRQ(ierr); 903 ierr = ISDestroy(&to);CHKERRQ(ierr); 904 ierr = ISDestroy(&from);CHKERRQ(ierr); 905 } 906 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 907 908 /* Count total number of neigh per node */ 909 k = 0; 910 for (i=1;i<n_neigh;i++) { 911 k += n_shared[i]; 912 for (j=0;j<n_shared[i];j++) { 913 graph->count[shared[i][j]] += 1; 914 } 915 } 916 /* Allocate space for storing the set of neighbours for each node */ 917 if (graph->nvtxs) { 918 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 919 } 920 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 921 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 922 } 923 /* Get information for sharing subdomains */ 924 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 925 for (i=1;i<n_neigh;i++) { /* dont count myself */ 926 s = n_shared[i]; 927 for (j=0;j<s;j++) { 928 k = shared[i][j]; 929 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 930 graph->count[k] += 1; 931 } 932 } 933 /* sort set of sharing subdomains */ 934 for (i=0;i<graph->nvtxs;i++) { 935 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 936 } 937 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 938 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 939 940 /* 941 Get info for dofs splitting 942 User can specify just a subset; an additional field is considered as a complementary field 943 */ 944 for (i=0;i<graph->nvtxs;i++) graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 945 for (i=0;i<n_ISForDofs;i++) { 946 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 947 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 948 for (j=0;j<is_size;j++) { 949 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 950 graph->which_dof[is_indices[j]] = i; 951 } 952 } 953 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 954 } 955 956 /* Take into account Neumann nodes */ 957 if (neumann_is) { 958 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 959 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 960 for (i=0;i<is_size;i++) { 961 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 962 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_NEUMANN_MARK; 963 } 964 } 965 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 966 } 967 /* Take into account Dirichlet nodes (they overwrite any neumann boundary mark previously set) */ 968 if (dirichlet_is) { 969 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 970 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 971 for (i=0;i<is_size;i++){ 972 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 973 if (commsize > graph->commsizelimit) { /* dirichlet nodes treated as internal */ 974 ierr = PetscBTSet(graph->touched,is_indices[i]);CHKERRQ(ierr); 975 graph->subset[is_indices[i]] = 0; 976 } 977 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_DIRICHLET_MARK; 978 } 979 } 980 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 981 } 982 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 983 if (graph->mirrors) { 984 for (i=0;i<graph->nvtxs;i++) 985 if (graph->mirrors[i]) 986 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 987 988 if (graph->xadj) { 989 PetscInt *new_xadj,*new_adjncy; 990 /* sort CSR graph */ 991 for (i=0;i<graph->nvtxs;i++) 992 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 993 994 /* adapt local CSR graph in case of local periodicity */ 995 k = 0; 996 for (i=0;i<graph->nvtxs;i++) 997 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 998 k += graph->mirrors[graph->adjncy[j]]; 999 1000 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 1001 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 1002 new_xadj[0] = 0; 1003 for (i=0;i<graph->nvtxs;i++) { 1004 k = graph->xadj[i+1]-graph->xadj[i]; 1005 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 1006 new_xadj[i+1] = new_xadj[i]+k; 1007 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 1008 k = graph->mirrors[graph->adjncy[j]]; 1009 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 1010 new_xadj[i+1] += k; 1011 } 1012 k = new_xadj[i+1]-new_xadj[i]; 1013 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 1014 new_xadj[i+1] = new_xadj[i]+k; 1015 } 1016 /* set new CSR into graph */ 1017 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1018 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1019 graph->xadj = new_xadj; 1020 graph->adjncy = new_adjncy; 1021 } 1022 } 1023 1024 /* mark special nodes (if any) -> each will become a single node equivalence class */ 1025 if (custom_primal_vertices) { 1026 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 1027 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1028 for (i=0,j=0;i<is_size;i++){ 1029 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs && graph->special_dof[is_indices[i]] != PCBDDCGRAPH_DIRICHLET_MARK) { /* out of bounds indices (if any) are skipped */ 1030 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_SPECIAL_MARK-j; 1031 j++; 1032 } 1033 } 1034 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1035 } 1036 1037 /* mark interior nodes (if commsize > graph->commsizelimit) as touched and belonging to partition number 0 */ 1038 if (commsize > graph->commsizelimit) { 1039 for (i=0;i<graph->nvtxs;i++) { 1040 if (!graph->count[i]) { 1041 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1042 graph->subset[i] = 0; 1043 } 1044 } 1045 } 1046 1047 /* init graph structure and compute default subsets */ 1048 nodes_touched = 0; 1049 for (i=0;i<graph->nvtxs;i++) { 1050 if (PetscBTLookup(graph->touched,i)) { 1051 nodes_touched++; 1052 } 1053 } 1054 i = 0; 1055 graph->ncc = 0; 1056 total_counts = 0; 1057 1058 /* allocated space for queues */ 1059 if (commsize == graph->commsizelimit) { 1060 ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,graph->nvtxs,&graph->queue);CHKERRQ(ierr); 1061 } else { 1062 PetscInt nused = graph->nvtxs - nodes_touched; 1063 ierr = PetscMalloc2(nused+1,&graph->cptr,nused,&graph->queue);CHKERRQ(ierr); 1064 } 1065 1066 while (nodes_touched<graph->nvtxs) { 1067 /* find first untouched node in local ordering */ 1068 while (PetscBTLookup(graph->touched,i)) i++; 1069 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1070 graph->subset[i] = graph->ncc+1; 1071 graph->cptr[graph->ncc] = total_counts; 1072 graph->queue[total_counts] = i; 1073 total_counts++; 1074 nodes_touched++; 1075 /* now find all other nodes having the same set of sharing subdomains */ 1076 for (j=i+1;j<graph->nvtxs;j++) { 1077 /* check for same number of sharing subdomains, dof number and same special mark */ 1078 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]) { 1079 /* check for same set of sharing subdomains */ 1080 same_set = PETSC_TRUE; 1081 for (k=0;k<graph->count[j];k++){ 1082 if (graph->neighbours_set[i][k] != graph->neighbours_set[j][k]) { 1083 same_set = PETSC_FALSE; 1084 } 1085 } 1086 /* I have found a friend of mine */ 1087 if (same_set) { 1088 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1089 graph->subset[j] = graph->ncc+1; 1090 nodes_touched++; 1091 graph->queue[total_counts] = j; 1092 total_counts++; 1093 } 1094 } 1095 } 1096 graph->ncc++; 1097 } 1098 /* set default number of subsets (at this point no info on csr and/or local_subs has been taken into account, so n_subsets = ncc */ 1099 graph->n_subsets = graph->ncc; 1100 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1101 for (i=0;i<graph->n_subsets;i++) { 1102 graph->subset_ncc[i] = 1; 1103 } 1104 /* final pointer */ 1105 graph->cptr[graph->ncc] = total_counts; 1106 1107 /* For consistency reasons (among neighbours), I need to sort (by global ordering) each connected component */ 1108 /* Get a reference node (min index in global ordering) for each subset for tagging messages */ 1109 ierr = PetscMalloc1(graph->ncc,&graph->subset_ref_node);CHKERRQ(ierr); 1110 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1111 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1112 for (j=0;j<graph->ncc;j++) { 1113 ierr = PetscSortIntWithArray(graph->cptr[j+1]-graph->cptr[j],&queue_global[graph->cptr[j]],&graph->queue[graph->cptr[j]]);CHKERRQ(ierr); 1114 graph->subset_ref_node[j] = graph->queue[graph->cptr[j]]; 1115 } 1116 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1117 graph->queue_sorted = PETSC_TRUE; 1118 1119 /* save information on subsets (needed when analyzing the connected components) */ 1120 if (graph->ncc) { 1121 ierr = PetscMalloc2(graph->ncc,&graph->subset_size,graph->ncc,&graph->subset_idxs);CHKERRQ(ierr); 1122 ierr = PetscMalloc1(graph->cptr[graph->ncc],&graph->subset_idxs[0]);CHKERRQ(ierr); 1123 ierr = PetscMemzero(graph->subset_idxs[0],graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1124 for (j=1;j<graph->ncc;j++) { 1125 graph->subset_size[j-1] = graph->cptr[j] - graph->cptr[j-1]; 1126 graph->subset_idxs[j] = graph->subset_idxs[j-1] + graph->subset_size[j-1]; 1127 } 1128 graph->subset_size[graph->ncc-1] = graph->cptr[graph->ncc] - graph->cptr[graph->ncc-1]; 1129 ierr = PetscMemcpy(graph->subset_idxs[0],graph->queue,graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1130 } 1131 1132 /* renumber reference nodes */ 1133 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->ncc,graph->subset_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 1134 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,subset_n,&subset);CHKERRQ(ierr); 1135 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1136 ierr = ISRenumber(subset,NULL,NULL,&subset_n);CHKERRQ(ierr); 1137 ierr = ISDestroy(&subset);CHKERRQ(ierr); 1138 ierr = ISGetLocalSize(subset_n,&k);CHKERRQ(ierr); 1139 if (k != graph->ncc) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",k,graph->ncc); 1140 ierr = ISGetIndices(subset_n,&is_indices);CHKERRQ(ierr); 1141 ierr = PetscMemcpy(graph->subset_ref_node,is_indices,graph->ncc*sizeof(PetscInt));CHKERRQ(ierr); 1142 ierr = ISRestoreIndices(subset_n,&is_indices);CHKERRQ(ierr); 1143 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1144 1145 /* free workspace */ 1146 graph->setupcalled = PETSC_TRUE; 1147 PetscFunctionReturn(0); 1148 } 1149 1150 PetscErrorCode PCBDDCGraphResetCoords(PCBDDCGraph graph) 1151 { 1152 PetscErrorCode ierr; 1153 1154 PetscFunctionBegin; 1155 if (!graph) PetscFunctionReturn(0); 1156 ierr = PetscFree(graph->coords);CHKERRQ(ierr); 1157 graph->cdim = 0; 1158 graph->cnloc = 0; 1159 graph->cloc = PETSC_FALSE; 1160 PetscFunctionReturn(0); 1161 } 1162 1163 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1164 { 1165 PetscErrorCode ierr; 1166 1167 PetscFunctionBegin; 1168 if (!graph) PetscFunctionReturn(0); 1169 if (graph->freecsr) { 1170 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1171 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1172 } else { 1173 graph->xadj = NULL; 1174 graph->adjncy = NULL; 1175 } 1176 graph->freecsr = PETSC_FALSE; 1177 graph->nvtxs_csr = 0; 1178 PetscFunctionReturn(0); 1179 } 1180 1181 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1182 { 1183 PetscErrorCode ierr; 1184 1185 PetscFunctionBegin; 1186 if (!graph) PetscFunctionReturn(0); 1187 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1188 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1189 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1190 if (graph->nvtxs) { 1191 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1192 } 1193 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1194 ierr = PetscFree5(graph->count, 1195 graph->neighbours_set, 1196 graph->subset, 1197 graph->which_dof, 1198 graph->special_dof);CHKERRQ(ierr); 1199 ierr = PetscFree2(graph->cptr,graph->queue);CHKERRQ(ierr); 1200 if (graph->mirrors) { 1201 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1202 } 1203 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1204 if (graph->subset_idxs) { 1205 ierr = PetscFree(graph->subset_idxs[0]);CHKERRQ(ierr); 1206 } 1207 ierr = PetscFree2(graph->subset_size,graph->subset_idxs);CHKERRQ(ierr); 1208 ierr = ISDestroy(&graph->dirdofs);CHKERRQ(ierr); 1209 ierr = ISDestroy(&graph->dirdofsB);CHKERRQ(ierr); 1210 if (graph->n_local_subs) { 1211 ierr = PetscFree(graph->local_subs);CHKERRQ(ierr); 1212 } 1213 graph->has_dirichlet = PETSC_FALSE; 1214 graph->twodimset = PETSC_FALSE; 1215 graph->twodim = PETSC_FALSE; 1216 graph->nvtxs = 0; 1217 graph->nvtxs_global = 0; 1218 graph->n_subsets = 0; 1219 graph->custom_minimal_size = 1; 1220 graph->n_local_subs = 0; 1221 graph->maxcount = PETSC_MAX_INT; 1222 graph->setupcalled = PETSC_FALSE; 1223 PetscFunctionReturn(0); 1224 } 1225 1226 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap, PetscInt N, PetscInt maxcount) 1227 { 1228 PetscInt n; 1229 PetscErrorCode ierr; 1230 1231 PetscFunctionBegin; 1232 PetscValidPointer(graph,1); 1233 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1234 PetscValidLogicalCollectiveInt(l2gmap,N,3); 1235 PetscValidLogicalCollectiveInt(l2gmap,maxcount,4); 1236 /* raise an error if already allocated */ 1237 if (graph->nvtxs_global) SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1238 /* set number of vertices */ 1239 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1240 graph->l2gmap = l2gmap; 1241 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1242 graph->nvtxs = n; 1243 graph->nvtxs_global = N; 1244 /* allocate used space */ 1245 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1246 ierr = PetscMalloc5(graph->nvtxs,&graph->count, 1247 graph->nvtxs,&graph->neighbours_set, 1248 graph->nvtxs,&graph->subset, 1249 graph->nvtxs,&graph->which_dof, 1250 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1251 /* zeroes memory */ 1252 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1253 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1254 /* use -1 as a default value for which_dof array */ 1255 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1256 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1257 /* zeroes first pointer to neighbour set */ 1258 if (graph->nvtxs) { 1259 graph->neighbours_set[0] = 0; 1260 } 1261 /* zeroes workspace for values of ncc */ 1262 graph->subset_ncc = 0; 1263 graph->subset_ref_node = 0; 1264 /* maxcount for cc */ 1265 graph->maxcount = maxcount; 1266 PetscFunctionReturn(0); 1267 } 1268 1269 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1270 { 1271 PetscErrorCode ierr; 1272 1273 PetscFunctionBegin; 1274 ierr = PCBDDCGraphResetCSR(*graph);CHKERRQ(ierr); 1275 ierr = PCBDDCGraphResetCoords(*graph);CHKERRQ(ierr); 1276 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1277 ierr = PetscFree(*graph);CHKERRQ(ierr); 1278 PetscFunctionReturn(0); 1279 } 1280 1281 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1282 { 1283 PCBDDCGraph new_graph; 1284 PetscErrorCode ierr; 1285 1286 PetscFunctionBegin; 1287 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1288 new_graph->custom_minimal_size = 1; 1289 new_graph->commsizelimit = 1; 1290 *graph = new_graph; 1291 PetscFunctionReturn(0); 1292 } 1293