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 <petsc-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 petsc_geo_mg_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 . agg_lists_1 - list of aggregates 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 static 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 const PetscCoarsenData *agg_lists_1, /* 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; 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_GAMG_USE_LOG 308 ierr = PetscLogEventBegin(petsc_gamg_setup_events[FIND_V],0,0,0,0);CHKERRQ(ierr); 309 #endif 310 { /* form P - setup some maps */ 311 PetscInt clid,mm,*nTri,*node_tri; 312 313 ierr = PetscMalloc( nselected_2*sizeof(PetscInt), &node_tri ); CHKERRQ(ierr); 314 ierr = PetscMalloc( nselected_2*sizeof(PetscInt), &nTri ); CHKERRQ(ierr); 315 316 /* need list of triangles on node */ 317 for(kk=0;kk<nselected_2;kk++) nTri[kk] = 0; 318 for(tid=0,kk=0;tid<mid.numberoftriangles;tid++){ 319 for(jj=0;jj<3;jj++) { 320 PetscInt cid = mid.trianglelist[kk++]; 321 if( nTri[cid] == 0 ) node_tri[cid] = tid; 322 nTri[cid]++; 323 } 324 } 325 #define EPS 1.e-12 326 /* find points and set prolongation */ 327 for( mm = clid = 0 ; mm < nFineLoc ; mm++ ){ 328 if( (jj=AILSizeAt(agg_lists_1,mm)) > 0 ) { 329 const PetscInt lid = mm; 330 //for(clid_iterator=0;clid_iterator<nselected_1;clid_iterator++){ 331 //PetscInt flid = clid_lid_1[clid_iterator]; assert(flid != -1); 332 PetscScalar AA[3][3]; 333 PetscBLASInt N=3,NRHS=1,LDA=3,IPIV[3],LDB=3,INFO; 334 LLNPos pos; 335 336 for( pos=AILGetHeadPos(agg_lists_1,lid) ; 337 pos ; 338 pos=AILGetNextPos(agg_lists_1,lid,pos) ){ 339 PetscInt flid = LLNGetID(pos); 340 if( flid < nFineLoc ) { /* could be a ghost */ 341 PetscInt bestTID = -1; PetscReal best_alpha = 1.e10; 342 const PetscInt fgid = flid + myFine0; 343 /* compute shape function for gid */ 344 const PetscReal fcoord[3] = {coords[flid],coords[nnodes+flid],1.0}; 345 PetscBool haveit=PETSC_FALSE; PetscScalar alpha[3]; PetscInt clids[3]; 346 /* look for it */ 347 for( tid = node_tri[clid], jj=0; 348 jj < 5 && !haveit && tid != -1; 349 jj++ ){ 350 for(tt=0;tt<3;tt++){ 351 PetscInt cid2 = mid.trianglelist[3*tid + tt]; 352 PetscInt lid2 = selected_idx_2[cid2]; 353 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 354 clids[tt] = cid2; /* store for interp */ 355 } 356 357 for(tt=0;tt<3;tt++) alpha[tt] = (PetscScalar)fcoord[tt]; 358 359 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 360 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 361 { 362 PetscBool have=PETSC_TRUE; PetscReal lowest=1.e10; 363 for( tt = 0, idx = 0 ; tt < 3 ; tt++ ) { 364 if( PetscRealPart(alpha[tt]) > (1.0+EPS) || PetscRealPart(alpha[tt]) < -EPS ) have = PETSC_FALSE; 365 if( PetscRealPart(alpha[tt]) < lowest ){ 366 lowest = PetscRealPart(alpha[tt]); 367 idx = tt; 368 } 369 } 370 haveit = have; 371 } 372 tid = mid.neighborlist[3*tid + idx]; 373 } 374 375 if( !haveit ) { 376 /* brute force */ 377 for(tid=0 ; tid<mid.numberoftriangles && !haveit ; tid++ ){ 378 for(tt=0;tt<3;tt++){ 379 PetscInt cid2 = mid.trianglelist[3*tid + tt]; 380 PetscInt lid2 = selected_idx_2[cid2]; 381 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 382 clids[tt] = cid2; /* store for interp */ 383 } 384 for(tt=0;tt<3;tt++) alpha[tt] = fcoord[tt]; 385 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 386 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 387 { 388 PetscBool have=PETSC_TRUE; PetscReal worst=0.0, v; 389 for(tt=0; tt<3 && have ;tt++) { 390 if( PetscRealPart(alpha[tt]) > 1.0+EPS || PetscRealPart(alpha[tt]) < -EPS ) have=PETSC_FALSE; 391 if( (v=PetscAbs(PetscRealPart(alpha[tt])-0.5)) > worst ) worst = v; 392 } 393 if( worst < best_alpha ) { 394 best_alpha = worst; bestTID = tid; 395 } 396 haveit = have; 397 } 398 } 399 } 400 if( !haveit ) { 401 if( best_alpha > *a_worst_best ) *a_worst_best = best_alpha; 402 /* use best one */ 403 for(tt=0;tt<3;tt++){ 404 PetscInt cid2 = mid.trianglelist[3*bestTID + tt]; 405 PetscInt lid2 = selected_idx_2[cid2]; 406 AA[tt][0] = coords[lid2]; AA[tt][1] = coords[nnodes + lid2]; AA[tt][2] = 1.0; 407 clids[tt] = cid2; /* store for interp */ 408 } 409 for(tt=0;tt<3;tt++) alpha[tt] = fcoord[tt]; 410 /* SUBROUTINE DGESV( N, NRHS, A, LDA, IPIV, B, LDB, INFO ) */ 411 LAPACKgesv_(&N, &NRHS, (PetscScalar*)AA, &LDA, IPIV, alpha, &LDB, &INFO); 412 } 413 414 /* put in row of P */ 415 for(idx=0;idx<3;idx++){ 416 PetscScalar shp = alpha[idx]; 417 if( PetscAbs(PetscRealPart(shp)) > 1.e-6 ) { 418 PetscInt cgid = crsGID[clids[idx]]; 419 PetscInt jj = cgid*bs, ii = fgid*bs; /* need to gloalize */ 420 for(tt=0 ; tt < bs ; tt++, ii++, jj++ ){ 421 ierr = MatSetValues(a_Prol,1,&ii,1,&jj,&shp,INSERT_VALUES); CHKERRQ(ierr); 422 } 423 } 424 } 425 } 426 } /* aggregates iterations */ 427 clid++; 428 } /* a coarse agg */ 429 } /* for all fine nodes */ 430 431 ierr = ISRestoreIndices( selected_2, &selected_idx_2 ); CHKERRQ(ierr); 432 ierr = MatAssemblyBegin(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 433 ierr = MatAssemblyEnd(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 434 435 ierr = PetscFree( node_tri ); CHKERRQ(ierr); 436 ierr = PetscFree( nTri ); CHKERRQ(ierr); 437 } 438 #if defined PETSC_GAMG_USE_LOG 439 ierr = PetscLogEventEnd(petsc_gamg_setup_events[FIND_V],0,0,0,0);CHKERRQ(ierr); 440 #endif 441 free( mid.trianglelist ); 442 free( mid.neighborlist ); 443 ierr = PetscFree( in.pointlist ); CHKERRQ(ierr); 444 445 PetscFunctionReturn(0); 446 #else 447 SETERRQ(((PetscObject)a_Prol)->comm,PETSC_ERR_LIB,"configure with TRIANGLE to use geometric MG"); 448 #endif 449 } 450 /* -------------------------------------------------------------------------- */ 451 /* 452 getGIDsOnSquareGraph - square graph, get 453 454 Input Parameter: 455 . nselected_1 - selected local indices (includes ghosts in input Gmat1) 456 . clid_lid_1 - [nselected_1] lids of selected nodes 457 . Gmat1 - graph that goes with 'selected_1' 458 Output Parameter: 459 . a_selected_2 - selected local indices (includes ghosts in output a_Gmat_2) 460 . a_Gmat_2 - graph that is squared of 'Gmat_1' 461 . a_crsGID[a_selected_2.size()] - map of global IDs of coarse grid nodes 462 */ 463 #undef __FUNCT__ 464 #define __FUNCT__ "getGIDsOnSquareGraph" 465 static PetscErrorCode getGIDsOnSquareGraph( const PetscInt nselected_1, 466 const PetscInt clid_lid_1[], 467 const Mat Gmat1, 468 IS *a_selected_2, 469 Mat *a_Gmat_2, 470 PetscInt **a_crsGID 471 ) 472 { 473 PetscErrorCode ierr; 474 PetscMPIInt mype,npe; 475 PetscInt *crsGID, kk,my0,Iend,nloc; 476 MPI_Comm wcomm = ((PetscObject)Gmat1)->comm; 477 478 PetscFunctionBegin; 479 ierr = MPI_Comm_rank(wcomm,&mype);CHKERRQ(ierr); 480 ierr = MPI_Comm_size(wcomm,&npe);CHKERRQ(ierr); 481 ierr = MatGetOwnershipRange(Gmat1,&my0,&Iend); CHKERRQ(ierr); /* AIJ */ 482 nloc = Iend - my0; /* this does not change */ 483 484 if (npe == 1) { /* not much to do in serial */ 485 ierr = PetscMalloc( nselected_1*sizeof(PetscInt), &crsGID ); CHKERRQ(ierr); 486 for(kk=0;kk<nselected_1;kk++) crsGID[kk] = kk; 487 *a_Gmat_2 = 0; 488 ierr = ISCreateGeneral(PETSC_COMM_SELF,nselected_1,clid_lid_1,PETSC_COPY_VALUES,a_selected_2); 489 CHKERRQ(ierr); 490 } 491 else { 492 PetscInt idx,num_fine_ghosts,num_crs_ghost,myCrs0; 493 Mat_MPIAIJ *mpimat2; 494 Mat Gmat2; 495 Vec locState; 496 PetscScalar *cpcol_state; 497 498 /* scan my coarse zero gid, set 'lid_state' with coarse GID */ 499 kk = nselected_1; 500 MPI_Scan( &kk, &myCrs0, 1, MPIU_INT, MPIU_SUM, wcomm ); 501 myCrs0 -= nselected_1; 502 503 if( a_Gmat_2 ) { /* output */ 504 /* grow graph to get wider set of selected vertices to cover fine grid, invalidates 'llist' */ 505 ierr = MatTransposeMatMult(Gmat1, Gmat1, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Gmat2 ); CHKERRQ(ierr); 506 *a_Gmat_2 = Gmat2; /* output */ 507 } 508 else Gmat2 = Gmat1; /* use local to get crsGIDs at least */ 509 /* get coarse grid GIDS for selected (locals and ghosts) */ 510 mpimat2 = (Mat_MPIAIJ*)Gmat2->data; 511 ierr = MatGetVecs( Gmat2, &locState, 0 ); CHKERRQ(ierr); 512 ierr = VecSet( locState, (PetscScalar)(PetscReal)(-1) ); CHKERRQ(ierr); /* set with UNKNOWN state */ 513 for(kk=0;kk<nselected_1;kk++){ 514 PetscInt fgid = clid_lid_1[kk] + my0; 515 PetscScalar v = (PetscScalar)(kk+myCrs0); 516 ierr = VecSetValues( locState, 1, &fgid, &v, INSERT_VALUES ); CHKERRQ(ierr); /* set with PID */ 517 } 518 ierr = VecAssemblyBegin( locState ); CHKERRQ(ierr); 519 ierr = VecAssemblyEnd( locState ); CHKERRQ(ierr); 520 ierr = VecScatterBegin(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 521 ierr = VecScatterEnd(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 522 ierr = VecGetLocalSize( mpimat2->lvec, &num_fine_ghosts ); CHKERRQ(ierr); 523 ierr = VecGetArray( mpimat2->lvec, &cpcol_state ); CHKERRQ(ierr); 524 for(kk=0,num_crs_ghost=0;kk<num_fine_ghosts;kk++){ 525 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ) num_crs_ghost++; 526 } 527 ierr = PetscMalloc( (nselected_1+num_crs_ghost)*sizeof(PetscInt), &crsGID ); CHKERRQ(ierr); /* output */ 528 { 529 PetscInt *selected_set; 530 ierr = PetscMalloc( (nselected_1+num_crs_ghost)*sizeof(PetscInt), &selected_set ); CHKERRQ(ierr); 531 /* do ghost of 'crsGID' */ 532 for(kk=0,idx=nselected_1;kk<num_fine_ghosts;kk++){ 533 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ){ 534 PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]); 535 selected_set[idx] = nloc + kk; 536 crsGID[idx++] = cgid; 537 } 538 } 539 assert(idx==(nselected_1+num_crs_ghost)); 540 ierr = VecRestoreArray( mpimat2->lvec, &cpcol_state ); CHKERRQ(ierr); 541 /* do locals in 'crsGID' */ 542 ierr = VecGetArray( locState, &cpcol_state ); CHKERRQ(ierr); 543 for(kk=0,idx=0;kk<nloc;kk++){ 544 if( (PetscInt)PetscRealPart(cpcol_state[kk]) != -1 ){ 545 PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]); 546 selected_set[idx] = kk; 547 crsGID[idx++] = cgid; 548 } 549 } 550 assert(idx==nselected_1); 551 ierr = VecRestoreArray( locState, &cpcol_state ); CHKERRQ(ierr); 552 553 if( a_selected_2 != 0 ) { /* output */ 554 ierr = ISCreateGeneral(PETSC_COMM_SELF,(nselected_1+num_crs_ghost),selected_set,PETSC_OWN_POINTER,a_selected_2); 555 CHKERRQ(ierr); 556 } 557 else { 558 ierr = PetscFree( selected_set ); CHKERRQ(ierr); 559 } 560 } 561 ierr = VecDestroy( &locState ); CHKERRQ(ierr); 562 } 563 *a_crsGID = crsGID; /* output */ 564 565 PetscFunctionReturn(0); 566 } 567 568 /* -------------------------------------------------------------------------- */ 569 /* 570 PCGAMGgraph_GEO 571 572 Input Parameter: 573 . pc - this 574 . Amat - matrix on this fine level 575 Output Parameter: 576 . a_Gmat 577 */ 578 #undef __FUNCT__ 579 #define __FUNCT__ "PCGAMGgraph_GEO" 580 PetscErrorCode PCGAMGgraph_GEO( PC pc, 581 const Mat Amat, 582 Mat *a_Gmat 583 ) 584 { 585 PetscErrorCode ierr; 586 PC_MG *mg = (PC_MG*)pc->data; 587 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 588 const PetscInt verbose = pc_gamg->verbose; 589 const PetscReal vfilter = pc_gamg->threshold; 590 PetscMPIInt mype,npe; 591 MPI_Comm wcomm = ((PetscObject)Amat)->comm; 592 Mat Gmat; 593 PetscBool set,flg,symm; 594 PetscFunctionBegin; 595 #if defined PETSC_USE_LOG 596 ierr = PetscLogEventBegin(PC_GAMGGgraph_GEO,0,0,0,0);CHKERRQ(ierr); 597 #endif 598 ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 599 ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 600 601 ierr = MatIsSymmetricKnown(Amat, &set, &flg); CHKERRQ(ierr); 602 symm = !(set && flg); 603 604 ierr = PCGAMGCreateSimpleGraph( Amat, &Gmat ); CHKERRQ( ierr ); 605 ierr = PCGAMGScaleFilterGraph( &Gmat, vfilter, symm, verbose ); CHKERRQ( ierr ); 606 607 *a_Gmat = Gmat; 608 #if defined PETSC_USE_LOG 609 ierr = PetscLogEventEnd(PC_GAMGGgraph_GEO,0,0,0,0);CHKERRQ(ierr); 610 #endif 611 PetscFunctionReturn(0); 612 } 613 614 /* -------------------------------------------------------------------------- */ 615 /* 616 PCGAMGcoarsen_GEO 617 618 Input Parameter: 619 . a_pc - this 620 . a_Gmat - graph 621 Output Parameter: 622 . a_llist_parent - linked list from selected indices for data locality only 623 */ 624 #undef __FUNCT__ 625 #define __FUNCT__ "PCGAMGcoarsen_GEO" 626 PetscErrorCode PCGAMGcoarsen_GEO( PC a_pc, 627 Mat *a_Gmat, 628 PetscCoarsenData **a_llist_parent 629 ) 630 { 631 PetscErrorCode ierr; 632 PC_MG *mg = (PC_MG*)a_pc->data; 633 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 634 PetscInt Istart,Iend,nloc,kk,Ii,ncols; 635 PetscMPIInt mype,npe; 636 IS perm; 637 GAMGNode *gnodes; 638 PetscInt *permute; 639 Mat Gmat = *a_Gmat; 640 MPI_Comm wcomm = ((PetscObject)Gmat)->comm; 641 MatCoarsen crs; 642 643 PetscFunctionBegin; 644 #if defined PETSC_USE_LOG 645 ierr = PetscLogEventBegin(PC_GAMGCoarsen_GEO,0,0,0,0);CHKERRQ(ierr); 646 #endif 647 ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 648 ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 649 ierr = MatGetOwnershipRange( Gmat, &Istart, &Iend ); CHKERRQ(ierr); 650 nloc = (Iend-Istart); 651 652 /* create random permutation with sort for geo-mg */ 653 ierr = PetscMalloc( nloc*sizeof(GAMGNode), &gnodes ); CHKERRQ(ierr); 654 ierr = PetscMalloc( nloc*sizeof(PetscInt), &permute ); CHKERRQ(ierr); 655 656 for (Ii=Istart; Ii<Iend; Ii++) { /* locals only? */ 657 ierr = MatGetRow(Gmat,Ii,&ncols,0,0); CHKERRQ(ierr); 658 { 659 PetscInt lid = Ii - Istart; 660 gnodes[lid].lid = lid; 661 gnodes[lid].degree = ncols; 662 } 663 ierr = MatRestoreRow(Gmat,Ii,&ncols,0,0); CHKERRQ(ierr); 664 } 665 /* randomize */ 666 srand(1); /* make deterministic */ 667 if( PETSC_TRUE ) { 668 PetscBool *bIndexSet; 669 ierr = PetscMalloc( nloc*sizeof(PetscBool), &bIndexSet ); CHKERRQ(ierr); 670 for ( Ii = 0; Ii < nloc ; Ii++) bIndexSet[Ii] = PETSC_FALSE; 671 for ( Ii = 0; Ii < nloc ; Ii++) 672 { 673 PetscInt iSwapIndex = rand()%nloc; 674 if (!bIndexSet[iSwapIndex] && iSwapIndex != Ii) 675 { 676 GAMGNode iTemp = gnodes[iSwapIndex]; 677 gnodes[iSwapIndex] = gnodes[Ii]; 678 gnodes[Ii] = iTemp; 679 bIndexSet[Ii] = PETSC_TRUE; 680 bIndexSet[iSwapIndex] = PETSC_TRUE; 681 } 682 } 683 ierr = PetscFree( bIndexSet ); CHKERRQ(ierr); 684 } 685 /* only sort locals */ 686 qsort( gnodes, nloc, sizeof(GAMGNode), petsc_geo_mg_compare ); 687 /* create IS of permutation */ 688 for(kk=0;kk<nloc;kk++) { /* locals only */ 689 permute[kk] = gnodes[kk].lid; 690 } 691 ierr = ISCreateGeneral(PETSC_COMM_SELF, nloc, permute, PETSC_OWN_POINTER, &perm); 692 CHKERRQ(ierr); 693 694 ierr = PetscFree( gnodes ); CHKERRQ(ierr); 695 696 /* get MIS aggs */ 697 698 ierr = MatCoarsenCreate( wcomm, &crs ); CHKERRQ(ierr); 699 ierr = MatCoarsenSetType( crs, MATCOARSENMIS ); CHKERRQ(ierr); 700 ierr = MatCoarsenSetGreedyOrdering( crs, perm ); CHKERRQ(ierr); 701 ierr = MatCoarsenSetAdjacency( crs, Gmat ); CHKERRQ(ierr); 702 ierr = MatCoarsenSetVerbose( crs, pc_gamg->verbose ); CHKERRQ(ierr); 703 ierr = MatCoarsenSetStrictAggs( crs, PETSC_FALSE ); CHKERRQ(ierr); 704 ierr = MatCoarsenApply( crs ); CHKERRQ(ierr); 705 ierr = MatCoarsenGetData( crs, a_llist_parent ); CHKERRQ(ierr); 706 ierr = MatCoarsenDestroy( &crs ); CHKERRQ(ierr); 707 708 ierr = ISDestroy( &perm ); CHKERRQ(ierr); 709 #if defined PETSC_USE_LOG 710 ierr = PetscLogEventEnd(PC_GAMGCoarsen_GEO,0,0,0,0);CHKERRQ(ierr); 711 #endif 712 PetscFunctionReturn(0); 713 } 714 715 /* -------------------------------------------------------------------------- */ 716 /* 717 PCGAMGProlongator_GEO 718 719 Input Parameter: 720 . pc - this 721 . Amat - matrix on this fine level 722 . Graph - used to get ghost data for nodes in 723 . selected_1 - [nselected] 724 . agg_lists - [nselected] 725 Output Parameter: 726 . a_P_out - prolongation operator to the next level 727 */ 728 #undef __FUNCT__ 729 #define __FUNCT__ "PCGAMGProlongator_GEO" 730 PetscErrorCode PCGAMGProlongator_GEO( PC pc, 731 const Mat Amat, 732 const Mat Gmat, 733 PetscCoarsenData *agg_lists, 734 Mat *a_P_out 735 ) 736 { 737 PC_MG *mg = (PC_MG*)pc->data; 738 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 739 const PetscInt verbose = pc_gamg->verbose; 740 const PetscInt dim = pc_gamg->data_cell_cols, data_cols = pc_gamg->data_cell_cols; 741 PetscErrorCode ierr; 742 PetscInt Istart,Iend,nloc,my0,jj,kk,ncols,nLocalSelected,bs,*clid_flid; 743 Mat Prol; 744 PetscMPIInt mype, npe; 745 MPI_Comm wcomm = ((PetscObject)Amat)->comm; 746 IS selected_2,selected_1; 747 const PetscInt *selected_idx; 748 749 PetscFunctionBegin; 750 #if defined PETSC_USE_LOG 751 ierr = PetscLogEventBegin(PC_GAMGProlongator_GEO,0,0,0,0);CHKERRQ(ierr); 752 #endif 753 ierr = MPI_Comm_rank(wcomm,&mype);CHKERRQ(ierr); 754 ierr = MPI_Comm_size(wcomm,&npe);CHKERRQ(ierr); 755 ierr = MatGetOwnershipRange( Amat, &Istart, &Iend ); CHKERRQ(ierr); 756 ierr = MatGetBlockSize( Amat, &bs ); CHKERRQ( ierr ); 757 nloc = (Iend-Istart)/bs; my0 = Istart/bs; assert((Iend-Istart)%bs==0); 758 759 /* get 'nLocalSelected' */ 760 ierr = AILGetMIS( agg_lists, &selected_1 ); CHKERRQ(ierr); 761 ierr = ISGetSize( selected_1, &jj ); CHKERRQ(ierr); 762 ierr = PetscMalloc( jj*sizeof(PetscInt), &clid_flid ); CHKERRQ(ierr); 763 ierr = ISGetIndices( selected_1, &selected_idx ); CHKERRQ(ierr); 764 for(kk=0,nLocalSelected=0;kk<jj;kk++) { 765 PetscInt lid = selected_idx[kk]; 766 if( lid<nloc ) { 767 ierr = MatGetRow(Gmat,lid+my0,&ncols,0,0); CHKERRQ(ierr); 768 if( ncols>1 ) { /* fiter out singletons */ 769 clid_flid[nLocalSelected++] = lid; 770 } 771 else assert(0); /* filtered in coarsening */ 772 ierr = MatRestoreRow(Gmat,lid+my0,&ncols,0,0); CHKERRQ(ierr); 773 } 774 } 775 ierr = ISRestoreIndices( selected_1, &selected_idx ); CHKERRQ(ierr); 776 777 /* create prolongator, create P matrix */ 778 ierr = MatCreateAIJ(wcomm, 779 nloc*bs, nLocalSelected*bs, 780 PETSC_DETERMINE, PETSC_DETERMINE, 781 3*data_cols, PETSC_NULL, /* don't have a good way to set this!!! */ 782 3*data_cols, PETSC_NULL, 783 &Prol ); 784 CHKERRQ(ierr); 785 786 /* can get all points "removed" - but not on geomg */ 787 ierr = MatGetSize( Prol, &kk, &jj ); CHKERRQ(ierr); 788 if( jj==0 ) { 789 if( verbose ) { 790 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s ERROE: no selected points on coarse grid\n",mype,__FUNCT__); 791 } 792 ierr = PetscFree( clid_flid ); CHKERRQ(ierr); 793 ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 794 *a_P_out = PETSC_NULL; /* out */ 795 PetscFunctionReturn(0); 796 } 797 798 { 799 PetscReal *coords; 800 PetscInt nnodes; 801 PetscInt *crsGID; 802 Mat Gmat2; 803 804 assert(dim==data_cols); 805 /* grow ghost data for better coarse grid cover of fine grid */ 806 #if defined PETSC_GAMG_USE_LOG 807 ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr); 808 #endif 809 ierr = getGIDsOnSquareGraph( nLocalSelected, clid_flid, Gmat, &selected_2, &Gmat2, &crsGID ); 810 CHKERRQ(ierr); 811 /* llist is now not valid wrt squared graph, but will work as iterator in 'triangulateAndFormProl' */ 812 #if defined PETSC_GAMG_USE_LOG 813 ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr); 814 #endif 815 /* create global vector of coorindates in 'coords' */ 816 if (npe > 1) { 817 ierr = PCGAMGGetDataWithGhosts( Gmat2, dim, pc_gamg->data, &nnodes, &coords ); 818 CHKERRQ(ierr); 819 } 820 else { 821 coords = (PetscReal*)pc_gamg->data; 822 nnodes = nloc; 823 } 824 ierr = MatDestroy( &Gmat2 ); CHKERRQ(ierr); 825 826 /* triangulate */ 827 if( dim == 2 ) { 828 PetscReal metric,tm; 829 #if defined PETSC_GAMG_USE_LOG 830 ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET6],0,0,0,0);CHKERRQ(ierr); 831 #endif 832 ierr = triangulateAndFormProl( selected_2, nnodes, coords, 833 nLocalSelected, clid_flid, agg_lists, crsGID, bs, Prol, &metric ); 834 CHKERRQ(ierr); 835 #if defined PETSC_GAMG_USE_LOG 836 ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET6],0,0,0,0); CHKERRQ(ierr); 837 #endif 838 ierr = PetscFree( crsGID ); CHKERRQ(ierr); 839 840 /* clean up and create coordinates for coarse grid (output) */ 841 if (npe > 1) ierr = PetscFree( coords ); CHKERRQ(ierr); 842 843 ierr = MPI_Allreduce( &metric, &tm, 1, MPIU_REAL, MPIU_MAX, wcomm ); CHKERRQ(ierr); 844 if( tm > 1. ) { /* needs to be globalized - should not happen */ 845 if( verbose ) { 846 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s failed metric for coarse grid %e\n",mype,__FUNCT__,tm); 847 } 848 ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 849 Prol = PETSC_NULL; 850 } 851 else if( metric > .0 ) { 852 if( verbose ) { 853 PetscPrintf(PETSC_COMM_WORLD,"[%d]%s worst metric for coarse grid = %e\n",mype,__FUNCT__,metric); 854 } 855 } 856 } else { 857 SETERRQ(wcomm,PETSC_ERR_LIB,"3D not implemented for 'geo' AMG"); 858 } 859 { /* create next coords - output */ 860 PetscReal *crs_crds; 861 ierr = PetscMalloc( dim*nLocalSelected*sizeof(PetscReal), &crs_crds ); 862 CHKERRQ(ierr); 863 for(kk=0;kk<nLocalSelected;kk++){/* grab local select nodes to promote - output */ 864 PetscInt lid = clid_flid[kk]; 865 for(jj=0;jj<dim;jj++) crs_crds[jj*nLocalSelected + kk] = pc_gamg->data[jj*nloc + lid]; 866 } 867 868 ierr = PetscFree( pc_gamg->data ); CHKERRQ( ierr ); 869 pc_gamg->data = crs_crds; /* out */ 870 pc_gamg->data_sz = dim*nLocalSelected; 871 } 872 ierr = ISDestroy( &selected_2 ); CHKERRQ(ierr); /* this is selected_1 in serial */ 873 } 874 *a_P_out = Prol; /* out */ 875 ierr = PetscFree( clid_flid ); CHKERRQ(ierr); 876 #if defined PETSC_USE_LOG 877 ierr = PetscLogEventEnd(PC_GAMGProlongator_GEO,0,0,0,0);CHKERRQ(ierr); 878 #endif 879 PetscFunctionReturn(0); 880 } 881 882 /* -------------------------------------------------------------------------- */ 883 /* 884 PCCreateGAMG_GEO 885 886 Input Parameter: 887 . pc - 888 */ 889 #undef __FUNCT__ 890 #define __FUNCT__ "PCCreateGAMG_GEO" 891 PetscErrorCode PCCreateGAMG_GEO( PC pc ) 892 { 893 PetscErrorCode ierr; 894 PC_MG *mg = (PC_MG*)pc->data; 895 PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 896 897 PetscFunctionBegin; 898 pc->ops->setfromoptions = PCSetFromOptions_GEO; 899 /* pc->ops->destroy = PCDestroy_GEO; */ 900 /* reset does not do anything; setup not virtual */ 901 902 /* set internal function pointers */ 903 pc_gamg->graph = PCGAMGgraph_GEO; 904 pc_gamg->coarsen = PCGAMGcoarsen_GEO; 905 pc_gamg->prolongator = PCGAMGProlongator_GEO; 906 pc_gamg->optprol = 0; 907 908 pc_gamg->createdefaultdata = PCSetData_GEO; 909 910 ierr = PetscObjectComposeFunctionDynamic( (PetscObject)pc, 911 "PCSetCoordinates_C", 912 "PCSetCoordinates_GEO", 913 PCSetCoordinates_GEO);CHKERRQ(ierr); 914 915 PetscFunctionReturn(0); 916 } 917