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 #if 0 375 printf("RANK %d: start %d (%g %g) (size %d)\n",PetscGlobalRank,shared[ns][0],wdist[0],wdist[1],n_shared[ns]); 376 printf("RANK %d: point1 %d (%g %g)\n",PetscGlobalRank,shared[ns][point1],wdist[point1*cdim + 0],wdist[point1*cdim + 1]); 377 printf("RANK %d: point2 %d (%g %g)\n",PetscGlobalRank,shared[ns][point2],wdist[point2*cdim + 0],wdist[point2*cdim + 1]); 378 printf("RANK %d: point3 %d (%g %g)\n",PetscGlobalRank,shared[ns][point3],wdist[point3*cdim + 0],wdist[point3*cdim + 1]); 379 #endif 380 /* all dofs having the same coordinates will be primal */ 381 for (j=0;j<n_shared[ns];j++) { 382 PetscBool same[3] = {PETSC_TRUE,PETSC_TRUE,PETSC_TRUE}; 383 384 for (d=0;d<cdim;d++) { 385 same[0] = (PetscBool)(same[0] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point1*cdim+d]) < PETSC_SMALL)); 386 same[1] = (PetscBool)(same[1] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point2*cdim+d]) < PETSC_SMALL)); 387 same[2] = (PetscBool)(same[2] && (PetscAbsReal(wdist[j*cdim + d]-wdist[point3*cdim+d]) < PETSC_SMALL)); 388 } 389 if (same[0] || same[1] || same[2]) { 390 ierr = PetscBTSet(cornerp,shared[ns][j]);CHKERRQ(ierr); 391 } 392 } 393 } 394 ierr = PetscFree(wdist);CHKERRQ(ierr); 395 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 396 } 397 398 /* check consistency of connected components among neighbouring subdomains -> it adapt them in case it is needed */ 399 ierr = MPI_Comm_size(interface_comm,&size);CHKERRQ(ierr); 400 adapt_interface_reduced = PETSC_FALSE; 401 if (size > 1) { 402 PetscInt i; 403 PetscBool adapt_interface = cornerp ? PETSC_TRUE : PETSC_FALSE; 404 for (i=0;i<graph->n_subsets && !adapt_interface;i++) { 405 /* We are not sure that on a given subset of the local interface, 406 with two connected components, the latters be the same among sharing subdomains */ 407 if (graph->subset_ncc[i] > 1) adapt_interface = PETSC_TRUE; 408 } 409 ierr = MPIU_Allreduce(&adapt_interface,&adapt_interface_reduced,1,MPIU_BOOL,MPI_LOR,interface_comm);CHKERRQ(ierr); 410 } 411 412 if (graph->n_subsets && adapt_interface_reduced) { 413 PetscBT subset_cc_adapt; 414 MPI_Request *send_requests,*recv_requests; 415 PetscInt *send_buffer,*recv_buffer; 416 PetscInt sum_requests,start_of_recv,start_of_send; 417 PetscInt *cum_recv_counts; 418 PetscInt *labels; 419 PetscInt ncc,cum_queue,mss,mns,j,k,s; 420 PetscInt **refine_buffer=NULL,*private_labels = NULL; 421 PetscBool *subset_has_corn,*recv_buffer_bool; 422 423 ierr = PetscCalloc1(graph->n_subsets,&subset_has_corn);CHKERRQ(ierr); 424 if (cornerp) { 425 for (i=0;i<graph->n_subsets;i++) { 426 for (j=0;j<graph->subset_size[i];j++) { 427 if (PetscBTLookup(cornerp,graph->subset_idxs[i][j])) { 428 subset_has_corn[i] = PETSC_TRUE; 429 break; 430 } 431 } 432 } 433 } 434 ierr = PetscMalloc1(graph->nvtxs,&labels);CHKERRQ(ierr); 435 ierr = PetscMemzero(labels,graph->nvtxs*sizeof(*labels));CHKERRQ(ierr); 436 for (i=0,k=0;i<graph->ncc;i++) { 437 PetscInt s = 1; 438 for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 439 if (cornerp && PetscBTLookup(cornerp,graph->queue[j])) { 440 labels[graph->queue[j]] = k+s; 441 s += 1; 442 } else { 443 labels[graph->queue[j]] = k; 444 } 445 } 446 k += s; 447 } 448 449 /* allocate some space */ 450 ierr = PetscMalloc1(graph->n_subsets+1,&cum_recv_counts);CHKERRQ(ierr); 451 ierr = PetscMemzero(cum_recv_counts,(graph->n_subsets+1)*sizeof(*cum_recv_counts));CHKERRQ(ierr); 452 453 /* first count how many neighbours per connected component I will receive from */ 454 cum_recv_counts[0] = 0; 455 for (i=0;i<graph->n_subsets;i++) cum_recv_counts[i+1] = cum_recv_counts[i]+graph->count[graph->subset_idxs[i][0]]; 456 ierr = PetscMalloc1(cum_recv_counts[graph->n_subsets],&recv_buffer_bool);CHKERRQ(ierr); 457 ierr = PetscMalloc2(cum_recv_counts[graph->n_subsets],&send_requests,cum_recv_counts[graph->n_subsets],&recv_requests);CHKERRQ(ierr); 458 for (i=0;i<cum_recv_counts[graph->n_subsets];i++) { 459 send_requests[i] = MPI_REQUEST_NULL; 460 recv_requests[i] = MPI_REQUEST_NULL; 461 } 462 463 /* exchange with my neighbours the number of my connected components on the subset of interface */ 464 sum_requests = 0; 465 for (i=0;i<graph->n_subsets;i++) { 466 PetscMPIInt neigh,tag; 467 PetscInt count,*neighs; 468 469 count = graph->count[graph->subset_idxs[i][0]]; 470 neighs = graph->neighbours_set[graph->subset_idxs[i][0]]; 471 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i],&tag);CHKERRQ(ierr); 472 for (k=0;k<count;k++) { 473 PetscBool ss = (PetscBool)(graph->subset_ncc[i] > 1 || subset_has_corn[i]); 474 475 ierr = PetscMPIIntCast(neighs[k],&neigh);CHKERRQ(ierr); 476 ierr = MPI_Isend(&ss,1,MPIU_BOOL,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 477 ierr = MPI_Irecv(&recv_buffer_bool[sum_requests],1,MPIU_BOOL,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 478 sum_requests++; 479 } 480 } 481 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 482 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 483 484 /* determine the subsets I have to adapt (those having more than 1 cc) */ 485 ierr = PetscBTCreate(graph->n_subsets,&subset_cc_adapt);CHKERRQ(ierr); 486 ierr = PetscBTMemzero(graph->n_subsets,subset_cc_adapt);CHKERRQ(ierr); 487 for (i=0;i<graph->n_subsets;i++) { 488 if (graph->subset_ncc[i] > 1 || subset_has_corn[i]) { 489 ierr = PetscBTSet(subset_cc_adapt,i);CHKERRQ(ierr); 490 continue; 491 } 492 for (j=cum_recv_counts[i];j<cum_recv_counts[i+1];j++){ 493 if (recv_buffer_bool[j]) { 494 ierr = PetscBTSet(subset_cc_adapt,i);CHKERRQ(ierr); 495 break; 496 } 497 } 498 } 499 ierr = PetscFree(recv_buffer_bool);CHKERRQ(ierr); 500 ierr = PetscFree(subset_has_corn);CHKERRQ(ierr); 501 502 /* determine send/recv buffers sizes */ 503 j = 0; 504 mss = 0; 505 for (i=0;i<graph->n_subsets;i++) { 506 if (PetscBTLookup(subset_cc_adapt,i)) { 507 j += graph->subset_size[i]; 508 mss = PetscMax(graph->subset_size[i],mss); 509 } 510 } 511 k = 0; 512 mns = 0; 513 for (i=0;i<graph->n_subsets;i++) { 514 if (PetscBTLookup(subset_cc_adapt,i)) { 515 k += (cum_recv_counts[i+1]-cum_recv_counts[i])*graph->subset_size[i]; 516 mns = PetscMax(cum_recv_counts[i+1]-cum_recv_counts[i],mns); 517 } 518 } 519 ierr = PetscMalloc2(j,&send_buffer,k,&recv_buffer);CHKERRQ(ierr); 520 521 /* fill send buffer (order matters: subset_idxs ordered by global ordering) */ 522 j = 0; 523 for (i=0;i<graph->n_subsets;i++) 524 if (PetscBTLookup(subset_cc_adapt,i)) 525 for (k=0;k<graph->subset_size[i];k++) 526 send_buffer[j++] = labels[graph->subset_idxs[i][k]]; 527 528 /* now exchange the data */ 529 start_of_recv = 0; 530 start_of_send = 0; 531 sum_requests = 0; 532 for (i=0;i<graph->n_subsets;i++) { 533 if (PetscBTLookup(subset_cc_adapt,i)) { 534 PetscMPIInt neigh,tag; 535 PetscInt size_of_send = graph->subset_size[i]; 536 537 j = graph->subset_idxs[i][0]; 538 ierr = PetscMPIIntCast(2*graph->subset_ref_node[i]+1,&tag);CHKERRQ(ierr); 539 for (k=0;k<graph->count[j];k++) { 540 ierr = PetscMPIIntCast(graph->neighbours_set[j][k],&neigh);CHKERRQ(ierr); 541 ierr = MPI_Isend(&send_buffer[start_of_send],size_of_send,MPIU_INT,neigh,tag,interface_comm,&send_requests[sum_requests]);CHKERRQ(ierr); 542 ierr = MPI_Irecv(&recv_buffer[start_of_recv],size_of_send,MPIU_INT,neigh,tag,interface_comm,&recv_requests[sum_requests]);CHKERRQ(ierr); 543 start_of_recv += size_of_send; 544 sum_requests++; 545 } 546 start_of_send += size_of_send; 547 } 548 } 549 ierr = MPI_Waitall(sum_requests,recv_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 550 551 /* refine connected components */ 552 start_of_recv = 0; 553 /* allocate some temporary space */ 554 if (mss) { 555 ierr = PetscMalloc1(mss,&refine_buffer);CHKERRQ(ierr); 556 ierr = PetscMalloc2(mss*(mns+1),&refine_buffer[0],mss,&private_labels);CHKERRQ(ierr); 557 } 558 ncc = 0; 559 cum_queue = 0; 560 graph->cptr[0] = 0; 561 for (i=0;i<graph->n_subsets;i++) { 562 if (PetscBTLookup(subset_cc_adapt,i)) { 563 PetscInt subset_counter = 0; 564 PetscInt sharingprocs = cum_recv_counts[i+1]-cum_recv_counts[i]+1; /* count myself */ 565 PetscInt buffer_size = graph->subset_size[i]; 566 567 /* compute pointers */ 568 for (j=1;j<buffer_size;j++) refine_buffer[j] = refine_buffer[j-1] + sharingprocs; 569 /* analyze contributions from subdomains that share the i-th subset 570 The structure of refine_buffer is suitable to find intersections of ccs among sharingprocs. 571 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) 572 sharing procs connected components: 573 neigh 0: [0 1 4], [2 3], labels [4,7] (2 connected components) 574 neigh 1: [0 1], [2 3 4], labels [3 2] (2 connected components) 575 neigh 2: [0 4], [1], [2 3], labels [1 5 6] (3 connected components) 576 refine_buffer will be filled as: 577 [ 4, 3, 1; 578 4, 2, 1; 579 7, 2, 6; 580 4, 3, 5; 581 7, 2, 6; ]; 582 The connected components in local ordering are [0], [1], [2 3], [4] */ 583 /* fill temp_buffer */ 584 for (k=0;k<buffer_size;k++) refine_buffer[k][0] = labels[graph->subset_idxs[i][k]]; 585 for (j=0;j<sharingprocs-1;j++) { 586 for (k=0;k<buffer_size;k++) refine_buffer[k][j+1] = recv_buffer[start_of_recv+k]; 587 start_of_recv += buffer_size; 588 } 589 ierr = PetscMemzero(private_labels,buffer_size*sizeof(PetscInt));CHKERRQ(ierr); 590 for (j=0;j<buffer_size;j++) { 591 if (!private_labels[j]) { /* found a new cc */ 592 PetscBool same_set; 593 594 graph->cptr[ncc] = cum_queue; 595 ncc++; 596 subset_counter++; 597 private_labels[j] = subset_counter; 598 graph->queue[cum_queue++] = graph->subset_idxs[i][j]; 599 for (k=j+1;k<buffer_size;k++) { /* check for other nodes in new cc */ 600 same_set = PETSC_TRUE; 601 for (s=0;s<sharingprocs;s++) { 602 if (refine_buffer[j][s] != refine_buffer[k][s]) { 603 same_set = PETSC_FALSE; 604 break; 605 } 606 } 607 if (same_set) { 608 private_labels[k] = subset_counter; 609 graph->queue[cum_queue++] = graph->subset_idxs[i][k]; 610 } 611 } 612 } 613 } 614 graph->cptr[ncc] = cum_queue; 615 graph->subset_ncc[i] = subset_counter; 616 graph->queue_sorted = PETSC_FALSE; 617 } else { /* this subset does not need to be adapted */ 618 ierr = PetscMemcpy(graph->queue+cum_queue,graph->subset_idxs[i],graph->subset_size[i]*sizeof(PetscInt));CHKERRQ(ierr); 619 ncc++; 620 cum_queue += graph->subset_size[i]; 621 graph->cptr[ncc] = cum_queue; 622 } 623 } 624 graph->cptr[ncc] = cum_queue; 625 graph->ncc = ncc; 626 if (mss) { 627 ierr = PetscFree2(refine_buffer[0],private_labels);CHKERRQ(ierr); 628 ierr = PetscFree(refine_buffer);CHKERRQ(ierr); 629 } 630 ierr = PetscFree(labels);CHKERRQ(ierr); 631 ierr = MPI_Waitall(sum_requests,send_requests,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 632 ierr = PetscFree2(send_requests,recv_requests);CHKERRQ(ierr); 633 ierr = PetscFree2(send_buffer,recv_buffer);CHKERRQ(ierr); 634 ierr = PetscFree(cum_recv_counts);CHKERRQ(ierr); 635 ierr = PetscBTDestroy(&subset_cc_adapt);CHKERRQ(ierr); 636 } 637 ierr = PetscBTDestroy(&cornerp);CHKERRQ(ierr); 638 639 /* Determine if we are in 2D or 3D */ 640 if (!graph->twodimset) { 641 PetscBool twodim = PETSC_TRUE; 642 for (i=0;i<graph->ncc;i++) { 643 PetscInt repdof = graph->queue[graph->cptr[i]]; 644 PetscInt ccsize = graph->cptr[i+1]-graph->cptr[i]; 645 if (graph->count[repdof] > 1 && ccsize > graph->custom_minimal_size) { 646 twodim = PETSC_FALSE; 647 break; 648 } 649 } 650 ierr = MPIU_Allreduce(&twodim,&graph->twodim,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)graph->l2gmap));CHKERRQ(ierr); 651 graph->twodimset = PETSC_TRUE; 652 } 653 PetscFunctionReturn(0); 654 } 655 656 657 PETSC_STATIC_INLINE PetscErrorCode PCBDDCGraphComputeCC_Private(PCBDDCGraph graph,PetscInt pid,PetscInt* queue_tip,PetscInt n_prev,PetscInt* n_added) 658 { 659 PetscInt i,j,n; 660 PetscInt *xadj = graph->xadj,*adjncy = graph->adjncy; 661 PetscBT touched = graph->touched; 662 PetscBool havecsr = (PetscBool)(!!xadj); 663 PetscBool havesubs = (PetscBool)(!!graph->n_local_subs); 664 PetscErrorCode ierr; 665 666 PetscFunctionBegin; 667 n = 0; 668 if (havecsr && !havesubs) { 669 for (i=-n_prev;i<0;i++) { 670 PetscInt start_dof = queue_tip[i]; 671 /* we assume that if a dof has a size 1 adjacency list and the corresponding entry is negative, it is connected to all dofs */ 672 if (xadj[start_dof+1]-xadj[start_dof] == 1 && adjncy[xadj[start_dof]] < 0) { 673 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 674 PetscInt dof = graph->subset_idxs[pid-1][j]; 675 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 676 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 677 queue_tip[n] = dof; 678 n++; 679 } 680 } 681 } else { 682 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 683 PetscInt dof = adjncy[j]; 684 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 685 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 686 queue_tip[n] = dof; 687 n++; 688 } 689 } 690 } 691 } 692 } else if (havecsr && havesubs) { 693 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 694 for (i=-n_prev;i<0;i++) { 695 PetscInt start_dof = queue_tip[i]; 696 /* 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 */ 697 if (xadj[start_dof+1]-xadj[start_dof] == 1 && adjncy[xadj[start_dof]] < 0) { 698 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 699 PetscInt dof = graph->subset_idxs[pid-1][j]; 700 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 701 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 702 queue_tip[n] = dof; 703 n++; 704 } 705 } 706 } else { 707 for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 708 PetscInt dof = adjncy[j]; 709 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 710 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 711 queue_tip[n] = dof; 712 n++; 713 } 714 } 715 } 716 } 717 } else if (havesubs) { /* sub info only */ 718 PetscInt sid = graph->local_subs[queue_tip[-n_prev]]; 719 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 720 PetscInt dof = graph->subset_idxs[pid-1][j]; 721 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid && graph->local_subs[dof] == sid) { 722 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 723 queue_tip[n] = dof; 724 n++; 725 } 726 } 727 } else { 728 for (j=0;j<graph->subset_size[pid-1];j++) { /* pid \in [1,graph->n_subsets] */ 729 PetscInt dof = graph->subset_idxs[pid-1][j]; 730 if (!PetscBTLookup(touched,dof) && graph->subset[dof] == pid) { 731 ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 732 queue_tip[n] = dof; 733 n++; 734 } 735 } 736 } 737 *n_added = n; 738 PetscFunctionReturn(0); 739 } 740 741 PetscErrorCode PCBDDCGraphComputeConnectedComponentsLocal(PCBDDCGraph graph) 742 { 743 PetscInt ncc,cum_queue,n; 744 PetscMPIInt commsize; 745 PetscErrorCode ierr; 746 747 PetscFunctionBegin; 748 if (!graph->setupcalled) SETERRQ(PetscObjectComm((PetscObject)graph->l2gmap),PETSC_ERR_ORDER,"PCBDDCGraphSetUp should be called first"); 749 /* quiet return if there isn't any local info */ 750 if (!graph->xadj && !graph->n_local_subs) { 751 PetscFunctionReturn(0); 752 } 753 754 /* reset any previous search of connected components */ 755 ierr = PetscBTMemzero(graph->nvtxs,graph->touched);CHKERRQ(ierr); 756 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)graph->l2gmap),&commsize);CHKERRQ(ierr); 757 if (commsize > graph->commsizelimit) { 758 PetscInt i; 759 for (i=0;i<graph->nvtxs;i++) { 760 if (graph->special_dof[i] == PCBDDCGRAPH_DIRICHLET_MARK || !graph->count[i]) { 761 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 762 } 763 } 764 } 765 766 /* begin search for connected components */ 767 cum_queue = 0; 768 ncc = 0; 769 for (n=0;n<graph->n_subsets;n++) { 770 PetscInt pid = n+1; /* partition labeled by 0 is discarded */ 771 PetscInt found = 0,prev = 0,first = 0,ncc_pid = 0; 772 while (found != graph->subset_size[n]) { 773 PetscInt added = 0; 774 if (!prev) { /* search for new starting dof */ 775 while (PetscBTLookup(graph->touched,graph->subset_idxs[n][first])) first++; 776 ierr = PetscBTSet(graph->touched,graph->subset_idxs[n][first]);CHKERRQ(ierr); 777 graph->queue[cum_queue] = graph->subset_idxs[n][first]; 778 graph->cptr[ncc] = cum_queue; 779 prev = 1; 780 cum_queue++; 781 found++; 782 ncc_pid++; 783 ncc++; 784 } 785 ierr = PCBDDCGraphComputeCC_Private(graph,pid,graph->queue + cum_queue,prev,&added);CHKERRQ(ierr); 786 if (!added) { 787 graph->subset_ncc[n] = ncc_pid; 788 graph->cptr[ncc] = cum_queue; 789 } 790 prev = added; 791 found += added; 792 cum_queue += added; 793 if (added && found == graph->subset_size[n]) { 794 graph->subset_ncc[n] = ncc_pid; 795 graph->cptr[ncc] = cum_queue; 796 } 797 } 798 } 799 graph->ncc = ncc; 800 graph->queue_sorted = PETSC_FALSE; 801 PetscFunctionReturn(0); 802 } 803 804 PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices) 805 { 806 IS subset,subset_n; 807 MPI_Comm comm; 808 const PetscInt *is_indices; 809 PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global; 810 PetscInt i,j,k,s,total_counts,nodes_touched,is_size; 811 PetscMPIInt commsize; 812 PetscBool same_set,mirrors_found; 813 PetscErrorCode ierr; 814 815 PetscFunctionBegin; 816 PetscValidLogicalCollectiveInt(graph->l2gmap,custom_minimal_size,2); 817 if (neumann_is) { 818 PetscValidHeaderSpecific(neumann_is,IS_CLASSID,3); 819 PetscCheckSameComm(graph->l2gmap,1,neumann_is,3); 820 } 821 graph->has_dirichlet = PETSC_FALSE; 822 if (dirichlet_is) { 823 PetscValidHeaderSpecific(dirichlet_is,IS_CLASSID,4); 824 PetscCheckSameComm(graph->l2gmap,1,dirichlet_is,4); 825 graph->has_dirichlet = PETSC_TRUE; 826 } 827 PetscValidLogicalCollectiveInt(graph->l2gmap,n_ISForDofs,5); 828 for (i=0;i<n_ISForDofs;i++) { 829 PetscValidHeaderSpecific(ISForDofs[i],IS_CLASSID,6); 830 PetscCheckSameComm(graph->l2gmap,1,ISForDofs[i],6); 831 } 832 if (custom_primal_vertices) { 833 PetscValidHeaderSpecific(custom_primal_vertices,IS_CLASSID,7); 834 PetscCheckSameComm(graph->l2gmap,1,custom_primal_vertices,7); 835 } 836 ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr); 837 ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 838 839 /* custom_minimal_size */ 840 graph->custom_minimal_size = custom_minimal_size; 841 /* get info l2gmap and allocate work vectors */ 842 ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 843 /* check if we have any local periodic nodes (periodic BCs) */ 844 mirrors_found = PETSC_FALSE; 845 if (graph->nvtxs && n_neigh) { 846 for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1; 847 for (i=0; i<n_shared[0]; i++) { 848 if (graph->count[shared[0][i]] > 1) { 849 mirrors_found = PETSC_TRUE; 850 break; 851 } 852 } 853 } 854 /* compute local mirrors (if any) */ 855 if (mirrors_found) { 856 IS to,from; 857 PetscInt *local_indices,*global_indices; 858 859 ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr); 860 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr); 861 /* get arrays of local and global indices */ 862 ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr); 863 ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 864 ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 865 ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr); 866 ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr); 867 ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 868 ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 869 ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr); 870 /* allocate space for mirrors */ 871 ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr); 872 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 873 graph->mirrors_set[0] = 0; 874 875 k=0; 876 for (i=0;i<n_shared[0];i++) { 877 j=shared[0][i]; 878 if (graph->count[j] > 1) { 879 graph->mirrors[j]++; 880 k++; 881 } 882 } 883 /* allocate space for set of mirrors */ 884 ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr); 885 for (i=1;i<graph->nvtxs;i++) 886 graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1]; 887 888 /* fill arrays */ 889 ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 890 for (j=0;j<n_shared[0];j++) { 891 i=shared[0][j]; 892 if (graph->count[i] > 1) 893 graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i]; 894 } 895 ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr); 896 for (i=0;i<graph->nvtxs;i++) { 897 if (graph->mirrors[i] > 0) { 898 ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr); 899 j = global_indices[k]; 900 while ( k > 0 && global_indices[k-1] == j) k--; 901 for (j=0;j<graph->mirrors[i];j++) { 902 graph->mirrors_set[i][j]=local_indices[k+j]; 903 } 904 ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr); 905 } 906 } 907 ierr = PetscFree(local_indices);CHKERRQ(ierr); 908 ierr = PetscFree(global_indices);CHKERRQ(ierr); 909 ierr = ISDestroy(&to);CHKERRQ(ierr); 910 ierr = ISDestroy(&from);CHKERRQ(ierr); 911 } 912 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 913 914 /* Count total number of neigh per node */ 915 k = 0; 916 for (i=1;i<n_neigh;i++) { 917 k += n_shared[i]; 918 for (j=0;j<n_shared[i];j++) { 919 graph->count[shared[i][j]] += 1; 920 } 921 } 922 /* Allocate space for storing the set of neighbours for each node */ 923 if (graph->nvtxs) { 924 ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr); 925 } 926 for (i=1;i<graph->nvtxs;i++) { /* dont count myself */ 927 graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1]; 928 } 929 /* Get information for sharing subdomains */ 930 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr); 931 for (i=1;i<n_neigh;i++) { /* dont count myself */ 932 s = n_shared[i]; 933 for (j=0;j<s;j++) { 934 k = shared[i][j]; 935 graph->neighbours_set[k][graph->count[k]] = neigh[i]; 936 graph->count[k] += 1; 937 } 938 } 939 /* sort set of sharing subdomains */ 940 for (i=0;i<graph->nvtxs;i++) { 941 ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr); 942 } 943 /* free memory allocated by ISLocalToGlobalMappingGetInfo */ 944 ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 945 946 /* 947 Get info for dofs splitting 948 User can specify just a subset; an additional field is considered as a complementary field 949 */ 950 for (i=0;i<graph->nvtxs;i++) graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */ 951 for (i=0;i<n_ISForDofs;i++) { 952 ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr); 953 ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 954 for (j=0;j<is_size;j++) { 955 if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 956 graph->which_dof[is_indices[j]] = i; 957 } 958 } 959 ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr); 960 } 961 962 /* Take into account Neumann nodes */ 963 if (neumann_is) { 964 ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr); 965 ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 966 for (i=0;i<is_size;i++) { 967 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 968 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_NEUMANN_MARK; 969 } 970 } 971 ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 972 } 973 /* Take into account Dirichlet nodes (they overwrite any neumann boundary mark previously set) */ 974 if (dirichlet_is) { 975 ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr); 976 ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 977 for (i=0;i<is_size;i++){ 978 if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */ 979 if (commsize > graph->commsizelimit) { /* dirichlet nodes treated as internal */ 980 ierr = PetscBTSet(graph->touched,is_indices[i]);CHKERRQ(ierr); 981 graph->subset[is_indices[i]] = 0; 982 } 983 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_DIRICHLET_MARK; 984 } 985 } 986 ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 987 } 988 /* mark local periodic nodes (if any) and adapt CSR graph (if any) */ 989 if (graph->mirrors) { 990 for (i=0;i<graph->nvtxs;i++) 991 if (graph->mirrors[i]) 992 graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK; 993 994 if (graph->xadj) { 995 PetscInt *new_xadj,*new_adjncy; 996 /* sort CSR graph */ 997 for (i=0;i<graph->nvtxs;i++) 998 ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr); 999 1000 /* adapt local CSR graph in case of local periodicity */ 1001 k = 0; 1002 for (i=0;i<graph->nvtxs;i++) 1003 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) 1004 k += graph->mirrors[graph->adjncy[j]]; 1005 1006 ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr); 1007 ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr); 1008 new_xadj[0] = 0; 1009 for (i=0;i<graph->nvtxs;i++) { 1010 k = graph->xadj[i+1]-graph->xadj[i]; 1011 ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr); 1012 new_xadj[i+1] = new_xadj[i]+k; 1013 for (j=graph->xadj[i];j<graph->xadj[i+1];j++) { 1014 k = graph->mirrors[graph->adjncy[j]]; 1015 ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr); 1016 new_xadj[i+1] += k; 1017 } 1018 k = new_xadj[i+1]-new_xadj[i]; 1019 ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr); 1020 new_xadj[i+1] = new_xadj[i]+k; 1021 } 1022 /* set new CSR into graph */ 1023 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1024 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1025 graph->xadj = new_xadj; 1026 graph->adjncy = new_adjncy; 1027 } 1028 } 1029 1030 /* mark special nodes (if any) -> each will become a single node equivalence class */ 1031 if (custom_primal_vertices) { 1032 ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr); 1033 ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1034 for (i=0,j=0;i<is_size;i++){ 1035 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 */ 1036 graph->special_dof[is_indices[i]] = PCBDDCGRAPH_SPECIAL_MARK-j; 1037 j++; 1038 } 1039 } 1040 ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1041 } 1042 1043 /* mark interior nodes (if commsize > graph->commsizelimit) as touched and belonging to partition number 0 */ 1044 if (commsize > graph->commsizelimit) { 1045 for (i=0;i<graph->nvtxs;i++) { 1046 if (!graph->count[i]) { 1047 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1048 graph->subset[i] = 0; 1049 } 1050 } 1051 } 1052 1053 /* init graph structure and compute default subsets */ 1054 nodes_touched = 0; 1055 for (i=0;i<graph->nvtxs;i++) { 1056 if (PetscBTLookup(graph->touched,i)) { 1057 nodes_touched++; 1058 } 1059 } 1060 i = 0; 1061 graph->ncc = 0; 1062 total_counts = 0; 1063 1064 /* allocated space for queues */ 1065 if (commsize == graph->commsizelimit) { 1066 ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,graph->nvtxs,&graph->queue);CHKERRQ(ierr); 1067 } else { 1068 PetscInt nused = graph->nvtxs - nodes_touched; 1069 ierr = PetscMalloc2(nused+1,&graph->cptr,nused,&graph->queue);CHKERRQ(ierr); 1070 } 1071 1072 while (nodes_touched<graph->nvtxs) { 1073 /* find first untouched node in local ordering */ 1074 while (PetscBTLookup(graph->touched,i)) i++; 1075 ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr); 1076 graph->subset[i] = graph->ncc+1; 1077 graph->cptr[graph->ncc] = total_counts; 1078 graph->queue[total_counts] = i; 1079 total_counts++; 1080 nodes_touched++; 1081 /* now find all other nodes having the same set of sharing subdomains */ 1082 for (j=i+1;j<graph->nvtxs;j++) { 1083 /* check for same number of sharing subdomains, dof number and same special mark */ 1084 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]) { 1085 /* check for same set of sharing subdomains */ 1086 same_set = PETSC_TRUE; 1087 for (k=0;k<graph->count[j];k++){ 1088 if (graph->neighbours_set[i][k] != graph->neighbours_set[j][k]) { 1089 same_set = PETSC_FALSE; 1090 } 1091 } 1092 /* I have found a friend of mine */ 1093 if (same_set) { 1094 ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr); 1095 graph->subset[j] = graph->ncc+1; 1096 nodes_touched++; 1097 graph->queue[total_counts] = j; 1098 total_counts++; 1099 } 1100 } 1101 } 1102 graph->ncc++; 1103 } 1104 /* 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 */ 1105 graph->n_subsets = graph->ncc; 1106 ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr); 1107 for (i=0;i<graph->n_subsets;i++) { 1108 graph->subset_ncc[i] = 1; 1109 } 1110 /* final pointer */ 1111 graph->cptr[graph->ncc] = total_counts; 1112 1113 /* For consistency reasons (among neighbours), I need to sort (by global ordering) each connected component */ 1114 /* Get a reference node (min index in global ordering) for each subset for tagging messages */ 1115 ierr = PetscMalloc1(graph->ncc,&graph->subset_ref_node);CHKERRQ(ierr); 1116 ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr); 1117 ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr); 1118 for (j=0;j<graph->ncc;j++) { 1119 ierr = PetscSortIntWithArray(graph->cptr[j+1]-graph->cptr[j],&queue_global[graph->cptr[j]],&graph->queue[graph->cptr[j]]);CHKERRQ(ierr); 1120 graph->subset_ref_node[j] = graph->queue[graph->cptr[j]]; 1121 } 1122 ierr = PetscFree(queue_global);CHKERRQ(ierr); 1123 graph->queue_sorted = PETSC_TRUE; 1124 1125 /* save information on subsets (needed when analyzing the connected components) */ 1126 if (graph->ncc) { 1127 ierr = PetscMalloc2(graph->ncc,&graph->subset_size,graph->ncc,&graph->subset_idxs);CHKERRQ(ierr); 1128 ierr = PetscMalloc1(graph->cptr[graph->ncc],&graph->subset_idxs[0]);CHKERRQ(ierr); 1129 ierr = PetscMemzero(graph->subset_idxs[0],graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1130 for (j=1;j<graph->ncc;j++) { 1131 graph->subset_size[j-1] = graph->cptr[j] - graph->cptr[j-1]; 1132 graph->subset_idxs[j] = graph->subset_idxs[j-1] + graph->subset_size[j-1]; 1133 } 1134 graph->subset_size[graph->ncc-1] = graph->cptr[graph->ncc] - graph->cptr[graph->ncc-1]; 1135 ierr = PetscMemcpy(graph->subset_idxs[0],graph->queue,graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr); 1136 } 1137 1138 /* renumber reference nodes */ 1139 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->ncc,graph->subset_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 1140 ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,subset_n,&subset);CHKERRQ(ierr); 1141 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1142 ierr = ISRenumber(subset,NULL,NULL,&subset_n);CHKERRQ(ierr); 1143 ierr = ISDestroy(&subset);CHKERRQ(ierr); 1144 ierr = ISGetLocalSize(subset_n,&k);CHKERRQ(ierr); 1145 if (k != graph->ncc) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",k,graph->ncc); 1146 ierr = ISGetIndices(subset_n,&is_indices);CHKERRQ(ierr); 1147 ierr = PetscMemcpy(graph->subset_ref_node,is_indices,graph->ncc*sizeof(PetscInt));CHKERRQ(ierr); 1148 ierr = ISRestoreIndices(subset_n,&is_indices);CHKERRQ(ierr); 1149 ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 1150 1151 /* free workspace */ 1152 graph->setupcalled = PETSC_TRUE; 1153 PetscFunctionReturn(0); 1154 } 1155 1156 PetscErrorCode PCBDDCGraphResetCoords(PCBDDCGraph graph) 1157 { 1158 PetscErrorCode ierr; 1159 1160 PetscFunctionBegin; 1161 if (!graph) PetscFunctionReturn(0); 1162 ierr = PetscFree(graph->coords);CHKERRQ(ierr); 1163 graph->cdim = 0; 1164 graph->cnloc = 0; 1165 graph->cloc = PETSC_FALSE; 1166 PetscFunctionReturn(0); 1167 } 1168 1169 PetscErrorCode PCBDDCGraphResetCSR(PCBDDCGraph graph) 1170 { 1171 PetscErrorCode ierr; 1172 1173 PetscFunctionBegin; 1174 if (!graph) PetscFunctionReturn(0); 1175 if (graph->freecsr) { 1176 ierr = PetscFree(graph->xadj);CHKERRQ(ierr); 1177 ierr = PetscFree(graph->adjncy);CHKERRQ(ierr); 1178 } else { 1179 graph->xadj = NULL; 1180 graph->adjncy = NULL; 1181 } 1182 graph->freecsr = PETSC_FALSE; 1183 graph->nvtxs_csr = 0; 1184 PetscFunctionReturn(0); 1185 } 1186 1187 PetscErrorCode PCBDDCGraphReset(PCBDDCGraph graph) 1188 { 1189 PetscErrorCode ierr; 1190 1191 PetscFunctionBegin; 1192 if (!graph) PetscFunctionReturn(0); 1193 ierr = ISLocalToGlobalMappingDestroy(&graph->l2gmap);CHKERRQ(ierr); 1194 ierr = PetscFree(graph->subset_ncc);CHKERRQ(ierr); 1195 ierr = PetscFree(graph->subset_ref_node);CHKERRQ(ierr); 1196 if (graph->nvtxs) { 1197 ierr = PetscFree(graph->neighbours_set[0]);CHKERRQ(ierr); 1198 } 1199 ierr = PetscBTDestroy(&graph->touched);CHKERRQ(ierr); 1200 ierr = PetscFree5(graph->count, 1201 graph->neighbours_set, 1202 graph->subset, 1203 graph->which_dof, 1204 graph->special_dof);CHKERRQ(ierr); 1205 ierr = PetscFree2(graph->cptr,graph->queue);CHKERRQ(ierr); 1206 if (graph->mirrors) { 1207 ierr = PetscFree(graph->mirrors_set[0]);CHKERRQ(ierr); 1208 } 1209 ierr = PetscFree2(graph->mirrors,graph->mirrors_set);CHKERRQ(ierr); 1210 if (graph->subset_idxs) { 1211 ierr = PetscFree(graph->subset_idxs[0]);CHKERRQ(ierr); 1212 } 1213 ierr = PetscFree2(graph->subset_size,graph->subset_idxs);CHKERRQ(ierr); 1214 ierr = ISDestroy(&graph->dirdofs);CHKERRQ(ierr); 1215 ierr = ISDestroy(&graph->dirdofsB);CHKERRQ(ierr); 1216 if (graph->n_local_subs) { 1217 ierr = PetscFree(graph->local_subs);CHKERRQ(ierr); 1218 } 1219 graph->has_dirichlet = PETSC_FALSE; 1220 graph->twodimset = PETSC_FALSE; 1221 graph->twodim = PETSC_FALSE; 1222 graph->nvtxs = 0; 1223 graph->nvtxs_global = 0; 1224 graph->n_subsets = 0; 1225 graph->custom_minimal_size = 1; 1226 graph->n_local_subs = 0; 1227 graph->maxcount = PETSC_MAX_INT; 1228 graph->setupcalled = PETSC_FALSE; 1229 PetscFunctionReturn(0); 1230 } 1231 1232 PetscErrorCode PCBDDCGraphInit(PCBDDCGraph graph, ISLocalToGlobalMapping l2gmap, PetscInt N, PetscInt maxcount) 1233 { 1234 PetscInt n; 1235 PetscErrorCode ierr; 1236 1237 PetscFunctionBegin; 1238 PetscValidPointer(graph,1); 1239 PetscValidHeaderSpecific(l2gmap,IS_LTOGM_CLASSID,2); 1240 PetscValidLogicalCollectiveInt(l2gmap,N,3); 1241 PetscValidLogicalCollectiveInt(l2gmap,maxcount,4); 1242 /* raise an error if already allocated */ 1243 if (graph->nvtxs_global) SETERRQ(PetscObjectComm((PetscObject)l2gmap),PETSC_ERR_PLIB,"BDDCGraph already initialized"); 1244 /* set number of vertices */ 1245 ierr = PetscObjectReference((PetscObject)l2gmap);CHKERRQ(ierr); 1246 graph->l2gmap = l2gmap; 1247 ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 1248 graph->nvtxs = n; 1249 graph->nvtxs_global = N; 1250 /* allocate used space */ 1251 ierr = PetscBTCreate(graph->nvtxs,&graph->touched);CHKERRQ(ierr); 1252 ierr = PetscMalloc5(graph->nvtxs,&graph->count, 1253 graph->nvtxs,&graph->neighbours_set, 1254 graph->nvtxs,&graph->subset, 1255 graph->nvtxs,&graph->which_dof, 1256 graph->nvtxs,&graph->special_dof);CHKERRQ(ierr); 1257 /* zeroes memory */ 1258 ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1259 ierr = PetscMemzero(graph->subset,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1260 /* use -1 as a default value for which_dof array */ 1261 for (n=0;n<graph->nvtxs;n++) graph->which_dof[n] = -1; 1262 ierr = PetscMemzero(graph->special_dof,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr); 1263 /* zeroes first pointer to neighbour set */ 1264 if (graph->nvtxs) { 1265 graph->neighbours_set[0] = 0; 1266 } 1267 /* zeroes workspace for values of ncc */ 1268 graph->subset_ncc = 0; 1269 graph->subset_ref_node = 0; 1270 /* maxcount for cc */ 1271 graph->maxcount = maxcount; 1272 PetscFunctionReturn(0); 1273 } 1274 1275 PetscErrorCode PCBDDCGraphDestroy(PCBDDCGraph* graph) 1276 { 1277 PetscErrorCode ierr; 1278 1279 PetscFunctionBegin; 1280 ierr = PCBDDCGraphResetCSR(*graph);CHKERRQ(ierr); 1281 ierr = PCBDDCGraphResetCoords(*graph);CHKERRQ(ierr); 1282 ierr = PCBDDCGraphReset(*graph);CHKERRQ(ierr); 1283 ierr = PetscFree(*graph);CHKERRQ(ierr); 1284 PetscFunctionReturn(0); 1285 } 1286 1287 PetscErrorCode PCBDDCGraphCreate(PCBDDCGraph *graph) 1288 { 1289 PCBDDCGraph new_graph; 1290 PetscErrorCode ierr; 1291 1292 PetscFunctionBegin; 1293 ierr = PetscNew(&new_graph);CHKERRQ(ierr); 1294 new_graph->custom_minimal_size = 1; 1295 new_graph->commsizelimit = 1; 1296 *graph = new_graph; 1297 PetscFunctionReturn(0); 1298 } 1299