1 /* 2 GAMG geometric-algebric multiogrid PC - Mark Adams 2011 3 */ 4 5 #include <../src/ksp/pc/impls/gamg/gamg.h> /*I "petscpc.h" I*/ 6 #include <private/kspimpl.h> 7 8 #if defined(PETSC_HAVE_TRIANGLE) 9 #define REAL PetscReal 10 #include <triangle.h> 11 #endif 12 13 #include <assert.h> 14 #include <petscblaslapack.h> 15 16 /* Private context for the GAMG preconditioner */ 17 typedef struct{ 18 PetscInt lid; /* local vertex index */ 19 PetscInt degree; /* vertex degree */ 20 } GAMGNode; 21 int compare (const void *a, const void *b) 22 { 23 return (((GAMGNode*)a)->degree - ((GAMGNode*)b)->degree); 24 } 25 26 /* -------------------------------------------------------------------------- */ 27 /* 28 PCSetCoordinates_GEO 29 30 Input Parameter: 31 . pc - the preconditioner context 32 */ 33 EXTERN_C_BEGIN 34 #undef __FUNCT__ 35 #define __FUNCT__ "PCSetCoordinates_GEO" 36 PetscErrorCode PCSetCoordinates_GEO( PC pc, PetscInt ndm, PetscReal *coords ) 37 { 38 PC_MG *mg = (PC_MG*)pc->data; 39 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 40 PetscErrorCode ierr; 41 PetscInt arrsz,bs,my0,kk,ii,nloc,Iend; 42 Mat Amat = pc->pmat; 43 44 PetscFunctionBegin; 45 PetscValidHeaderSpecific( Amat, MAT_CLASSID, 1 ); 46 ierr = MatGetBlockSize( Amat, &bs ); CHKERRQ( ierr ); 47 ierr = MatGetOwnershipRange( Amat, &my0, &Iend ); CHKERRQ(ierr); 48 nloc = (Iend-my0)/bs; 49 if((Iend-my0)%bs!=0) SETERRQ1(((PetscObject)Amat)->comm,PETSC_ERR_ARG_WRONG, "Bad local size %d.",nloc); 50 51 pc_gamg->data_cell_rows = 1; 52 if( coords==0 && nloc > 0 ) { 53 SETERRQ(((PetscObject)Amat)->comm,PETSC_ERR_ARG_WRONG, "Need coordinates for pc_gamg_type 'geo'."); 54 } 55 pc_gamg->data_cell_cols = ndm; /* coordinates */ 56 57 arrsz = nloc*pc_gamg->data_cell_rows*pc_gamg->data_cell_cols; 58 59 /* create data - syntactic sugar that should be refactored at some point */ 60 if (pc_gamg->data==0 || (pc_gamg->data_sz != arrsz)) { 61 ierr = PetscFree( pc_gamg->data ); CHKERRQ(ierr); 62 ierr = PetscMalloc((arrsz+1)*sizeof(PetscReal), &pc_gamg->data ); CHKERRQ(ierr); 63 } 64 for(kk=0;kk<arrsz;kk++)pc_gamg->data[kk] = -999.; 65 pc_gamg->data[arrsz] = -99.; 66 /* copy data in - column oriented */ 67 for( kk = 0 ; kk < nloc ; kk++ ){ 68 for( ii = 0 ; ii < ndm ; ii++ ) { 69 pc_gamg->data[ii*nloc + kk] = coords[kk*ndm + ii]; 70 } 71 } 72 assert(pc_gamg->data[arrsz] == -99.); 73 74 pc_gamg->data_sz = arrsz; 75 76 PetscFunctionReturn(0); 77 } 78 EXTERN_C_END 79 80 /* -------------------------------------------------------------------------- */ 81 /* 82 PCSetData_GEO 83 84 Input Parameter: 85 . pc - 86 */ 87 #undef __FUNCT__ 88 #define __FUNCT__ "PCSetData_GEO" 89 PetscErrorCode PCSetData_GEO( PC pc ) 90 { 91 PetscFunctionBegin; 92 SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_LIB,"GEO MG needs coordinates"); 93 } 94 95 /* -------------------------------------------------------------------------- */ 96 /* 97 PCSetFromOptions_GEO 98 99 Input Parameter: 100 . pc - 101 */ 102 #undef __FUNCT__ 103 #define __FUNCT__ "PCSetFromOptions_GEO" 104 PetscErrorCode PCSetFromOptions_GEO( PC pc ) 105 { 106 PetscErrorCode ierr; 107 PC_MG *mg = (PC_MG*)pc->data; 108 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 109 110 PetscFunctionBegin; 111 ierr = PetscOptionsHead("GAMG-GEO options"); CHKERRQ(ierr); 112 { 113 /* -pc_gamg_sa_nsmooths */ 114 /* pc_gamg_sa->smooths = 0; */ 115 /* ierr = PetscOptionsInt("-pc_gamg_agg_nsmooths", */ 116 /* "smoothing steps for smoothed aggregation, usually 1 (0)", */ 117 /* "PCGAMGSetNSmooths_AGG", */ 118 /* pc_gamg_sa->smooths, */ 119 /* &pc_gamg_sa->smooths, */ 120 /* &flag); */ 121 /* CHKERRQ(ierr); */ 122 } 123 ierr = PetscOptionsTail();CHKERRQ(ierr); 124 125 /* call base class */ 126 ierr = PCSetFromOptions_GAMG( pc ); CHKERRQ(ierr); 127 128 if( pc_gamg->verbose ) { 129 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s done\n",0,__FUNCT__); 130 } 131 132 PetscFunctionReturn(0); 133 } 134 135 /* -------------------------------------------------------------------------- */ 136 /* 137 triangulateAndFormProl 138 139 Input Parameter: 140 . selected_2 - list of selected local ID, includes selected ghosts 141 . nnodes - 142 . coords[2*nnodes] - column vector of local coordinates w/ ghosts 143 . nselected_1 - selected IDs that go with base (1) graph 144 . clid_lid_1 - [nselected_1] lids of selected (c) nodes 145 . locals_llist - linked list with (some) locality info of base graph 146 . crsGID[selected.size()] - global index for prolongation operator 147 . bs - block size 148 Output Parameter: 149 . a_Prol - prolongation operator 150 . a_worst_best - measure of worst missed fine vertex, 0 is no misses 151 */ 152 #undef __FUNCT__ 153 #define __FUNCT__ "triangulateAndFormProl" 154 PetscErrorCode triangulateAndFormProl( IS selected_2, /* list of selected local ID, includes selected ghosts */ 155 const PetscInt nnodes, 156 const PetscReal coords[], /* column vector of local coordinates w/ ghosts */ 157 const PetscInt nselected_1, /* list of selected local ID, includes selected ghosts */ 158 const PetscInt clid_lid_1[], 159 IS locals_llist, /* linked list from selected_1 vertices of aggregate unselected vertices */ 160 const PetscInt crsGID[], 161 const PetscInt bs, 162 Mat a_Prol, /* prolongation operator (output) */ 163 PetscReal *a_worst_best /* measure of worst missed fine vertex, 0 is no misses */ 164 ) 165 { 166 #if defined(PETSC_HAVE_TRIANGLE) 167 PetscErrorCode ierr; 168 PetscInt jj,tid,tt,idx,nselected_2; 169 struct triangulateio in,mid; 170 const PetscInt *selected_idx_2,*llist_idx; 171 PetscMPIInt mype,npe; 172 PetscInt Istart,Iend,nFineLoc,myFine0; 173 int kk,nPlotPts,sid; 174 MPI_Comm wcomm = ((PetscObject)a_Prol)->comm; 175 PetscReal tm; 176 PetscFunctionBegin; 177 178 ierr = MPI_Comm_rank(wcomm,&mype); CHKERRQ(ierr); 179 ierr = MPI_Comm_size(wcomm,&npe); CHKERRQ(ierr); 180 ierr = ISGetSize( selected_2, &nselected_2 ); CHKERRQ(ierr); 181 if(nselected_2 == 1 || nselected_2 == 2 ){ /* 0 happens on idle processors */ 182 *a_worst_best = 100.0; /* this will cause a stop, but not globalized (should not happen) */ 183 } 184 else *a_worst_best = 0.0; 185 ierr = MPI_Allreduce( a_worst_best, &tm, 1, MPIU_REAL, MPIU_MAX, wcomm ); CHKERRQ(ierr); 186 if( tm > 0.0 ) { 187 *a_worst_best = 100.0; 188 PetscFunctionReturn(0); 189 } 190 ierr = MatGetOwnershipRange( a_Prol, &Istart, &Iend ); CHKERRQ(ierr); 191 nFineLoc = (Iend-Istart)/bs; myFine0 = Istart/bs; 192 nPlotPts = nFineLoc; /* locals */ 193 /* traingle */ 194 /* Define input points - in*/ 195 in.numberofpoints = nselected_2; 196 in.numberofpointattributes = 0; 197 /* get nselected points */ 198 ierr = PetscMalloc( 2*(nselected_2)*sizeof(REAL), &in.pointlist ); CHKERRQ(ierr); 199 ierr = ISGetIndices( selected_2, &selected_idx_2 ); CHKERRQ(ierr); 200 201 for(kk=0,sid=0;kk<nselected_2;kk++,sid += 2){ 202 PetscInt lid = selected_idx_2[kk]; 203 in.pointlist[sid] = coords[lid]; 204 in.pointlist[sid+1] = coords[nnodes + lid]; 205 if(lid>=nFineLoc) nPlotPts++; 206 } 207 assert(sid==2*nselected_2); 208 209 in.numberofsegments = 0; 210 in.numberofedges = 0; 211 in.numberofholes = 0; 212 in.numberofregions = 0; 213 in.trianglelist = 0; 214 in.segmentmarkerlist = 0; 215 in.pointattributelist = 0; 216 in.pointmarkerlist = 0; 217 in.triangleattributelist = 0; 218 in.trianglearealist = 0; 219 in.segmentlist = 0; 220 in.holelist = 0; 221 in.regionlist = 0; 222 in.edgelist = 0; 223 in.edgemarkerlist = 0; 224 in.normlist = 0; 225 /* triangulate */ 226 mid.pointlist = 0; /* Not needed if -N switch used. */ 227 /* Not needed if -N switch used or number of point attributes is zero: */ 228 mid.pointattributelist = 0; 229 mid.pointmarkerlist = 0; /* Not needed if -N or -B switch used. */ 230 mid.trianglelist = 0; /* Not needed if -E switch used. */ 231 /* Not needed if -E switch used or number of triangle attributes is zero: */ 232 mid.triangleattributelist = 0; 233 mid.neighborlist = 0; /* Needed only if -n switch used. */ 234 /* Needed only if segments are output (-p or -c) and -P not used: */ 235 mid.segmentlist = 0; 236 /* Needed only if segments are output (-p or -c) and -P and -B not used: */ 237 mid.segmentmarkerlist = 0; 238 mid.edgelist = 0; /* Needed only if -e switch used. */ 239 mid.edgemarkerlist = 0; /* Needed if -e used and -B not used. */ 240 mid.numberoftriangles = 0; 241 242 /* Triangulate the points. Switches are chosen to read and write a */ 243 /* PSLG (p), preserve the convex hull (c), number everything from */ 244 /* zero (z), assign a regional attribute to each element (A), and */ 245 /* produce an edge list (e), a Voronoi diagram (v), and a triangle */ 246 /* neighbor list (n). */ 247 if(nselected_2 != 0){ /* inactive processor */ 248 char args[] = "npczQ"; /* c is needed ? */ 249 triangulate(args, &in, &mid, (struct triangulateio *) NULL ); 250 /* output .poly files for 'showme' */ 251 if( !PETSC_TRUE ) { 252 static int level = 1; 253 FILE *file; char fname[32]; 254 255 sprintf(fname,"C%d_%d.poly",level,mype); file = fopen(fname, "w"); 256 /*First line: <# of vertices> <dimension (must be 2)> <# of attributes> <# of boundary markers (0 or 1)>*/ 257 fprintf(file, "%d %d %d %d\n",in.numberofpoints,2,0,0); 258 /*Following lines: <vertex #> <x> <y> */ 259 for(kk=0,sid=0;kk<in.numberofpoints;kk++,sid += 2){ 260 fprintf(file, "%d %e %e\n",kk,in.pointlist[sid],in.pointlist[sid+1]); 261 } 262 /*One line: <# of segments> <# of boundary markers (0 or 1)> */ 263 fprintf(file, "%d %d\n",0,0); 264 /*Following lines: <segment #> <endpoint> <endpoint> [boundary marker] */ 265 /* One line: <# of holes> */ 266 fprintf(file, "%d\n",0); 267 /* Following lines: <hole #> <x> <y> */ 268 /* Optional line: <# of regional attributes and/or area constraints> */ 269 /* Optional following lines: <region #> <x> <y> <attribute> <maximum area> */ 270 fclose(file); 271 272 /* elems */ 273 sprintf(fname,"C%d_%d.ele",level,mype); file = fopen(fname, "w"); 274 /* First line: <# of triangles> <nodes per triangle> <# of attributes> */ 275 fprintf(file, "%d %d %d\n",mid.numberoftriangles,3,0); 276 /* Remaining lines: <triangle #> <node> <node> <node> ... [attributes] */ 277 for(kk=0,sid=0;kk<mid.numberoftriangles;kk++,sid += 3){ 278 fprintf(file, "%d %d %d %d\n",kk,mid.trianglelist[sid],mid.trianglelist[sid+1],mid.trianglelist[sid+2]); 279 } 280 fclose(file); 281 282 sprintf(fname,"C%d_%d.node",level,mype); file = fopen(fname, "w"); 283 /* First line: <# of vertices> <dimension (must be 2)> <# of attributes> <# of boundary markers (0 or 1)> */ 284 /* fprintf(file, "%d %d %d %d\n",in.numberofpoints,2,0,0); */ 285 fprintf(file, "%d %d %d %d\n",nPlotPts,2,0,0); 286 /*Following lines: <vertex #> <x> <y> */ 287 for(kk=0,sid=0;kk<in.numberofpoints;kk++,sid+=2){ 288 fprintf(file, "%d %e %e\n",kk,in.pointlist[sid],in.pointlist[sid+1]); 289 } 290 291 sid /= 2; 292 for(jj=0;jj<nFineLoc;jj++){ 293 PetscBool sel = PETSC_TRUE; 294 for( kk=0 ; kk<nselected_2 && sel ; kk++ ){ 295 PetscInt lid = selected_idx_2[kk]; 296 if( lid == jj ) sel = PETSC_FALSE; 297 } 298 if( sel ) { 299 fprintf(file, "%d %e %e\n",sid++,coords[jj],coords[nnodes + jj]); 300 } 301 } 302 fclose(file); 303 assert(sid==nPlotPts); 304 level++; 305 } 306 } 307 #if defined PETSC_USE_LOG 308 ierr = PetscLogEventBegin(gamg_setup_events[FIND_V],0,0,0,0);CHKERRQ(ierr); 309 #endif 310 { /* form P - setup some maps */ 311 PetscInt clid_iterator; 312 PetscInt *nTri, *node_tri; 313 314 ierr = PetscMalloc( nselected_2*sizeof(PetscInt), &node_tri ); CHKERRQ(ierr); 315 ierr = PetscMalloc( nselected_2*sizeof(PetscInt), &nTri ); CHKERRQ(ierr); 316 317 /* need list of triangles on node */ 318 for(kk=0;kk<nselected_2;kk++) nTri[kk] = 0; 319 for(tid=0,kk=0;tid<mid.numberoftriangles;tid++){ 320 for(jj=0;jj<3;jj++) { 321 PetscInt cid = mid.trianglelist[kk++]; 322 if( nTri[cid] == 0 ) node_tri[cid] = tid; 323 nTri[cid]++; 324 } 325 } 326 #define EPS 1.e-12 327 /* find points and set prolongation */ 328 ierr = ISGetIndices( locals_llist, &llist_idx ); CHKERRQ(ierr); 329 for( clid_iterator = 0 ; clid_iterator < nselected_1 ; clid_iterator++ ){ 330 PetscInt flid = clid_lid_1[clid_iterator]; assert(flid != -1); 331 PetscScalar AA[3][3]; 332 PetscBLASInt N=3,NRHS=1,LDA=3,IPIV[3],LDB=3,INFO; 333 do{ 334 if( flid < nFineLoc ) { /* could be a ghost */ 335 PetscInt bestTID = -1; PetscReal best_alpha = 1.e10; 336 const PetscInt fgid = flid + myFine0; 337 /* compute shape function for gid */ 338 const PetscReal fcoord[3] = { coords[flid], coords[nnodes + flid], 1.0 }; 339 PetscBool haveit = PETSC_FALSE; PetscScalar alpha[3]; PetscInt clids[3]; 340 /* look for it */ 341 for( tid = node_tri[clid_iterator], jj=0; 342 jj < 5 && !haveit && tid != -1; 343 jj++ ){ 344 for(tt=0;tt<3;tt++){ 345 PetscInt cid2 = mid.trianglelist[3*tid + tt]; 346 PetscInt lid2 = selected_idx_2[cid2]; 347 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 348 clids[tt] = cid2; /* store for interp */ 349 } 350 351 for(tt=0;tt<3;tt++) alpha[tt] = (PetscScalar)fcoord[tt]; 352 353 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 354 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 355 { 356 PetscBool have=PETSC_TRUE; PetscReal lowest=1.e10; 357 for( tt = 0, idx = 0 ; tt < 3 ; tt++ ) { 358 if( PetscRealPart(alpha[tt]) > (1.0+EPS) || PetscRealPart(alpha[tt]) < -EPS ) have = PETSC_FALSE; 359 if( PetscRealPart(alpha[tt]) < lowest ){ 360 lowest = PetscRealPart(alpha[tt]); 361 idx = tt; 362 } 363 } 364 haveit = have; 365 } 366 tid = mid.neighborlist[3*tid + idx]; 367 } 368 369 if( !haveit ) { 370 /* brute force */ 371 for(tid=0 ; tid<mid.numberoftriangles && !haveit ; tid++ ){ 372 for(tt=0;tt<3;tt++){ 373 PetscInt cid2 = mid.trianglelist[3*tid + tt]; 374 PetscInt lid2 = selected_idx_2[cid2]; 375 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 376 clids[tt] = cid2; /* store for interp */ 377 } 378 for(tt=0;tt<3;tt++) alpha[tt] = fcoord[tt]; 379 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 380 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 381 { 382 PetscBool have=PETSC_TRUE; PetscReal worst=0.0, v; 383 for(tt=0; tt<3 && have ;tt++) { 384 if( PetscRealPart(alpha[tt]) > 1.0+EPS || PetscRealPart(alpha[tt]) < -EPS ) have=PETSC_FALSE; 385 if( (v=PetscAbs(PetscRealPart(alpha[tt])-0.5)) > worst ) worst = v; 386 } 387 if( worst < best_alpha ) { 388 best_alpha = worst; bestTID = tid; 389 } 390 haveit = have; 391 } 392 } 393 } 394 if( !haveit ) { 395 if( best_alpha > *a_worst_best ) *a_worst_best = best_alpha; 396 /* use best one */ 397 for(tt=0;tt<3;tt++){ 398 PetscInt cid2 = mid.trianglelist[3*bestTID + tt]; 399 PetscInt lid2 = selected_idx_2[cid2]; 400 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 401 clids[tt] = cid2; /* store for interp */ 402 } 403 for(tt=0;tt<3;tt++) alpha[tt] = fcoord[tt]; 404 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 405 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 406 } 407 408 /* put in row of P */ 409 for(idx=0;idx<3;idx++){ 410 PetscScalar shp = alpha[idx]; 411 if( PetscAbs(PetscRealPart(shp)) > 1.e-6 ) { 412 PetscInt cgid = crsGID[clids[idx]]; 413 PetscInt jj = cgid*bs, ii = fgid*bs; /* need to gloalize */ 414 for(tt=0 ; tt < bs ; tt++, ii++, jj++ ){ 415 ierr = MatSetValues(a_Prol,1,&ii,1,&jj,&shp,INSERT_VALUES); CHKERRQ(ierr); 416 } 417 } 418 } 419 } /* local vertex test */ 420 } while( (flid=llist_idx[flid]) != -1 ); 421 } 422 ierr = ISRestoreIndices( selected_2, &selected_idx_2 ); CHKERRQ(ierr); 423 ierr = ISRestoreIndices( locals_llist, &llist_idx ); CHKERRQ(ierr); 424 ierr = MatAssemblyBegin(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 425 ierr = MatAssemblyEnd(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 426 427 ierr = PetscFree( node_tri ); CHKERRQ(ierr); 428 ierr = PetscFree( nTri ); CHKERRQ(ierr); 429 } 430 #if defined PETSC_USE_LOG 431 ierr = PetscLogEventEnd(gamg_setup_events[FIND_V],0,0,0,0);CHKERRQ(ierr); 432 #endif 433 free( mid.trianglelist ); 434 free( mid.neighborlist ); 435 ierr = PetscFree( in.pointlist ); CHKERRQ(ierr); 436 437 PetscFunctionReturn(0); 438 #else 439 SETERRQ(((PetscObject)a_Prol)->comm,PETSC_ERR_LIB,"configure with TRIANGLE to use geometric MG"); 440 #endif 441 } 442 /* -------------------------------------------------------------------------- */ 443 /* 444 getGIDsOnSquareGraph - square graph, get 445 446 Input Parameter: 447 . nselected_1 - selected local indices (includes ghosts in input Gmat_1) 448 . clid_lid_1 - [nselected_1] lids of selected nodes 449 . Gmat1 - graph that goes with 'selected_1' 450 Output Parameter: 451 . a_selected_2 - selected local indices (includes ghosts in output a_Gmat_2) 452 . a_Gmat_2 - graph that is squared of 'Gmat_1' 453 . a_crsGID[a_selected_2.size()] - map of global IDs of coarse grid nodes 454 */ 455 #undef __FUNCT__ 456 #define __FUNCT__ "getGIDsOnSquareGraph" 457 PetscErrorCode getGIDsOnSquareGraph( const PetscInt nselected_1, 458 const PetscInt clid_lid_1[], 459 const Mat Gmat1, 460 IS *a_selected_2, 461 Mat *a_Gmat_2, 462 PetscInt **a_crsGID 463 ) 464 { 465 PetscErrorCode ierr; 466 PetscMPIInt mype,npe; 467 PetscInt *crsGID, kk,my0,Iend,nloc; 468 MPI_Comm wcomm = ((PetscObject)Gmat1)->comm; 469 470 PetscFunctionBegin; 471 ierr = MPI_Comm_rank(wcomm,&mype);CHKERRQ(ierr); 472 ierr = MPI_Comm_size(wcomm,&npe);CHKERRQ(ierr); 473 ierr = MatGetOwnershipRange(Gmat1,&my0,&Iend); CHKERRQ(ierr); /* AIJ */ 474 nloc = Iend - my0; /* this does not change */ 475 476 if (npe == 1) { /* not much to do in serial */ 477 ierr = PetscMalloc( nselected_1*sizeof(PetscInt), &crsGID ); CHKERRQ(ierr); 478 for(kk=0;kk<nselected_1;kk++) crsGID[kk] = kk; 479 *a_Gmat_2 = 0; 480 ierr = ISCreateGeneral(PETSC_COMM_SELF,nselected_1,clid_lid_1,PETSC_COPY_VALUES,a_selected_2); 481 CHKERRQ(ierr); 482 } 483 else { 484 PetscInt idx,num_fine_ghosts,num_crs_ghost,myCrs0; 485 Mat_MPIAIJ *mpimat2; 486 Mat Gmat2; 487 Vec locState; 488 PetscScalar *cpcol_state; 489 490 /* scan my coarse zero gid, set 'lid_state' with coarse GID */ 491 kk = nselected_1; 492 MPI_Scan( &kk, &myCrs0, 1, MPIU_INT, MPIU_SUM, wcomm ); 493 myCrs0 -= nselected_1; 494 495 if( a_Gmat_2 ) { /* output */ 496 /* grow graph to get wider set of selected vertices to cover fine grid, invalidates 'llist' */ 497 ierr = MatTransposeMatMult(Gmat1, Gmat1, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Gmat2 ); CHKERRQ(ierr); 498 *a_Gmat_2 = Gmat2; /* output */ 499 } 500 else Gmat2 = Gmat1; /* use local to get crsGIDs at least */ 501 /* get coarse grid GIDS for selected (locals and ghosts) */ 502 mpimat2 = (Mat_MPIAIJ*)Gmat2->data; 503 ierr = MatGetVecs( Gmat2, &locState, 0 ); CHKERRQ(ierr); 504 ierr = VecSet( locState, (PetscScalar)(PetscReal)(-1) ); CHKERRQ(ierr); /* set with UNKNOWN state */ 505 for(kk=0;kk<nselected_1;kk++){ 506 PetscInt fgid = clid_lid_1[kk] + my0; 507 PetscScalar v = (PetscScalar)(kk+myCrs0); 508 ierr = VecSetValues( locState, 1, &fgid, &v, INSERT_VALUES ); CHKERRQ(ierr); /* set with PID */ 509 } 510 ierr = VecAssemblyBegin( locState ); CHKERRQ(ierr); 511 ierr = VecAssemblyEnd( locState ); CHKERRQ(ierr); 512 ierr = VecScatterBegin(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 513 ierr = VecScatterEnd(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 514 ierr = VecGetLocalSize( mpimat2->lvec, &num_fine_ghosts ); CHKERRQ(ierr); 515 ierr = VecGetArray( mpimat2->lvec, &cpcol_state ); CHKERRQ(ierr); 516 for(kk=0,num_crs_ghost=0;kk<num_fine_ghosts;kk++){ 517 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ) num_crs_ghost++; 518 } 519 ierr = PetscMalloc( (nselected_1+num_crs_ghost)*sizeof(PetscInt), &crsGID ); CHKERRQ(ierr); /* output */ 520 { 521 PetscInt *selected_set; 522 ierr = PetscMalloc( (nselected_1+num_crs_ghost)*sizeof(PetscInt), &selected_set ); CHKERRQ(ierr); 523 /* do ghost of 'crsGID' */ 524 for(kk=0,idx=nselected_1;kk<num_fine_ghosts;kk++){ 525 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ){ 526 PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]); 527 selected_set[idx] = nloc + kk; 528 crsGID[idx++] = cgid; 529 } 530 } 531 assert(idx==(nselected_1+num_crs_ghost)); 532 ierr = VecRestoreArray( mpimat2->lvec, &cpcol_state ); CHKERRQ(ierr); 533 /* do locals in 'crsGID' */ 534 ierr = VecGetArray( locState, &cpcol_state ); CHKERRQ(ierr); 535 for(kk=0,idx=0;kk<nloc;kk++){ 536 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ){ 537 PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]); 538 selected_set[idx] = kk; 539 crsGID[idx++] = cgid; 540 } 541 } 542 assert(idx==nselected_1); 543 ierr = VecRestoreArray( locState, &cpcol_state ); CHKERRQ(ierr); 544 545 if( a_selected_2 != 0 ) { /* output */ 546 ierr = ISCreateGeneral(PETSC_COMM_SELF,(nselected_1+num_crs_ghost),selected_set,PETSC_COPY_VALUES,a_selected_2); 547 CHKERRQ(ierr); 548 } 549 ierr = PetscFree( selected_set ); CHKERRQ(ierr); 550 } 551 ierr = VecDestroy( &locState ); CHKERRQ(ierr); 552 } 553 *a_crsGID = crsGID; /* output */ 554 555 PetscFunctionReturn(0); 556 } 557 558 /* -------------------------------------------------------------------------- */ 559 /* 560 PCGAMGgraph_GEO 561 562 Input Parameter: 563 . pc - this 564 . Amat - matrix on this fine level 565 Output Parameter: 566 . a_Gmat 567 */ 568 #undef __FUNCT__ 569 #define __FUNCT__ "PCGAMGgraph_GEO" 570 PetscErrorCode PCGAMGgraph_GEO( PC pc, 571 const Mat Amat, 572 Mat *a_Gmat 573 ) 574 { 575 PetscErrorCode ierr; 576 PC_MG *mg = (PC_MG*)pc->data; 577 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 578 const PetscInt verbose = pc_gamg->verbose; 579 const PetscReal vfilter = pc_gamg->threshold; 580 PetscMPIInt mype,npe; 581 MPI_Comm wcomm = ((PetscObject)Amat)->comm; 582 Mat Gmat; 583 584 PetscFunctionBegin; 585 ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 586 ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 587 588 ierr = createSimpleGraph( Amat, &Gmat ); CHKERRQ( ierr ); 589 ierr = scaleFilterGraph( &Gmat, vfilter, PETSC_TRUE, verbose ); CHKERRQ( ierr ); 590 591 *a_Gmat = Gmat; 592 593 PetscFunctionReturn(0); 594 } 595 596 /* -------------------------------------------------------------------------- */ 597 /* 598 PCGAMGcoarsen_GEO 599 600 Input Parameter: 601 . pc - this 602 . Gmat - graph 603 Output Parameter: 604 . a_selected - selected indices (local) 605 . a_llist_parent - linked list from selected indices for data locality only 606 */ 607 #undef __FUNCT__ 608 #define __FUNCT__ "PCGAMGcoarsen_GEO" 609 PetscErrorCode PCGAMGcoarsen_GEO( PC pc, 610 const Mat Gmat, 611 IS *a_selected, 612 IS *a_llist_parent 613 ) 614 { 615 PetscErrorCode ierr; 616 PC_MG *mg = (PC_MG*)pc->data; 617 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 618 PetscInt Istart,Iend,nloc,kk,Ii,ncols; 619 PetscMPIInt mype,npe; 620 IS perm,selected,llist_parent; 621 GAMGNode *gnodes; 622 PetscInt *permute; 623 MPI_Comm wcomm = ((PetscObject)Gmat)->comm; 624 MatCoarsen crs; 625 626 PetscFunctionBegin; 627 ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 628 ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 629 ierr = MatGetOwnershipRange( Gmat, &Istart, &Iend ); CHKERRQ(ierr); 630 nloc = (Iend-Istart); 631 632 /* create random permutation with sort for geo-mg */ 633 ierr = PetscMalloc( nloc*sizeof(GAMGNode), &gnodes ); CHKERRQ(ierr); 634 ierr = PetscMalloc( nloc*sizeof(PetscInt), &permute ); CHKERRQ(ierr); 635 636 for (Ii=Istart; Ii<Iend; Ii++) { /* locals only? */ 637 ierr = MatGetRow(Gmat,Ii,&ncols,0,0); CHKERRQ(ierr); 638 { 639 PetscInt lid = Ii - Istart; 640 gnodes[lid].lid = lid; 641 gnodes[lid].degree = ncols; 642 } 643 ierr = MatRestoreRow(Gmat,Ii,&ncols,0,0); CHKERRQ(ierr); 644 } 645 /* randomize */ 646 srand(1); /* make deterministic */ 647 if( PETSC_TRUE ) { 648 PetscBool *bIndexSet; 649 ierr = PetscMalloc( nloc*sizeof(PetscBool), &bIndexSet ); CHKERRQ(ierr); 650 for ( Ii = 0; Ii < nloc ; Ii++) bIndexSet[Ii] = PETSC_FALSE; 651 for ( Ii = 0; Ii < nloc ; Ii++) 652 { 653 PetscInt iSwapIndex = rand()%nloc; 654 if (!bIndexSet[iSwapIndex] && iSwapIndex != Ii) 655 { 656 GAMGNode iTemp = gnodes[iSwapIndex]; 657 gnodes[iSwapIndex] = gnodes[Ii]; 658 gnodes[Ii] = iTemp; 659 bIndexSet[Ii] = PETSC_TRUE; 660 bIndexSet[iSwapIndex] = PETSC_TRUE; 661 } 662 } 663 ierr = PetscFree( bIndexSet ); CHKERRQ(ierr); 664 } 665 /* only sort locals */ 666 qsort( gnodes, nloc, sizeof(GAMGNode), compare ); 667 /* create IS of permutation */ 668 for(kk=0;kk<nloc;kk++) { /* locals only */ 669 permute[kk] = gnodes[kk].lid; 670 } 671 ierr = ISCreateGeneral(PETSC_COMM_SELF, nloc, permute, PETSC_COPY_VALUES, &perm); 672 CHKERRQ(ierr); 673 674 ierr = PetscFree( gnodes ); CHKERRQ(ierr); 675 ierr = PetscFree( permute ); CHKERRQ(ierr); 676 677 /* get MIS aggs */ 678 /*ierr = maxIndSetAgg( perm, Gmat, PETSC_FALSE, pc_gamg->verbose, &selected, &llist_parent ); CHKERRQ(ierr);*/ 679 680 ierr = MatCoarsenCreate( wcomm, &crs ); CHKERRQ(ierr); 681 ierr = MatCoarsenSetType( crs, MATCOARSENMIS ); CHKERRQ(ierr); 682 ierr = MatCoarsenSetGreedyOrdering( crs, perm ); CHKERRQ(ierr); 683 ierr = MatCoarsenSetAdjacency( crs, Gmat ); CHKERRQ(ierr); 684 ierr = MatCoarsenSetVerbose( crs, pc_gamg->verbose ); CHKERRQ(ierr); 685 ierr = MatCoarsenSetStrictAggs( crs, PETSC_FALSE ); CHKERRQ(ierr); 686 ierr = MatCoarsenApply( crs ); CHKERRQ(ierr); 687 ierr = MatCoarsenGetMISAggLists( crs, &selected, &llist_parent ); CHKERRQ(ierr); 688 ierr = MatCoarsenDestroy( &crs ); CHKERRQ(ierr); 689 690 ierr = ISDestroy( &perm ); CHKERRQ(ierr); 691 692 *a_selected = selected; 693 *a_llist_parent = llist_parent; 694 695 PetscFunctionReturn(0); 696 } 697 698 /* -------------------------------------------------------------------------- */ 699 /* 700 PCGAMGprolongator_GEO 701 702 Input Parameter: 703 . pc - this 704 . Amat - matrix on this fine level 705 . Graph - used to get ghost data for nodes in 706 . selected_1 - [nselected inc. chosts] 707 . llist_1 - [nloc + Gmat.nghost] linked list 708 Output Parameter: 709 . a_P_out - prolongation operator to the next level 710 */ 711 #undef __FUNCT__ 712 #define __FUNCT__ "PCGAMGprolongator_GEO" 713 PetscErrorCode PCGAMGprolongator_GEO( PC pc, 714 const Mat Amat, 715 const Mat Gmat, 716 IS selected_1, 717 IS llist_1, 718 Mat *a_P_out 719 ) 720 { 721 PC_MG *mg = (PC_MG*)pc->data; 722 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 723 const PetscInt verbose = pc_gamg->verbose; 724 const PetscInt dim = pc_gamg->data_cell_cols, data_cols = pc_gamg->data_cell_cols; 725 PetscErrorCode ierr; 726 PetscInt Istart,Iend,nloc,my0,jj,kk,ncols,nLocalSelected,bs,*clid_flid; 727 Mat Prol; 728 PetscMPIInt mype, npe; 729 MPI_Comm wcomm = ((PetscObject)Amat)->comm; 730 IS selected_2; 731 const PetscInt *selected_idx; 732 733 PetscFunctionBegin; 734 ierr = MPI_Comm_rank(wcomm,&mype);CHKERRQ(ierr); 735 ierr = MPI_Comm_size(wcomm,&npe);CHKERRQ(ierr); 736 ierr = MatGetOwnershipRange( Amat, &Istart, &Iend ); CHKERRQ(ierr); 737 ierr = MatGetBlockSize( Amat, &bs ); CHKERRQ( ierr ); 738 nloc = (Iend-Istart)/bs; my0 = Istart/bs; assert((Iend-Istart)%bs==0); 739 740 /* get 'nLocalSelected' */ 741 ierr = ISGetSize( selected_1, &jj ); CHKERRQ(ierr); 742 ierr = PetscMalloc( jj*sizeof(PetscInt), &clid_flid ); CHKERRQ(ierr); 743 ierr = ISGetIndices( selected_1, &selected_idx ); CHKERRQ(ierr); 744 for(kk=0,nLocalSelected=0;kk<jj;kk++) { 745 PetscInt lid = selected_idx[kk]; 746 if( lid<nloc ) { 747 ierr = MatGetRow(Gmat,kk+my0,&ncols,0,0); CHKERRQ(ierr); 748 if( ncols>1 ) { /* fiter out singletons */ 749 clid_flid[nLocalSelected++] = lid; 750 } 751 ierr = MatRestoreRow(Gmat,kk+my0,&ncols,0,0); CHKERRQ(ierr); 752 } 753 } 754 ierr = ISRestoreIndices( selected_1, &selected_idx ); CHKERRQ(ierr); 755 756 /* create prolongator, create P matrix */ 757 ierr = MatCreateMPIAIJ(wcomm, 758 nloc*bs, nLocalSelected*bs, 759 PETSC_DETERMINE, PETSC_DETERMINE, 760 3*data_cols, PETSC_NULL, /* don't have a good way to set this!!! */ 761 3*data_cols, PETSC_NULL, 762 &Prol ); 763 CHKERRQ(ierr); 764 765 /* can get all points "removed" - but not on geomg */ 766 ierr = MatGetSize( Prol, &kk, &jj ); CHKERRQ(ierr); 767 if( jj==0 ) { 768 if( verbose ) { 769 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s ERROE: no selected points on coarse grid\n",mype,__FUNCT__); 770 } 771 ierr = PetscFree( clid_flid ); CHKERRQ(ierr); 772 ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 773 *a_P_out = PETSC_NULL; /* out */ 774 PetscFunctionReturn(0); 775 } 776 777 { 778 PetscReal *coords; 779 PetscInt nnodes; 780 PetscInt *crsGID; 781 Mat Gmat2; 782 783 assert(dim==data_cols); 784 /* grow ghost data for better coarse grid cover of fine grid */ 785 #if defined PETSC_USE_LOG 786 ierr = PetscLogEventBegin(gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr); 787 #endif 788 ierr = getGIDsOnSquareGraph( nLocalSelected, clid_flid, Gmat, &selected_2, &Gmat2, &crsGID ); 789 CHKERRQ(ierr); 790 /* llist is now not valid wrt squared graph, but will work as iterator in 'triangulateAndFormProl' */ 791 #if defined PETSC_USE_LOG 792 ierr = PetscLogEventEnd(gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr); 793 #endif 794 /* create global vector of coorindates in 'coords' */ 795 if (npe > 1) { 796 ierr = getDataWithGhosts( Gmat2, dim, pc_gamg->data, &nnodes, &coords ); 797 CHKERRQ(ierr); 798 } 799 else { 800 coords = (PetscReal*)pc_gamg->data; 801 nnodes = nloc; 802 } 803 ierr = MatDestroy( &Gmat2 ); CHKERRQ(ierr); 804 805 /* triangulate */ 806 if( dim == 2 ) { 807 PetscReal metric,tm; 808 #if defined PETSC_USE_LOG 809 ierr = PetscLogEventBegin(gamg_setup_events[SET6],0,0,0,0);CHKERRQ(ierr); 810 #endif 811 ierr = triangulateAndFormProl( selected_2, nnodes, coords, 812 nLocalSelected, clid_flid, llist_1, crsGID, bs, Prol, &metric ); 813 CHKERRQ(ierr); 814 #if defined PETSC_USE_LOG 815 ierr = PetscLogEventEnd(gamg_setup_events[SET6],0,0,0,0); CHKERRQ(ierr); 816 #endif 817 ierr = PetscFree( crsGID ); CHKERRQ(ierr); 818 819 /* clean up and create coordinates for coarse grid (output) */ 820 if (npe > 1) ierr = PetscFree( coords ); CHKERRQ(ierr); 821 822 ierr = MPI_Allreduce( &metric, &tm, 1, MPIU_REAL, MPIU_MAX, wcomm ); CHKERRQ(ierr); 823 if( tm > 1. ) { /* needs to be globalized - should not happen */ 824 if( verbose ) { 825 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s failed metric for coarse grid %e\n",mype,__FUNCT__,tm); 826 } 827 ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 828 Prol = PETSC_NULL; 829 } 830 else if( metric > .0 ) { 831 if( verbose ) { 832 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s worst metric for coarse grid = %e\n",mype,__FUNCT__,metric); 833 } 834 } 835 } else { 836 SETERRQ(wcomm,PETSC_ERR_LIB,"3D not implemented for 'geo' AMG"); 837 } 838 { /* create next coords - output */ 839 PetscReal *crs_crds; 840 ierr = PetscMalloc( dim*nLocalSelected*sizeof(PetscReal), &crs_crds ); 841 CHKERRQ(ierr); 842 for(kk=0;kk<nLocalSelected;kk++){/* grab local select nodes to promote - output */ 843 PetscInt lid = clid_flid[kk]; 844 for(jj=0;jj<dim;jj++) crs_crds[jj*nLocalSelected + kk] = pc_gamg->data[jj*nloc + lid]; 845 } 846 847 ierr = PetscFree( pc_gamg->data ); CHKERRQ( ierr ); 848 pc_gamg->data = crs_crds; /* out */ 849 pc_gamg->data_sz = dim*nLocalSelected; 850 } 851 ierr = ISDestroy( &selected_2 ); CHKERRQ(ierr); /* this is selected_1 in serial */ 852 } 853 *a_P_out = Prol; /* out */ 854 ierr = PetscFree( clid_flid ); CHKERRQ(ierr); 855 856 PetscFunctionReturn(0); 857 } 858 859 /* -------------------------------------------------------------------------- */ 860 /* 861 PCCreateGAMG_GEO 862 863 Input Parameter: 864 . pc - 865 */ 866 #undef __FUNCT__ 867 #define __FUNCT__ "PCCreateGAMG_GEO" 868 PetscErrorCode PCCreateGAMG_GEO( PC pc ) 869 { 870 PetscErrorCode ierr; 871 PC_MG *mg = (PC_MG*)pc->data; 872 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 873 874 PetscFunctionBegin; 875 pc->ops->setfromoptions = PCSetFromOptions_GEO; 876 /* pc->ops->destroy = PCDestroy_GEO; */ 877 /* reset does not do anything; setup not virtual */ 878 879 /* set internal function pointers */ 880 pc_gamg->graph = PCGAMGgraph_GEO; 881 pc_gamg->coarsen = PCGAMGcoarsen_GEO; 882 pc_gamg->prolongator = PCGAMGprolongator_GEO; 883 pc_gamg->optprol = 0; 884 885 pc_gamg->createdefaultdata = PCSetData_GEO; 886 887 ierr = PetscObjectComposeFunctionDynamic( (PetscObject)pc, 888 "PCSetCoordinates_C", 889 "PCSetCoordinates_GEO", 890 PCSetCoordinates_GEO);CHKERRQ(ierr); 891 892 PetscFunctionReturn(0); 893 } 894