12e68589bSMark F. Adams /* 22e68589bSMark F. Adams GAMG geometric-algebric multiogrid PC - Mark Adams 2011 32e68589bSMark F. Adams */ 42e68589bSMark F. Adams 52e68589bSMark F. Adams #include <../src/ksp/pc/impls/gamg/gamg.h> /*I "petscpc.h" I*/ 6b45d2f2cSJed Brown #include <petsc-private/kspimpl.h> 72e68589bSMark F. Adams 82e68589bSMark F. Adams #include <assert.h> 92e68589bSMark F. Adams #include <petscblaslapack.h> 102e68589bSMark F. Adams 112e68589bSMark F. Adams typedef struct { 12c8b0795cSMark F. Adams PetscInt nsmooths; 13c8b0795cSMark F. Adams PetscBool sym_graph; 14ef4ad70eSMark F. Adams PetscBool square_graph; 152e68589bSMark F. Adams }PC_GAMG_AGG; 162e68589bSMark F. Adams 172e68589bSMark F. Adams #undef __FUNCT__ 182e68589bSMark F. Adams #define __FUNCT__ "PCGAMGSetNSmooths" 192e68589bSMark F. Adams /*@ 202e68589bSMark F. Adams PCGAMGSetNSmooths - Set number of smoothing steps (1 is typical) 212e68589bSMark F. Adams 222e68589bSMark F. Adams Not Collective on PC 232e68589bSMark F. Adams 242e68589bSMark F. Adams Input Parameters: 252e68589bSMark F. Adams . pc - the preconditioner context 262e68589bSMark F. Adams 272e68589bSMark F. Adams Options Database Key: 282e68589bSMark F. Adams . -pc_gamg_agg_nsmooths 292e68589bSMark F. Adams 302e68589bSMark F. Adams Level: intermediate 312e68589bSMark F. Adams 322e68589bSMark F. Adams Concepts: Aggregation AMG preconditioner 332e68589bSMark F. Adams 342e68589bSMark F. Adams .seealso: () 352e68589bSMark F. Adams @*/ 362e68589bSMark F. Adams PetscErrorCode PCGAMGSetNSmooths(PC pc, PetscInt n) 372e68589bSMark F. Adams { 382e68589bSMark F. Adams PetscErrorCode ierr; 392e68589bSMark F. Adams 402e68589bSMark F. Adams PetscFunctionBegin; 412e68589bSMark F. Adams PetscValidHeaderSpecific(pc,PC_CLASSID,1); 422e68589bSMark F. Adams ierr = PetscTryMethod(pc,"PCGAMGSetNSmooths_C",(PC,PetscInt),(pc,n));CHKERRQ(ierr); 432e68589bSMark F. Adams PetscFunctionReturn(0); 442e68589bSMark F. Adams } 452e68589bSMark F. Adams 462e68589bSMark F. Adams EXTERN_C_BEGIN 472e68589bSMark F. Adams #undef __FUNCT__ 482e68589bSMark F. Adams #define __FUNCT__ "PCGAMGSetNSmooths_GAMG" 492e68589bSMark F. Adams PetscErrorCode PCGAMGSetNSmooths_GAMG(PC pc, PetscInt n) 502e68589bSMark F. Adams { 512e68589bSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 522e68589bSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 53c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 542e68589bSMark F. Adams 552e68589bSMark F. Adams PetscFunctionBegin; 56c8b0795cSMark F. Adams pc_gamg_agg->nsmooths = n; 57c8b0795cSMark F. Adams PetscFunctionReturn(0); 58c8b0795cSMark F. Adams } 59c8b0795cSMark F. Adams EXTERN_C_END 60c8b0795cSMark F. Adams 61c8b0795cSMark F. Adams #undef __FUNCT__ 62c8b0795cSMark F. Adams #define __FUNCT__ "PCGAMGSetSymGraph" 63c8b0795cSMark F. Adams /*@ 64c8b0795cSMark F. Adams PCGAMGSetSymGraph - 65c8b0795cSMark F. Adams 66c8b0795cSMark F. Adams Not Collective on PC 67c8b0795cSMark F. Adams 68c8b0795cSMark F. Adams Input Parameters: 69c8b0795cSMark F. Adams . pc - the preconditioner context 70c8b0795cSMark F. Adams 71c8b0795cSMark F. Adams Options Database Key: 72c8b0795cSMark F. Adams . -pc_gamg_sym_graph 73c8b0795cSMark F. Adams 74c8b0795cSMark F. Adams Level: intermediate 75c8b0795cSMark F. Adams 76c8b0795cSMark F. Adams Concepts: Aggregation AMG preconditioner 77c8b0795cSMark F. Adams 78c8b0795cSMark F. Adams .seealso: () 79c8b0795cSMark F. Adams @*/ 80c8b0795cSMark F. Adams PetscErrorCode PCGAMGSetSymGraph(PC pc, PetscBool n) 81c8b0795cSMark F. Adams { 82c8b0795cSMark F. Adams PetscErrorCode ierr; 83c8b0795cSMark F. Adams 84c8b0795cSMark F. Adams PetscFunctionBegin; 85c8b0795cSMark F. Adams PetscValidHeaderSpecific(pc,PC_CLASSID,1); 86c8b0795cSMark F. Adams ierr = PetscTryMethod(pc,"PCGAMGSetSymGraph_C",(PC,PetscBool),(pc,n));CHKERRQ(ierr); 87c8b0795cSMark F. Adams PetscFunctionReturn(0); 88c8b0795cSMark F. Adams } 89c8b0795cSMark F. Adams 90c8b0795cSMark F. Adams EXTERN_C_BEGIN 91c8b0795cSMark F. Adams #undef __FUNCT__ 92c8b0795cSMark F. Adams #define __FUNCT__ "PCGAMGSetSymGraph_GAMG" 93c8b0795cSMark F. Adams PetscErrorCode PCGAMGSetSymGraph_GAMG(PC pc, PetscBool n) 94c8b0795cSMark F. Adams { 95c8b0795cSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 96c8b0795cSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 97c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 98c8b0795cSMark F. Adams 99c8b0795cSMark F. Adams PetscFunctionBegin; 100c8b0795cSMark F. Adams pc_gamg_agg->sym_graph = n; 1012e68589bSMark F. Adams PetscFunctionReturn(0); 1022e68589bSMark F. Adams } 1032e68589bSMark F. Adams EXTERN_C_END 1042e68589bSMark F. Adams 105ef4ad70eSMark F. Adams #undef __FUNCT__ 106ef4ad70eSMark F. Adams #define __FUNCT__ "PCGAMGSetSquareGraph" 107ef4ad70eSMark F. Adams /*@ 108ef4ad70eSMark F. Adams PCGAMGSetSquareGraph - 109ef4ad70eSMark F. Adams 110ef4ad70eSMark F. Adams Not Collective on PC 111ef4ad70eSMark F. Adams 112ef4ad70eSMark F. Adams Input Parameters: 113ef4ad70eSMark F. Adams . pc - the preconditioner context 114ef4ad70eSMark F. Adams 115ef4ad70eSMark F. Adams Options Database Key: 116ef4ad70eSMark F. Adams . -pc_gamg_square_graph 117ef4ad70eSMark F. Adams 118ef4ad70eSMark F. Adams Level: intermediate 119ef4ad70eSMark F. Adams 120ef4ad70eSMark F. Adams Concepts: Aggregation AMG preconditioner 121ef4ad70eSMark F. Adams 122ef4ad70eSMark F. Adams .seealso: () 123ef4ad70eSMark F. Adams @*/ 124ef4ad70eSMark F. Adams PetscErrorCode PCGAMGSetSquareGraph(PC pc, PetscBool n) 125ef4ad70eSMark F. Adams { 126ef4ad70eSMark F. Adams PetscErrorCode ierr; 127ef4ad70eSMark F. Adams 128ef4ad70eSMark F. Adams PetscFunctionBegin; 129ef4ad70eSMark F. Adams PetscValidHeaderSpecific(pc,PC_CLASSID,1); 130ef4ad70eSMark F. Adams ierr = PetscTryMethod(pc,"PCGAMGSetSquareGraph_C",(PC,PetscBool),(pc,n));CHKERRQ(ierr); 131ef4ad70eSMark F. Adams PetscFunctionReturn(0); 132ef4ad70eSMark F. Adams } 133ef4ad70eSMark F. Adams 134ef4ad70eSMark F. Adams EXTERN_C_BEGIN 135ef4ad70eSMark F. Adams #undef __FUNCT__ 136ef4ad70eSMark F. Adams #define __FUNCT__ "PCGAMGSetSquareGraph_GAMG" 137ef4ad70eSMark F. Adams PetscErrorCode PCGAMGSetSquareGraph_GAMG(PC pc, PetscBool n) 138ef4ad70eSMark F. Adams { 139ef4ad70eSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 140ef4ad70eSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 141ef4ad70eSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 142ef4ad70eSMark F. Adams 143ef4ad70eSMark F. Adams PetscFunctionBegin; 144ef4ad70eSMark F. Adams pc_gamg_agg->square_graph = n; 145ef4ad70eSMark F. Adams PetscFunctionReturn(0); 146ef4ad70eSMark F. Adams } 147ef4ad70eSMark F. Adams EXTERN_C_END 148ef4ad70eSMark F. Adams 1492e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 1502e68589bSMark F. Adams /* 1512e68589bSMark F. Adams PCSetFromOptions_GAMG_AGG 1522e68589bSMark F. Adams 1532e68589bSMark F. Adams Input Parameter: 1542e68589bSMark F. Adams . pc - 1552e68589bSMark F. Adams */ 1562e68589bSMark F. Adams #undef __FUNCT__ 1572e68589bSMark F. Adams #define __FUNCT__ "PCSetFromOptions_GAMG_AGG" 1582e68589bSMark F. Adams PetscErrorCode PCSetFromOptions_GAMG_AGG( PC pc ) 1592e68589bSMark F. Adams { 1602e68589bSMark F. Adams PetscErrorCode ierr; 1612e68589bSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 1622e68589bSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 163c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 1642e68589bSMark F. Adams PetscBool flag; 1652e68589bSMark F. Adams 1662e68589bSMark F. Adams PetscFunctionBegin; 1672e68589bSMark F. Adams /* call base class */ 1682e68589bSMark F. Adams ierr = PCSetFromOptions_GAMG( pc ); CHKERRQ(ierr); 1692e68589bSMark F. Adams 1702e68589bSMark F. Adams ierr = PetscOptionsHead("GAMG-AGG options"); CHKERRQ(ierr); 1712e68589bSMark F. Adams { 1722e68589bSMark F. Adams /* -pc_gamg_agg_nsmooths */ 173c8b0795cSMark F. Adams pc_gamg_agg->nsmooths = 0; 1742e68589bSMark F. Adams ierr = PetscOptionsInt("-pc_gamg_agg_nsmooths", 1752e68589bSMark F. Adams "smoothing steps for smoothed aggregation, usually 1 (0)", 1762e68589bSMark F. Adams "PCGAMGSetNSmooths", 177c8b0795cSMark F. Adams pc_gamg_agg->nsmooths, 178c8b0795cSMark F. Adams &pc_gamg_agg->nsmooths, 179c8b0795cSMark F. Adams &flag); 180c8b0795cSMark F. Adams CHKERRQ(ierr); 181c8b0795cSMark F. Adams 182c8b0795cSMark F. Adams /* -pc_gamg_sym_graph */ 183c8b0795cSMark F. Adams pc_gamg_agg->sym_graph = PETSC_FALSE; 184c8b0795cSMark F. Adams ierr = PetscOptionsBool("-pc_gamg_sym_graph", 185581a99e3SJed Brown "Set for asymmetric matrices", 186c8b0795cSMark F. Adams "PCGAMGSetSymGraph", 187c8b0795cSMark F. Adams pc_gamg_agg->sym_graph, 188c8b0795cSMark F. Adams &pc_gamg_agg->sym_graph, 1892e68589bSMark F. Adams &flag); 1902e68589bSMark F. Adams CHKERRQ(ierr); 191ef4ad70eSMark F. Adams 192ef4ad70eSMark F. Adams /* -pc_gamg_square_graph */ 193ef4ad70eSMark F. Adams pc_gamg_agg->square_graph = PETSC_TRUE; 194ef4ad70eSMark F. Adams ierr = PetscOptionsBool("-pc_gamg_square_graph", 1950cbbd2e1SMark F. Adams "For faster coarsening and lower coarse grid complexity", 196ef4ad70eSMark F. Adams "PCGAMGSetSquareGraph", 197ef4ad70eSMark F. Adams pc_gamg_agg->square_graph, 198ef4ad70eSMark F. Adams &pc_gamg_agg->square_graph, 199ef4ad70eSMark F. Adams &flag); 200ef4ad70eSMark F. Adams CHKERRQ(ierr); 2012e68589bSMark F. Adams } 2022e68589bSMark F. Adams ierr = PetscOptionsTail();CHKERRQ(ierr); 2032e68589bSMark F. Adams 2042e68589bSMark F. Adams PetscFunctionReturn(0); 2052e68589bSMark F. Adams } 2062e68589bSMark F. Adams 2072e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 2082e68589bSMark F. Adams /* 2092e68589bSMark F. Adams PCDestroy_AGG 2102e68589bSMark F. Adams 2112e68589bSMark F. Adams Input Parameter: 2122e68589bSMark F. Adams . pc - 2132e68589bSMark F. Adams */ 2142e68589bSMark F. Adams #undef __FUNCT__ 2152e68589bSMark F. Adams #define __FUNCT__ "PCDestroy_AGG" 2162e68589bSMark F. Adams PetscErrorCode PCDestroy_AGG( PC pc ) 2172e68589bSMark F. Adams { 2182e68589bSMark F. Adams PetscErrorCode ierr; 2192e68589bSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 2202e68589bSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 221c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 2222e68589bSMark F. Adams 2232e68589bSMark F. Adams PetscFunctionBegin; 224c8b0795cSMark F. Adams if( pc_gamg_agg ) { 225c8b0795cSMark F. Adams ierr = PetscFree(pc_gamg_agg);CHKERRQ(ierr); 226c8b0795cSMark F. Adams pc_gamg_agg = 0; 2272e68589bSMark F. Adams } 2282e68589bSMark F. Adams 2292e68589bSMark F. Adams /* call base class */ 2302e68589bSMark F. Adams ierr = PCDestroy_GAMG( pc );CHKERRQ(ierr); 2312e68589bSMark F. Adams 2322e68589bSMark F. Adams PetscFunctionReturn(0); 2332e68589bSMark F. Adams } 2342e68589bSMark F. Adams 2352e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 2362e68589bSMark F. Adams /* 2372e68589bSMark F. Adams PCSetCoordinates_AGG 238*302f38e8SMark F. Adams - collective 2392e68589bSMark F. Adams 2402e68589bSMark F. Adams Input Parameter: 2412e68589bSMark F. Adams . pc - the preconditioner context 242*302f38e8SMark F. Adams . ndm - dimesion 243*302f38e8SMark F. Adams . a_nloc - number of vertices local 244*302f38e8SMark F. Adams . coords - [a_nloc][ndm] - interleaved coordinate data: {x_0, y_0, z_0, x_1, y_1, ...} 2452e68589bSMark F. Adams */ 2462e68589bSMark F. Adams EXTERN_C_BEGIN 2472e68589bSMark F. Adams #undef __FUNCT__ 2482e68589bSMark F. Adams #define __FUNCT__ "PCSetCoordinates_AGG" 249*302f38e8SMark F. Adams PetscErrorCode PCSetCoordinates_AGG( PC pc, PetscInt ndm, PetscInt a_nloc, PetscReal *coords ) 2502e68589bSMark F. Adams { 2512e68589bSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 2522e68589bSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 2532e68589bSMark F. Adams PetscErrorCode ierr; 254*302f38e8SMark F. Adams PetscInt arrsz,bs,kk,ii,jj,nloc; 2552e68589bSMark F. Adams Mat Amat = pc->pmat; 256*302f38e8SMark F. Adams MPI_Comm wcomm = ((PetscObject)pc)->comm; 2572e68589bSMark F. Adams 2582e68589bSMark F. Adams PetscFunctionBegin; 2592e68589bSMark F. Adams PetscValidHeaderSpecific( Amat, MAT_CLASSID, 1 ); 260*302f38e8SMark F. Adams /* set 'bs' and 'nloc' */ 2612e68589bSMark F. Adams ierr = MatGetBlockSize( Amat, &bs ); CHKERRQ( ierr ); 262*302f38e8SMark F. Adams if( a_nloc == -1 ) { 263*302f38e8SMark F. Adams PetscInt my0, Iend; 264*302f38e8SMark F. Adams /* stokes = PETSC_FALSE; */ 2652e68589bSMark F. Adams ierr = MatGetOwnershipRange( Amat, &my0, &Iend ); CHKERRQ(ierr); 2662e68589bSMark F. Adams nloc = (Iend-my0)/bs; 2672e68589bSMark F. Adams if((Iend-my0)%bs!=0) SETERRQ1(((PetscObject)Amat)->comm,PETSC_ERR_ARG_WRONG, "Bad local size %d.",nloc); 268*302f38e8SMark F. Adams } 269*302f38e8SMark F. Adams else { 270*302f38e8SMark F. Adams ierr = MPI_Allreduce( &a_nloc, &ii, 1, MPIU_INT, MPIU_SUM, wcomm ); CHKERRQ( ierr ); 271*302f38e8SMark F. Adams ierr = MatGetSize( Amat, &kk, &jj ); CHKERRQ( ierr ); 272*302f38e8SMark F. Adams if( bs==1 && ii!=kk ) { 273*302f38e8SMark F. Adams /* stokes = PETSC_TRUE; */ 274*302f38e8SMark F. Adams bs = ndm; 275*302f38e8SMark F. Adams nloc = a_nloc; 276*302f38e8SMark F. Adams } 277*302f38e8SMark F. Adams else{ 278*302f38e8SMark F. Adams PetscInt my0,Iend; 279*302f38e8SMark F. Adams /* stokes = PETSC_FALSE; */ 280*302f38e8SMark F. Adams ierr = MatGetOwnershipRange( Amat, &my0, &Iend ); CHKERRQ(ierr); 281*302f38e8SMark F. Adams nloc = (Iend-my0)/bs; 282*302f38e8SMark F. Adams if((Iend-my0)%bs!=0) SETERRQ1(((PetscObject)Amat)->comm,PETSC_ERR_ARG_WRONG, "Bad local size %d.",nloc); 283*302f38e8SMark F. Adams } 284*302f38e8SMark F. Adams } 2852e68589bSMark F. Adams 2862e68589bSMark F. Adams /* SA: null space vectors */ 287c8b0795cSMark F. Adams if( coords && bs==1 ) pc_gamg->data_cell_cols = 1; /* scalar w/ coords and SA (not needed) */ 288c8b0795cSMark F. Adams else if( coords ) pc_gamg->data_cell_cols = (ndm==2 ? 3 : 6); /* elasticity */ 289c8b0795cSMark F. Adams else pc_gamg->data_cell_cols = bs; /* no data, force SA with constant null space vectors */ 290c8b0795cSMark F. Adams pc_gamg->data_cell_rows = bs; 2912e68589bSMark F. Adams 292c8b0795cSMark F. Adams arrsz = nloc*pc_gamg->data_cell_rows*pc_gamg->data_cell_cols; 2932e68589bSMark F. Adams 2942e68589bSMark F. Adams /* create data - syntactic sugar that should be refactored at some point */ 2952e68589bSMark F. Adams if (pc_gamg->data==0 || (pc_gamg->data_sz != arrsz)) { 2962e68589bSMark F. Adams ierr = PetscFree( pc_gamg->data ); CHKERRQ(ierr); 297*302f38e8SMark F. Adams ierr = PetscMalloc((arrsz+1)*sizeof(PetscReal), &pc_gamg->data ); CHKERRQ(ierr); 298*302f38e8SMark F. Adams /* !nul if if nloc==0 */ 2992e68589bSMark F. Adams } 3002e68589bSMark F. Adams /* copy data in - column oriented */ 3012e68589bSMark F. Adams for(kk=0;kk<nloc;kk++){ 302*302f38e8SMark F. Adams const PetscInt M = nloc*pc_gamg->data_cell_rows; 3032e68589bSMark F. Adams PetscReal *data = &pc_gamg->data[kk*bs]; 304c8b0795cSMark F. Adams if( pc_gamg->data_cell_cols==1 ) *data = 1.0; 3052e68589bSMark F. Adams else { 3062e68589bSMark F. Adams for(ii=0;ii<bs;ii++) 3072e68589bSMark F. Adams for(jj=0;jj<bs;jj++) 3082e68589bSMark F. Adams if(ii==jj)data[ii*M + jj] = 1.0; /* translational modes */ 3092e68589bSMark F. Adams else data[ii*M + jj] = 0.0; 3102e68589bSMark F. Adams if( coords ) { 3112e68589bSMark F. Adams if( ndm == 2 ){ /* rotational modes */ 3122e68589bSMark F. Adams data += 2*M; 3132e68589bSMark F. Adams data[0] = -coords[2*kk+1]; 3142e68589bSMark F. Adams data[1] = coords[2*kk]; 3152e68589bSMark F. Adams } 3162e68589bSMark F. Adams else { 3172e68589bSMark F. Adams data += 3*M; 3182e68589bSMark F. Adams data[0] = 0.0; data[M+0] = coords[3*kk+2]; data[2*M+0] = -coords[3*kk+1]; 3192e68589bSMark F. Adams data[1] = -coords[3*kk+2]; data[M+1] = 0.0; data[2*M+1] = coords[3*kk]; 3202e68589bSMark F. Adams data[2] = coords[3*kk+1]; data[M+2] = -coords[3*kk]; data[2*M+2] = 0.0; 3212e68589bSMark F. Adams } 3222e68589bSMark F. Adams } 3232e68589bSMark F. Adams } 3242e68589bSMark F. Adams } 3252e68589bSMark F. Adams 3262e68589bSMark F. Adams pc_gamg->data_sz = arrsz; 3272e68589bSMark F. Adams 3282e68589bSMark F. Adams PetscFunctionReturn(0); 3292e68589bSMark F. Adams } 3302e68589bSMark F. Adams EXTERN_C_END 3312e68589bSMark F. Adams 332b43b03e9SMark F. Adams typedef PetscInt NState; 333b43b03e9SMark F. Adams static const NState NOT_DONE=-2; 334b43b03e9SMark F. Adams static const NState DELETED=-1; 335b43b03e9SMark F. Adams static const NState REMOVED=-3; 336b43b03e9SMark F. Adams #define IS_SELECTED(s) (s!=DELETED && s!=NOT_DONE && s!=REMOVED) 337b43b03e9SMark F. Adams 338c8b0795cSMark F. Adams /* -------------------------------------------------------------------------- */ 339c8b0795cSMark F. Adams /* 340b43b03e9SMark F. Adams smoothAggs - greedy grab of with G1 (unsquared graph) -- AIJ specific 341b43b03e9SMark F. Adams - AGG-MG specific: clears singletons out of 'selected_2' 342c8b0795cSMark F. Adams 343c8b0795cSMark F. Adams Input Parameter: 344c8b0795cSMark F. Adams . Gmat_2 - glabal matrix of graph (data not defined) 345c8b0795cSMark F. Adams . Gmat_1 - base graph to grab with 346c8b0795cSMark F. Adams Input/Output Parameter: 3470cbbd2e1SMark F. Adams . aggs_2 - linked list of aggs with gids ) 348c8b0795cSMark F. Adams */ 349c8b0795cSMark F. Adams #undef __FUNCT__ 350c8b0795cSMark F. Adams #define __FUNCT__ "smoothAggs" 3510cbbd2e1SMark F. Adams static PetscErrorCode smoothAggs( const Mat Gmat_2, /* base (squared) graph */ 3520cbbd2e1SMark F. Adams const Mat Gmat_1, /* base graph */ 3530cbbd2e1SMark F. Adams /* const IS selected_2, [nselected local] selected vertices */ 3540cbbd2e1SMark F. Adams PetscCoarsenData *aggs_2 /* [nselected local] global ID of aggregate */ 355c8b0795cSMark F. Adams ) 356c8b0795cSMark F. Adams { 357c8b0795cSMark F. Adams PetscErrorCode ierr; 358c8b0795cSMark F. Adams PetscBool isMPI; 359c8b0795cSMark F. Adams Mat_SeqAIJ *matA_1, *matB_1=0, *matA_2, *matB_2=0; 360c8b0795cSMark F. Adams MPI_Comm wcomm = ((PetscObject)Gmat_2)->comm; 3610cbbd2e1SMark F. Adams PetscMPIInt mype,npe; 3620cbbd2e1SMark F. Adams PetscInt lid,*ii,*idx,ix,Iend,my0,kk,n,j; 363c8b0795cSMark F. Adams Mat_MPIAIJ *mpimat_2 = 0, *mpimat_1=0; 364c8b0795cSMark F. Adams const PetscInt nloc = Gmat_2->rmap->n; 3650cbbd2e1SMark F. Adams PetscScalar *cpcol_1_state,*cpcol_2_state,*cpcol_2_par_orig,*lid_parent_gid; 3660cbbd2e1SMark F. Adams PetscInt *lid_cprowID_1; 367c8b0795cSMark F. Adams NState *lid_state; 3680cbbd2e1SMark F. Adams Vec ghost_par_orig2; 369c8b0795cSMark F. Adams 370c8b0795cSMark F. Adams PetscFunctionBegin; 371c8b0795cSMark F. Adams ierr = MPI_Comm_rank( wcomm, &mype ); CHKERRQ(ierr); 3720cbbd2e1SMark F. Adams ierr = MPI_Comm_size( wcomm, &npe ); CHKERRQ(ierr); 373c8b0795cSMark F. Adams ierr = MatGetOwnershipRange(Gmat_1,&my0,&Iend); CHKERRQ(ierr); 374c8b0795cSMark F. Adams 3750cbbd2e1SMark F. Adams if( PETSC_FALSE ) { 376c8b0795cSMark F. Adams PetscViewer viewer; char fname[32]; static int llev=0; 377c8b0795cSMark F. Adams sprintf(fname,"Gmat2_%d.m",llev++); 378c8b0795cSMark F. Adams PetscViewerASCIIOpen(wcomm,fname,&viewer); 379c8b0795cSMark F. Adams ierr = PetscViewerSetFormat( viewer, PETSC_VIEWER_ASCII_MATLAB); CHKERRQ(ierr); 380c8b0795cSMark F. Adams ierr = MatView(Gmat_2, viewer ); CHKERRQ(ierr); 381c8b0795cSMark F. Adams ierr = PetscViewerDestroy( &viewer ); 382c8b0795cSMark F. Adams } 383c8b0795cSMark F. Adams 384c8b0795cSMark F. Adams /* get submatrices */ 385c8b0795cSMark F. Adams ierr = PetscTypeCompare( (PetscObject)Gmat_1, MATMPIAIJ, &isMPI ); CHKERRQ(ierr); 386c8b0795cSMark F. Adams if(isMPI) { 387c8b0795cSMark F. Adams /* grab matrix objects */ 388c8b0795cSMark F. Adams mpimat_2 = (Mat_MPIAIJ*)Gmat_2->data; 389c8b0795cSMark F. Adams mpimat_1 = (Mat_MPIAIJ*)Gmat_1->data; 390c8b0795cSMark F. Adams matA_1 = (Mat_SeqAIJ*)mpimat_1->A->data; 391c8b0795cSMark F. Adams matB_1 = (Mat_SeqAIJ*)mpimat_1->B->data; 392c8b0795cSMark F. Adams matA_2 = (Mat_SeqAIJ*)mpimat_2->A->data; 393c8b0795cSMark F. Adams matB_2 = (Mat_SeqAIJ*)mpimat_2->B->data; 394c8b0795cSMark F. Adams 395c8b0795cSMark F. Adams /* force compressed row storage for B matrix in AuxMat */ 396c8b0795cSMark F. Adams matB_1->compressedrow.check = PETSC_TRUE; 397c8b0795cSMark F. Adams ierr = MatCheckCompressedRow(mpimat_1->B,&matB_1->compressedrow,matB_1->i,Gmat_1->rmap->n,-1.0); 398c8b0795cSMark F. Adams CHKERRQ(ierr); 399c8b0795cSMark F. Adams 400c8b0795cSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscInt), &lid_cprowID_1 ); CHKERRQ(ierr); 4010cbbd2e1SMark F. Adams for( lid = 0 ; lid < nloc ; lid++ ) lid_cprowID_1[lid] = -1; 402c8b0795cSMark F. Adams for (ix=0; ix<matB_1->compressedrow.nrows; ix++) { 403c8b0795cSMark F. Adams PetscInt lid = matB_1->compressedrow.rindex[ix]; 404c8b0795cSMark F. Adams lid_cprowID_1[lid] = ix; 405c8b0795cSMark F. Adams } 406c8b0795cSMark F. Adams } 407c8b0795cSMark F. Adams else { 408c8b0795cSMark F. Adams matA_1 = (Mat_SeqAIJ*)Gmat_1->data; 409c8b0795cSMark F. Adams matA_2 = (Mat_SeqAIJ*)Gmat_2->data; 4100cbbd2e1SMark F. Adams lid_cprowID_1 = PETSC_NULL; 411c8b0795cSMark F. Adams } 412c8b0795cSMark F. Adams assert( matA_1 && !matA_1->compressedrow.use ); 413c8b0795cSMark F. Adams assert( matB_1==0 || matB_1->compressedrow.use ); 414c8b0795cSMark F. Adams assert( matA_2 && !matA_2->compressedrow.use ); 415c8b0795cSMark F. Adams assert( matB_2==0 || matB_2->compressedrow.use ); 416c8b0795cSMark F. Adams 417c8b0795cSMark F. Adams /* get state of locals and selected gid for deleted */ 418c8b0795cSMark F. Adams ierr = PetscMalloc( nloc*sizeof(NState), &lid_state ); CHKERRQ(ierr); 4190cbbd2e1SMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscScalar), &lid_parent_gid ); CHKERRQ(ierr); 420c8b0795cSMark F. Adams for( lid = 0 ; lid < nloc ; lid++ ) { 4210cbbd2e1SMark F. Adams lid_parent_gid[lid] = -1.0; 422c8b0795cSMark F. Adams lid_state[lid] = DELETED; 423c8b0795cSMark F. Adams } 4240cbbd2e1SMark F. Adams 4250cbbd2e1SMark F. Adams /* set lid_state */ 4260cbbd2e1SMark F. Adams for( lid = 0 ; lid < nloc ; lid++ ) { 42741b27cdeSMark F. Adams PetscCDPos pos; 428e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,lid,&pos); CHKERRQ(ierr); 429e78576d6SMark F. Adams if( pos ) { 430e78576d6SMark F. Adams PetscInt gid1; 431ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid1 ); CHKERRQ(ierr); assert(gid1==lid+my0); 4320cbbd2e1SMark F. Adams lid_state[lid] = gid1; 433b43b03e9SMark F. Adams } 434b43b03e9SMark F. Adams } 4350cbbd2e1SMark F. Adams 4360cbbd2e1SMark F. Adams /* map local to selected local, DELETED means a ghost owns it */ 437c8b0795cSMark F. Adams for(lid=kk=0;lid<nloc;lid++){ 438c8b0795cSMark F. Adams NState state = lid_state[lid]; 439c8b0795cSMark F. Adams if( IS_SELECTED(state) ){ 44041b27cdeSMark F. Adams PetscCDPos pos; 441e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,lid,&pos); CHKERRQ(ierr); 442e78576d6SMark F. Adams while(pos){ 443e78576d6SMark F. Adams PetscInt gid1; 444ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid1 ); CHKERRQ(ierr); 445e78576d6SMark F. Adams ierr = PetscCDGetNextPos(aggs_2,lid,&pos); CHKERRQ(ierr); 446e78576d6SMark F. Adams 4470cbbd2e1SMark F. Adams if( gid1 >= my0 && gid1 < Iend ){ 4480cbbd2e1SMark F. Adams lid_parent_gid[gid1-my0] = (PetscScalar)(lid + my0); 449c8b0795cSMark F. Adams } 450c8b0795cSMark F. Adams } 4510cbbd2e1SMark F. Adams } 4520cbbd2e1SMark F. Adams } 4530cbbd2e1SMark F. Adams /* get 'cpcol_1/2_state' & cpcol_2_par_orig - uses mpimat_1/2->lvec for temp space */ 454c8b0795cSMark F. Adams if (isMPI) { 455c8b0795cSMark F. Adams Vec tempVec; 456c8b0795cSMark F. Adams /* get 'cpcol_1_state' */ 457c8b0795cSMark F. Adams ierr = MatGetVecs( Gmat_1, &tempVec, 0 ); CHKERRQ(ierr); 458c8b0795cSMark F. Adams for(kk=0,j=my0;kk<nloc;kk++,j++){ 459c8b0795cSMark F. Adams PetscScalar v = (PetscScalar)lid_state[kk]; 460c8b0795cSMark F. Adams ierr = VecSetValues( tempVec, 1, &j, &v, INSERT_VALUES ); CHKERRQ(ierr); 461c8b0795cSMark F. Adams } 462c8b0795cSMark F. Adams ierr = VecAssemblyBegin( tempVec ); CHKERRQ(ierr); 463c8b0795cSMark F. Adams ierr = VecAssemblyEnd( tempVec ); CHKERRQ(ierr); 464c8b0795cSMark F. Adams ierr = VecScatterBegin(mpimat_1->Mvctx,tempVec, mpimat_1->lvec,INSERT_VALUES,SCATTER_FORWARD); 465c8b0795cSMark F. Adams CHKERRQ(ierr); 466c8b0795cSMark F. Adams ierr = VecScatterEnd(mpimat_1->Mvctx,tempVec, mpimat_1->lvec,INSERT_VALUES,SCATTER_FORWARD); 467c8b0795cSMark F. Adams CHKERRQ(ierr); 468c8b0795cSMark F. Adams ierr = VecGetArray( mpimat_1->lvec, &cpcol_1_state ); CHKERRQ(ierr); 469c8b0795cSMark F. Adams /* get 'cpcol_2_state' */ 470c8b0795cSMark F. Adams ierr = VecScatterBegin(mpimat_2->Mvctx,tempVec, mpimat_2->lvec,INSERT_VALUES,SCATTER_FORWARD); 471c8b0795cSMark F. Adams CHKERRQ(ierr); 472c8b0795cSMark F. Adams ierr = VecScatterEnd(mpimat_2->Mvctx,tempVec, mpimat_2->lvec,INSERT_VALUES,SCATTER_FORWARD); 473c8b0795cSMark F. Adams CHKERRQ(ierr); 474c8b0795cSMark F. Adams ierr = VecGetArray( mpimat_2->lvec, &cpcol_2_state ); CHKERRQ(ierr); 4750cbbd2e1SMark F. Adams /* get 'cpcol_2_par_orig' */ 4760cbbd2e1SMark F. Adams for(kk=0,j=my0;kk<nloc;kk++,j++){ 4770cbbd2e1SMark F. Adams PetscScalar v = (PetscScalar)lid_parent_gid[kk]; 4780cbbd2e1SMark F. Adams ierr = VecSetValues( tempVec, 1, &j, &v, INSERT_VALUES ); CHKERRQ(ierr); 4790cbbd2e1SMark F. Adams } 4800cbbd2e1SMark F. Adams ierr = VecAssemblyBegin( tempVec ); CHKERRQ(ierr); 4810cbbd2e1SMark F. Adams ierr = VecAssemblyEnd( tempVec ); CHKERRQ(ierr); 4820cbbd2e1SMark F. Adams ierr = VecDuplicate( mpimat_2->lvec, &ghost_par_orig2 ); CHKERRQ(ierr); 4830cbbd2e1SMark F. Adams ierr = VecScatterBegin(mpimat_2->Mvctx,tempVec, ghost_par_orig2,INSERT_VALUES,SCATTER_FORWARD); 4840cbbd2e1SMark F. Adams CHKERRQ(ierr); 4850cbbd2e1SMark F. Adams ierr = VecScatterEnd(mpimat_2->Mvctx,tempVec, ghost_par_orig2,INSERT_VALUES,SCATTER_FORWARD); 4860cbbd2e1SMark F. Adams CHKERRQ(ierr); 4870cbbd2e1SMark F. Adams ierr = VecGetArray( ghost_par_orig2, &cpcol_2_par_orig ); CHKERRQ(ierr); 4880cbbd2e1SMark F. Adams 489c8b0795cSMark F. Adams ierr = VecDestroy( &tempVec ); CHKERRQ(ierr); 490c8b0795cSMark F. Adams } /* ismpi */ 491c8b0795cSMark F. Adams 492c8b0795cSMark F. Adams /* doit */ 493c8b0795cSMark F. Adams for(lid=0;lid<nloc;lid++){ 494c8b0795cSMark F. Adams NState state = lid_state[lid]; 4950cbbd2e1SMark F. Adams if( IS_SELECTED(state) ) { 4960cbbd2e1SMark F. Adams /* steal locals */ 497c8b0795cSMark F. Adams ii = matA_1->i; n = ii[lid+1] - ii[lid]; 498c8b0795cSMark F. Adams idx = matA_1->j + ii[lid]; 499c8b0795cSMark F. Adams for (j=0; j<n; j++) { 5000cbbd2e1SMark F. Adams PetscInt lidj = idx[j], sgid; 501c8b0795cSMark F. Adams NState statej = lid_state[lidj]; 5020cbbd2e1SMark F. Adams if (statej==DELETED && (sgid=(PetscInt)PetscRealPart(lid_parent_gid[lidj])) != lid+my0) { /* steal local */ 5030cbbd2e1SMark F. Adams lid_parent_gid[lidj] = (PetscScalar)(lid+my0); /* send this if sgid is not local */ 5040cbbd2e1SMark F. Adams if( sgid >= my0 && sgid < Iend ){ /* I'm stealing this local from a local sgid */ 5050cbbd2e1SMark F. Adams PetscInt hav=0,slid=sgid-my0,gidj=lidj+my0; 50641b27cdeSMark F. Adams PetscCDPos pos,last=PETSC_NULL; 507c8b0795cSMark F. Adams /* looking for local from local so id_llist_2 works */ 508e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,slid,&pos); CHKERRQ(ierr); 509e78576d6SMark F. Adams while(pos){ 510e78576d6SMark F. Adams PetscInt gid; 511ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid ); CHKERRQ(ierr); 5120cbbd2e1SMark F. Adams if( gid == gidj ) { 5130cbbd2e1SMark F. Adams assert(last); 51441b27cdeSMark F. Adams ierr = PetscCDRemoveNextNode( aggs_2, slid, last ); CHKERRQ(ierr); 51541b27cdeSMark F. Adams ierr = PetscCDAppendNode( aggs_2, lid, pos ); CHKERRQ(ierr); 5160cbbd2e1SMark F. Adams hav = 1; 517c8b0795cSMark F. Adams break; 518c8b0795cSMark F. Adams } 5190cbbd2e1SMark F. Adams else last = pos; 520e78576d6SMark F. Adams 521e78576d6SMark F. Adams ierr = PetscCDGetNextPos(aggs_2,slid,&pos); CHKERRQ(ierr); 522c8b0795cSMark F. Adams } 523c8b0795cSMark F. Adams if(hav!=1){ 524c8b0795cSMark F. Adams if(hav==0)SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"failed to find adj in 'selected' lists - structurally unsymmetric matrix"); 525c8b0795cSMark F. Adams SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"found node %d times???",hav); 526c8b0795cSMark F. Adams } 527c8b0795cSMark F. Adams } 5280cbbd2e1SMark F. Adams else{ /* I'm stealing this local, owned by a ghost */ 529c8b0795cSMark F. Adams assert(sgid==-1); 53041b27cdeSMark F. Adams ierr = PetscCDAppendID( aggs_2, lid, lidj+my0 ); CHKERRQ(ierr); 531c8b0795cSMark F. Adams } 532c8b0795cSMark F. Adams } 5330cbbd2e1SMark F. Adams } /* local neighbors */ 534c8b0795cSMark F. Adams } 535c8b0795cSMark F. Adams else if( state == DELETED && lid_cprowID_1 ) { 5360cbbd2e1SMark F. Adams PetscInt sgidold = (PetscInt)PetscRealPart(lid_parent_gid[lid]); 537c8b0795cSMark F. Adams /* see if I have a selected ghost neighbor that will steal me */ 538c8b0795cSMark F. Adams if( (ix=lid_cprowID_1[lid]) != -1 ){ 539c8b0795cSMark F. Adams ii = matB_1->compressedrow.i; n = ii[ix+1] - ii[ix]; 540c8b0795cSMark F. Adams idx = matB_1->j + ii[ix]; 541c8b0795cSMark F. Adams for( j=0 ; j<n ; j++ ) { 542c8b0795cSMark F. Adams PetscInt cpid = idx[j]; 543c8b0795cSMark F. Adams NState statej = (NState)PetscRealPart(cpcol_1_state[cpid]); 544c8b0795cSMark F. Adams if( IS_SELECTED(statej) && sgidold != (PetscInt)statej ) { /* ghost will steal this, remove from my list */ 5450cbbd2e1SMark F. Adams lid_parent_gid[lid] = (PetscScalar)statej; /* send who selected */ 5460cbbd2e1SMark F. Adams if( sgidold>=my0 && sgidold<Iend ) { /* this was mine */ 5470cbbd2e1SMark F. Adams PetscInt hav=0,oldslidj=sgidold-my0; 54841b27cdeSMark F. Adams PetscCDPos pos,last=PETSC_NULL; 5490cbbd2e1SMark F. Adams /* remove from 'oldslidj' list */ 550e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,oldslidj,&pos); CHKERRQ(ierr); 551e78576d6SMark F. Adams while( pos ) { 552e78576d6SMark F. Adams PetscInt gid; 553ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid ); CHKERRQ(ierr); 5540cbbd2e1SMark F. Adams if( lid+my0 == gid ) { 5550cbbd2e1SMark F. Adams /* id_llist_2[lastid] = id_llist_2[flid]; /\* remove lid from oldslidj list *\/ */ 5560cbbd2e1SMark F. Adams assert(last); 55741b27cdeSMark F. Adams ierr = PetscCDRemoveNextNode( aggs_2, oldslidj, last ); CHKERRQ(ierr); 5580cbbd2e1SMark F. Adams /* ghost (PetscScalar)statej will add this later */ 5590cbbd2e1SMark F. Adams hav = 1; 560c8b0795cSMark F. Adams break; 561c8b0795cSMark F. Adams } 5620cbbd2e1SMark F. Adams else last = pos; 563e78576d6SMark F. Adams 564e78576d6SMark F. Adams ierr = PetscCDGetNextPos(aggs_2,oldslidj,&pos); CHKERRQ(ierr); 565c8b0795cSMark F. Adams } 566c8b0795cSMark F. Adams if(hav!=1){ 567c8b0795cSMark F. Adams if(hav==0)SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"failed to find adj in 'selected' lists - structurally unsymmetric matrix"); 568c8b0795cSMark F. Adams SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"found node %d times???",hav); 569c8b0795cSMark F. Adams } 570c8b0795cSMark F. Adams } 5710cbbd2e1SMark F. Adams else { 5720cbbd2e1SMark F. Adams /* ghosts remove this later */ 5730cbbd2e1SMark F. Adams } 574c8b0795cSMark F. Adams } 575c8b0795cSMark F. Adams } 576c8b0795cSMark F. Adams } 577c8b0795cSMark F. Adams } /* selected/deleted */ 578c8b0795cSMark F. Adams } /* node loop */ 579c8b0795cSMark F. Adams 580c8b0795cSMark F. Adams if( isMPI ) { 5810cbbd2e1SMark F. Adams PetscScalar *cpcol_2_parent,*cpcol_2_gid; 5820cbbd2e1SMark F. Adams Vec tempVec,ghostgids2,ghostparents2; 5830cbbd2e1SMark F. Adams PetscInt cpid,nghost_2; 5840cbbd2e1SMark F. Adams GAMGHashTable gid_cpid; 585c8b0795cSMark F. Adams 5860cbbd2e1SMark F. Adams ierr = VecGetSize( mpimat_2->lvec, &nghost_2 ); CHKERRQ(ierr); 587c8b0795cSMark F. Adams ierr = MatGetVecs( Gmat_2, &tempVec, 0 ); CHKERRQ(ierr); 5880cbbd2e1SMark F. Adams 5890cbbd2e1SMark F. Adams /* get 'cpcol_2_parent' */ 590c8b0795cSMark F. Adams for(kk=0,j=my0;kk<nloc;kk++,j++){ 5910cbbd2e1SMark F. Adams ierr = VecSetValues( tempVec, 1, &j, &lid_parent_gid[kk], INSERT_VALUES ); CHKERRQ(ierr); 592c8b0795cSMark F. Adams } 593c8b0795cSMark F. Adams ierr = VecAssemblyBegin( tempVec ); CHKERRQ(ierr); 594c8b0795cSMark F. Adams ierr = VecAssemblyEnd( tempVec ); CHKERRQ(ierr); 5950cbbd2e1SMark F. Adams ierr = VecDuplicate( mpimat_2->lvec, &ghostparents2 ); CHKERRQ(ierr); 5960cbbd2e1SMark F. Adams ierr = VecScatterBegin(mpimat_2->Mvctx,tempVec, ghostparents2,INSERT_VALUES,SCATTER_FORWARD); 597c8b0795cSMark F. Adams CHKERRQ(ierr); 5980cbbd2e1SMark F. Adams ierr = VecScatterEnd(mpimat_2->Mvctx,tempVec, ghostparents2,INSERT_VALUES,SCATTER_FORWARD); 599c8b0795cSMark F. Adams CHKERRQ(ierr); 6000cbbd2e1SMark F. Adams ierr = VecGetArray( ghostparents2, &cpcol_2_parent ); CHKERRQ(ierr); 6010cbbd2e1SMark F. Adams 6020cbbd2e1SMark F. Adams /* get 'cpcol_2_gid' */ 6030cbbd2e1SMark F. Adams for(kk=0,j=my0;kk<nloc;kk++,j++){ 6040cbbd2e1SMark F. Adams PetscScalar v = (PetscScalar)j; 6050cbbd2e1SMark F. Adams ierr = VecSetValues( tempVec, 1, &j, &v, INSERT_VALUES ); CHKERRQ(ierr); 6060cbbd2e1SMark F. Adams } 6070cbbd2e1SMark F. Adams ierr = VecAssemblyBegin( tempVec ); CHKERRQ(ierr); 6080cbbd2e1SMark F. Adams ierr = VecAssemblyEnd( tempVec ); CHKERRQ(ierr); 6090cbbd2e1SMark F. Adams ierr = VecDuplicate( mpimat_2->lvec, &ghostgids2 ); CHKERRQ(ierr); 6100cbbd2e1SMark F. Adams ierr = VecScatterBegin(mpimat_2->Mvctx,tempVec, ghostgids2,INSERT_VALUES,SCATTER_FORWARD); 6110cbbd2e1SMark F. Adams CHKERRQ(ierr); 6120cbbd2e1SMark F. Adams ierr = VecScatterEnd(mpimat_2->Mvctx,tempVec, ghostgids2,INSERT_VALUES,SCATTER_FORWARD); 6130cbbd2e1SMark F. Adams CHKERRQ(ierr); 6140cbbd2e1SMark F. Adams ierr = VecGetArray( ghostgids2, &cpcol_2_gid ); CHKERRQ(ierr); 6150cbbd2e1SMark F. Adams 616c8b0795cSMark F. Adams ierr = VecDestroy( &tempVec ); CHKERRQ(ierr); 617c8b0795cSMark F. Adams 6180cbbd2e1SMark F. Adams /* look for deleted ghosts and add to table */ 6190cbbd2e1SMark F. Adams ierr = GAMGTableCreate( 2*nghost_2, &gid_cpid ); CHKERRQ(ierr); 6200cbbd2e1SMark F. Adams for( cpid = 0 ; cpid < nghost_2 ; cpid++ ) { 6210cbbd2e1SMark F. Adams NState state = (NState)PetscRealPart(cpcol_2_state[cpid]); 6220cbbd2e1SMark F. Adams if( state==DELETED ) { 6230cbbd2e1SMark F. Adams PetscInt sgid_new = (PetscInt)PetscRealPart(cpcol_2_parent[cpid]); 6240cbbd2e1SMark F. Adams PetscInt sgid_old = (PetscInt)PetscRealPart(cpcol_2_par_orig[cpid]); 6250cbbd2e1SMark F. Adams if( sgid_old == -1 && sgid_new != -1 ) { 6260cbbd2e1SMark F. Adams PetscInt gid = (PetscInt)PetscRealPart(cpcol_2_gid[cpid]); 6270cbbd2e1SMark F. Adams ierr = GAMGTableAdd( &gid_cpid, gid, cpid ); CHKERRQ(ierr); 6280cbbd2e1SMark F. Adams } 6290cbbd2e1SMark F. Adams } 6300cbbd2e1SMark F. Adams } 631c8b0795cSMark F. Adams 6320cbbd2e1SMark F. Adams /* look for deleted ghosts and see if they moved - remove it */ 633c8b0795cSMark F. Adams for(lid=0;lid<nloc;lid++){ 634c8b0795cSMark F. Adams NState state = lid_state[lid]; 635c8b0795cSMark F. Adams if( IS_SELECTED(state) ){ 63641b27cdeSMark F. Adams PetscCDPos pos,last=PETSC_NULL; 637c8b0795cSMark F. Adams /* look for deleted ghosts and see if they moved */ 638e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,lid,&pos); CHKERRQ(ierr); 639e78576d6SMark F. Adams while(pos){ 640e78576d6SMark F. Adams PetscInt gid; 641ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid ); CHKERRQ(ierr); 642e78576d6SMark F. Adams 6430cbbd2e1SMark F. Adams if( gid < my0 || gid >= Iend ) { 6440cbbd2e1SMark F. Adams ierr = GAMGTableFind( &gid_cpid, gid, &cpid ); CHKERRQ(ierr); 6450cbbd2e1SMark F. Adams if( cpid != -1 ) { 6460cbbd2e1SMark F. Adams /* a moved ghost - */ 6470cbbd2e1SMark F. Adams /* id_llist_2[lastid] = id_llist_2[flid]; /\* remove 'flid' from list *\/ */ 64841b27cdeSMark F. Adams ierr = PetscCDRemoveNextNode( aggs_2, lid, last ); CHKERRQ(ierr); 6490cbbd2e1SMark F. Adams } 6500cbbd2e1SMark F. Adams else last = pos; 6510cbbd2e1SMark F. Adams } 6520cbbd2e1SMark F. Adams else last = pos; 653e78576d6SMark F. Adams 654e78576d6SMark F. Adams ierr = PetscCDGetNextPos(aggs_2,lid,&pos); CHKERRQ(ierr); 655c8b0795cSMark F. Adams } /* loop over list of deleted */ 656c8b0795cSMark F. Adams } /* selected */ 657c8b0795cSMark F. Adams } 6580cbbd2e1SMark F. Adams ierr = GAMGTableDestroy( &gid_cpid ); CHKERRQ(ierr); 659c8b0795cSMark F. Adams 6600cbbd2e1SMark F. Adams /* look at ghosts, see if they changed - and it */ 6610cbbd2e1SMark F. Adams for( cpid = 0 ; cpid < nghost_2 ; cpid++ ){ 6620cbbd2e1SMark F. Adams PetscInt sgid_new = (PetscInt)PetscRealPart(cpcol_2_parent[cpid]); 6630cbbd2e1SMark F. Adams if( sgid_new >= my0 && sgid_new < Iend ) { /* this is mine */ 6640cbbd2e1SMark F. Adams PetscInt gid = (PetscInt)PetscRealPart(cpcol_2_gid[cpid]); 6650cbbd2e1SMark F. Adams PetscInt slid_new=sgid_new-my0,hav=0; 66641b27cdeSMark F. Adams PetscCDPos pos; 6670cbbd2e1SMark F. Adams /* search for this gid to see if I have it */ 668e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(aggs_2,slid_new,&pos); CHKERRQ(ierr); 669e78576d6SMark F. Adams while(pos){ 670e78576d6SMark F. Adams PetscInt gidj; 671ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gidj ); CHKERRQ(ierr); 672e78576d6SMark F. Adams ierr = PetscCDGetNextPos(aggs_2,slid_new,&pos); CHKERRQ(ierr); 673e78576d6SMark F. Adams 6740cbbd2e1SMark F. Adams if( gidj == gid ) { hav = 1; break; } 675c8b0795cSMark F. Adams } 676c8b0795cSMark F. Adams if( hav != 1 ){ 677ffc955d6SMark F. Adams /* insert 'flidj' into head of llist */ 67841b27cdeSMark F. Adams ierr = PetscCDAppendID( aggs_2, slid_new, gid ); CHKERRQ(ierr); 679c8b0795cSMark F. Adams } 680c8b0795cSMark F. Adams } 681c8b0795cSMark F. Adams } 682c8b0795cSMark F. Adams 6830cbbd2e1SMark F. Adams ierr = VecRestoreArray( mpimat_1->lvec, &cpcol_1_state ); CHKERRQ(ierr); 6840cbbd2e1SMark F. Adams ierr = VecRestoreArray( mpimat_2->lvec, &cpcol_2_state ); CHKERRQ(ierr); 6850cbbd2e1SMark F. Adams ierr = VecRestoreArray( ghostparents2, &cpcol_2_parent ); CHKERRQ(ierr); 6860cbbd2e1SMark F. Adams ierr = VecRestoreArray( ghostgids2, &cpcol_2_gid ); CHKERRQ(ierr); 687c8b0795cSMark F. Adams ierr = PetscFree( lid_cprowID_1 ); CHKERRQ(ierr); 6880cbbd2e1SMark F. Adams ierr = VecDestroy( &ghostgids2 ); CHKERRQ(ierr); 6890cbbd2e1SMark F. Adams ierr = VecDestroy( &ghostparents2 ); CHKERRQ(ierr); 6900cbbd2e1SMark F. Adams ierr = VecDestroy( &ghost_par_orig2 ); CHKERRQ(ierr); 691c8b0795cSMark F. Adams } 692c8b0795cSMark F. Adams 6930cbbd2e1SMark F. Adams ierr = PetscFree( lid_parent_gid ); CHKERRQ(ierr); 694c8b0795cSMark F. Adams ierr = PetscFree( lid_state ); CHKERRQ(ierr); 695c8b0795cSMark F. Adams 696c8b0795cSMark F. Adams PetscFunctionReturn(0); 697c8b0795cSMark F. Adams } 6982e68589bSMark F. Adams 6992e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 7002e68589bSMark F. Adams /* 7012e68589bSMark F. Adams PCSetData_AGG 7022e68589bSMark F. Adams 7032e68589bSMark F. Adams Input Parameter: 7042e68589bSMark F. Adams . pc - 7052e68589bSMark F. Adams */ 7062e68589bSMark F. Adams #undef __FUNCT__ 7072e68589bSMark F. Adams #define __FUNCT__ "PCSetData_AGG" 708b8cd405aSMark F. Adams PetscErrorCode PCSetData_AGG( PC pc, Mat a_A ) 7092e68589bSMark F. Adams { 7102e68589bSMark F. Adams PetscErrorCode ierr; 711b8cd405aSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 712b8cd405aSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 713b8cd405aSMark F. Adams MatNullSpace mnull; 714b8cd405aSMark F. Adams 7152e68589bSMark F. Adams PetscFunctionBegin; 716b8cd405aSMark F. Adams ierr = MatGetNearNullSpace( a_A, &mnull ); CHKERRQ(ierr); 717b8cd405aSMark F. Adams if( !mnull ) { 718*302f38e8SMark F. Adams ierr = PCSetCoordinates_AGG( pc, -1, -1, PETSC_NULL ); CHKERRQ(ierr); 719b8cd405aSMark F. Adams } 720b8cd405aSMark F. Adams else { 721b8cd405aSMark F. Adams PetscReal *nullvec; 722b8cd405aSMark F. Adams PetscBool has_const; 723b8cd405aSMark F. Adams PetscInt i,j,mlocal,nvec,bs; 724b8cd405aSMark F. Adams const Vec *vecs; const PetscScalar *v; 725b8cd405aSMark F. Adams ierr = MatGetLocalSize(a_A,&mlocal,PETSC_NULL);CHKERRQ(ierr); 726b8cd405aSMark F. Adams ierr = MatNullSpaceGetVecs(mnull,&has_const,&nvec,&vecs);CHKERRQ(ierr); 727b8cd405aSMark F. Adams ierr = MatGetBlockSize( a_A, &bs ); CHKERRQ( ierr ); 728b8cd405aSMark F. Adams ierr = PetscMalloc((nvec+!!has_const)*mlocal*sizeof *nullvec,&nullvec);CHKERRQ(ierr); 729b8cd405aSMark F. Adams if (has_const) for (i=0; i<mlocal; i++) nullvec[i] = 1.0; 730b8cd405aSMark F. Adams for (i=0; i<nvec; i++) { 731b8cd405aSMark F. Adams ierr = VecGetArrayRead(vecs[i],&v);CHKERRQ(ierr); 732b8cd405aSMark F. Adams for (j=0; j<mlocal; j++) nullvec[(i+!!has_const)*mlocal + j] = PetscRealPart(v[j]); 733b8cd405aSMark F. Adams ierr = VecRestoreArrayRead(vecs[i],&v);CHKERRQ(ierr); 734b8cd405aSMark F. Adams } 735b8cd405aSMark F. Adams pc_gamg->data = nullvec; 736b8cd405aSMark F. Adams pc_gamg->data_cell_cols = (nvec+!!has_const); 737b8cd405aSMark F. Adams pc_gamg->data_cell_rows = bs; 738b8cd405aSMark F. Adams } 7392e68589bSMark F. Adams PetscFunctionReturn(0); 7402e68589bSMark F. Adams } 7412e68589bSMark F. Adams 7422e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 7432e68589bSMark F. Adams /* 7442e68589bSMark F. Adams formProl0 7452e68589bSMark F. Adams 7462e68589bSMark F. Adams Input Parameter: 7470cbbd2e1SMark F. Adams . agg_llists - list of arrays with aggregates 7482e68589bSMark F. Adams . bs - block size 7490cbbd2e1SMark F. Adams . nSAvec - column bs of new P 7500cbbd2e1SMark F. Adams . my0crs - global index of start of locals 7512e68589bSMark F. Adams . data_stride - bs*(nloc nodes + ghost nodes) 7522e68589bSMark F. Adams . data_in[data_stride*nSAvec] - local data on fine grid 7532e68589bSMark F. Adams . flid_fgid[data_stride/bs] - make local to global IDs, includes ghosts in 'locals_llist' 7542e68589bSMark F. Adams Output Parameter: 7552e68589bSMark F. Adams . a_data_out - in with fine grid data (w/ghosts), out with coarse grid data 7562e68589bSMark F. Adams . a_Prol - prolongation operator 7572e68589bSMark F. Adams */ 7582e68589bSMark F. Adams #undef __FUNCT__ 7592e68589bSMark F. Adams #define __FUNCT__ "formProl0" 7600cbbd2e1SMark F. Adams static PetscErrorCode formProl0(const PetscCoarsenData *agg_llists,/* list from selected vertices of aggregate unselected vertices */ 7610cbbd2e1SMark F. Adams const PetscInt bs, /* (row) block size */ 7620cbbd2e1SMark F. Adams const PetscInt nSAvec, /* column bs */ 7630cbbd2e1SMark F. Adams const PetscInt my0crs, /* global index of start of locals */ 7640cbbd2e1SMark F. Adams const PetscInt data_stride, /* (nloc+nghost)*bs */ 7650cbbd2e1SMark F. Adams PetscReal data_in[], /* [data_stride][nSAvec] */ 7660cbbd2e1SMark F. Adams const PetscInt flid_fgid[], /* [data_stride/bs] */ 7672e68589bSMark F. Adams PetscReal **a_data_out, 7682e68589bSMark F. Adams Mat a_Prol /* prolongation operator (output)*/ 7692e68589bSMark F. Adams ) 7702e68589bSMark F. Adams { 7712e68589bSMark F. Adams PetscErrorCode ierr; 7720cbbd2e1SMark F. Adams PetscInt Istart,my0,Iend,nloc,clid,flid,aggID,kk,jj,ii,mm,ndone,nSelected,minsz,nghosts,out_data_stride; 7732e68589bSMark F. Adams MPI_Comm wcomm = ((PetscObject)a_Prol)->comm; 7742e68589bSMark F. Adams PetscMPIInt mype, npe; 7752e68589bSMark F. Adams PetscReal *out_data; 77641b27cdeSMark F. Adams PetscCDPos pos; 7770cbbd2e1SMark F. Adams GAMGHashTable fgid_flid; 7780cbbd2e1SMark F. Adams 779797e13b7SMark F. Adams /* #define OUT_AGGS */ 7809057884aSMark F. Adams #ifdef OUT_AGGS 781f7620de1SMatthew G Knepley static PetscInt llev = 0; char fname[32]; FILE *file = PETSC_NULL; PetscInt pM; 7829057884aSMark F. Adams #endif 7832e68589bSMark F. Adams 7842e68589bSMark F. Adams PetscFunctionBegin; 7852e68589bSMark F. Adams ierr = MPI_Comm_rank(wcomm,&mype);CHKERRQ(ierr); 7862e68589bSMark F. Adams ierr = MPI_Comm_size(wcomm,&npe);CHKERRQ(ierr); 7872e68589bSMark F. Adams ierr = MatGetOwnershipRange( a_Prol, &Istart, &Iend ); CHKERRQ(ierr); 7880cbbd2e1SMark F. Adams nloc = (Iend-Istart)/bs; my0 = Istart/bs; assert((Iend-Istart)%bs==0); 7890cbbd2e1SMark F. Adams Iend /= bs; 7900cbbd2e1SMark F. Adams nghosts = data_stride/bs - nloc; 7912e68589bSMark F. Adams 7920cbbd2e1SMark F. Adams ierr = GAMGTableCreate( 2*nghosts, &fgid_flid ); CHKERRQ(ierr); 7930cbbd2e1SMark F. Adams for(kk=0;kk<nghosts;kk++) { 7940cbbd2e1SMark F. Adams ierr = GAMGTableAdd( &fgid_flid, flid_fgid[nloc+kk], nloc+kk ); CHKERRQ(ierr); 7952e68589bSMark F. Adams } 7962e68589bSMark F. Adams 7970cbbd2e1SMark F. Adams #ifdef OUT_AGGS 7980cbbd2e1SMark F. Adams sprintf(fname,"aggs_%d_%d.m",llev++,mype); 7990cbbd2e1SMark F. Adams if(llev==1) { 8000cbbd2e1SMark F. Adams file = fopen(fname,"w"); 8010cbbd2e1SMark F. Adams } 8020cbbd2e1SMark F. Adams MatGetSize( a_Prol, &pM, &jj ); 8030cbbd2e1SMark F. Adams #endif 8040cbbd2e1SMark F. Adams 8050cbbd2e1SMark F. Adams /* count selected -- same as number of cols of P */ 8060cbbd2e1SMark F. Adams for(nSelected=mm=0;mm<nloc;mm++) { 807e78576d6SMark F. Adams PetscBool ise; 808e78576d6SMark F. Adams ierr = PetscCDEmptyAt( agg_llists, mm, &ise ); CHKERRQ(ierr); 809e78576d6SMark F. Adams if( !ise ) nSelected++; 8100cbbd2e1SMark F. Adams } 8110cbbd2e1SMark F. Adams ierr = MatGetOwnershipRangeColumn( a_Prol, &ii, &jj ); CHKERRQ(ierr); 8120cbbd2e1SMark F. Adams assert((ii/nSAvec)==my0crs); assert(nSelected==(jj-ii)/nSAvec); 8130cbbd2e1SMark F. Adams 8142e68589bSMark F. Adams /* aloc space for coarse point data (output) */ 8150cbbd2e1SMark F. Adams out_data_stride = nSelected*nSAvec; 8160cbbd2e1SMark F. Adams ierr = PetscMalloc( out_data_stride*nSAvec*sizeof(PetscReal), &out_data ); CHKERRQ(ierr); 8170cbbd2e1SMark F. Adams for(ii=0;ii<out_data_stride*nSAvec;ii++) { 8180cbbd2e1SMark F. Adams out_data[ii]=1.e300; 8190cbbd2e1SMark F. Adams } 8200cbbd2e1SMark F. Adams *a_data_out = out_data; /* output - stride nSelected*nSAvec */ 8212e68589bSMark F. Adams 8222e68589bSMark F. Adams /* find points and set prolongation */ 823c8b0795cSMark F. Adams minsz = 100; 8242e68589bSMark F. Adams ndone = 0; 8250cbbd2e1SMark F. Adams for( mm = clid = 0 ; mm < nloc ; mm++ ){ 826e78576d6SMark F. Adams ierr = PetscCDSizeAt( agg_llists, mm, &jj ); CHKERRQ(ierr); 827e78576d6SMark F. Adams if( jj > 0 ) { 8280cbbd2e1SMark F. Adams const PetscInt lid = mm, cgid = my0crs + clid; 8290cbbd2e1SMark F. Adams PetscInt cids[100]; /* max bs */ 8300cbbd2e1SMark F. Adams PetscBLASInt asz=jj,M=asz*bs,N=nSAvec,INFO; 8312e68589bSMark F. Adams PetscBLASInt Mdata=M+((N-M>0)?N-M:0),LDA=Mdata,LWORK=N*bs; 8322e68589bSMark F. Adams PetscScalar *qqc,*qqr,*TAU,*WORK; 8332e68589bSMark F. Adams PetscInt *fids; 83465d7b583SSatish Balay PetscReal *data; 8350cbbd2e1SMark F. Adams /* count agg */ 8360cbbd2e1SMark F. Adams if( asz<minsz ) minsz = asz; 8370cbbd2e1SMark F. Adams 8380cbbd2e1SMark F. Adams /* get block */ 8392e68589bSMark F. Adams ierr = PetscMalloc( (Mdata*N)*sizeof(PetscScalar), &qqc ); CHKERRQ(ierr); 8402e68589bSMark F. Adams ierr = PetscMalloc( (M*N)*sizeof(PetscScalar), &qqr ); CHKERRQ(ierr); 8412e68589bSMark F. Adams ierr = PetscMalloc( N*sizeof(PetscScalar), &TAU ); CHKERRQ(ierr); 8422e68589bSMark F. Adams ierr = PetscMalloc( LWORK*sizeof(PetscScalar), &WORK ); CHKERRQ(ierr); 8432e68589bSMark F. Adams ierr = PetscMalloc( M*sizeof(PetscInt), &fids ); CHKERRQ(ierr); 8442e68589bSMark F. Adams 8452e68589bSMark F. Adams aggID = 0; 846e78576d6SMark F. Adams ierr = PetscCDGetHeadPos(agg_llists,lid,&pos); CHKERRQ(ierr); 847e78576d6SMark F. Adams while(pos){ 848e78576d6SMark F. Adams PetscInt gid1; 849ffc955d6SMark F. Adams ierr = PetscLLNGetID( pos, &gid1 ); CHKERRQ(ierr); 850e78576d6SMark F. Adams ierr = PetscCDGetNextPos(agg_llists,lid,&pos); CHKERRQ(ierr); 851e78576d6SMark F. Adams 8520cbbd2e1SMark F. Adams if( gid1 >= my0 && gid1 < Iend ) flid = gid1 - my0; 8530cbbd2e1SMark F. Adams else { 8540cbbd2e1SMark F. Adams ierr = GAMGTableFind( &fgid_flid, gid1, &flid ); CHKERRQ(ierr); 8550cbbd2e1SMark F. Adams assert(flid>=0); 8560cbbd2e1SMark F. Adams } 8572e68589bSMark F. Adams /* copy in B_i matrix - column oriented */ 85865d7b583SSatish Balay data = &data_in[flid*bs]; 8592e68589bSMark F. Adams for( kk = ii = 0; ii < bs ; ii++ ) { 8602e68589bSMark F. Adams for( jj = 0; jj < N ; jj++ ) { 8610cbbd2e1SMark F. Adams PetscReal d = data[jj*data_stride + ii]; 8620cbbd2e1SMark F. Adams qqc[jj*Mdata + aggID*bs + ii] = d; 8632e68589bSMark F. Adams } 8642e68589bSMark F. Adams } 8659057884aSMark F. Adams #ifdef OUT_AGGS 866b2a4f308SMark F. Adams if(llev==1) { 8679057884aSMark F. Adams char str[] = "plot(%e,%e,'r*'), hold on,\n", col[] = "rgbkmc", sim[] = "*os+h>d<vx^"; 8680cbbd2e1SMark F. Adams PetscInt MM,pi,pj; 8690cbbd2e1SMark F. Adams str[12] = col[(clid+17*mype)%6]; str[13] = sim[(clid+17*mype)%11]; 870f7620de1SMatthew G Knepley MM = (PetscInt)(PetscSqrtReal((PetscReal)pM)); 8710cbbd2e1SMark F. Adams pj = gid1/MM; pi = gid1%MM; 872b2a4f308SMark F. Adams fprintf(file,str,(double)pi,(double)pj); 873b2a4f308SMark F. Adams /* fprintf(file,str,data[2*data_stride+1],-data[2*data_stride]); */ 8749057884aSMark F. Adams } 8759057884aSMark F. Adams #endif 8762e68589bSMark F. Adams /* set fine IDs */ 8772e68589bSMark F. Adams for(kk=0;kk<bs;kk++) fids[aggID*bs + kk] = flid_fgid[flid]*bs + kk; 8782e68589bSMark F. Adams 8792e68589bSMark F. Adams aggID++; 8800cbbd2e1SMark F. Adams } 8812e68589bSMark F. Adams 8822e68589bSMark F. Adams /* pad with zeros */ 8832e68589bSMark F. Adams for( ii = asz*bs; ii < Mdata ; ii++ ) { 8842e68589bSMark F. Adams for( jj = 0; jj < N ; jj++, kk++ ) { 8852e68589bSMark F. Adams qqc[jj*Mdata + ii] = .0; 8862e68589bSMark F. Adams } 8872e68589bSMark F. Adams } 8882e68589bSMark F. Adams 8892e68589bSMark F. Adams ndone += aggID; 8902e68589bSMark F. Adams /* QR */ 8912e68589bSMark F. Adams LAPACKgeqrf_( &Mdata, &N, qqc, &LDA, TAU, WORK, &LWORK, &INFO ); 8922e68589bSMark F. Adams if( INFO != 0 ) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"xGEQRS error"); 8932e68589bSMark F. Adams /* get R - column oriented - output B_{i+1} */ 8942e68589bSMark F. Adams { 8952e68589bSMark F. Adams PetscReal *data = &out_data[clid*nSAvec]; 8962e68589bSMark F. Adams for( jj = 0; jj < nSAvec ; jj++ ) { 8972e68589bSMark F. Adams for( ii = 0; ii < nSAvec ; ii++ ) { 8980cbbd2e1SMark F. Adams assert(data[jj*out_data_stride + ii] == 1.e300); 8990cbbd2e1SMark F. Adams if( ii <= jj ) data[jj*out_data_stride + ii] = PetscRealPart(qqc[jj*Mdata + ii]); 9000cbbd2e1SMark F. Adams else data[jj*out_data_stride + ii] = 0.; 9012e68589bSMark F. Adams } 9022e68589bSMark F. Adams } 9032e68589bSMark F. Adams } 9042e68589bSMark F. Adams 9052e68589bSMark F. Adams /* get Q - row oriented */ 9062e68589bSMark F. Adams LAPACKungqr_( &Mdata, &N, &N, qqc, &LDA, TAU, WORK, &LWORK, &INFO ); 9072e68589bSMark F. Adams if( INFO != 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"xORGQR error arg %d",-INFO); 9082e68589bSMark F. Adams 9092e68589bSMark F. Adams for( ii = 0 ; ii < M ; ii++ ){ 9102e68589bSMark F. Adams for( jj = 0 ; jj < N ; jj++ ) { 9112e68589bSMark F. Adams qqr[N*ii + jj] = qqc[jj*Mdata + ii]; 9122e68589bSMark F. Adams } 9132e68589bSMark F. Adams } 9142e68589bSMark F. Adams 9152e68589bSMark F. Adams /* add diagonal block of P0 */ 916c8b0795cSMark F. Adams for(kk=0;kk<N;kk++) { 917c8b0795cSMark F. Adams cids[kk] = N*cgid + kk; /* global col IDs in P0 */ 918c8b0795cSMark F. Adams } 9192e68589bSMark F. Adams ierr = MatSetValues(a_Prol,M,fids,N,cids,qqr,INSERT_VALUES); CHKERRQ(ierr); 9202e68589bSMark F. Adams 9212e68589bSMark F. Adams ierr = PetscFree( qqc ); CHKERRQ(ierr); 9222e68589bSMark F. Adams ierr = PetscFree( qqr ); CHKERRQ(ierr); 9232e68589bSMark F. Adams ierr = PetscFree( TAU ); CHKERRQ(ierr); 9242e68589bSMark F. Adams ierr = PetscFree( WORK ); CHKERRQ(ierr); 9252e68589bSMark F. Adams ierr = PetscFree( fids ); CHKERRQ(ierr); 926b43b03e9SMark F. Adams clid++; 9270cbbd2e1SMark F. Adams } /* coarse agg */ 9280cbbd2e1SMark F. Adams } /* for all fine nodes */ 9290cbbd2e1SMark F. Adams ierr = MatAssemblyBegin(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9300cbbd2e1SMark F. Adams ierr = MatAssemblyEnd(a_Prol,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9312e68589bSMark F. Adams 932c8b0795cSMark F. Adams /* ierr = MPI_Allreduce( &ndone, &ii, 1, MPIU_INT, MPIU_SUM, wcomm ); */ 9332e68589bSMark F. Adams /* MatGetSize( a_Prol, &kk, &jj ); */ 934c8b0795cSMark F. Adams /* ierr = MPI_Allreduce( &minsz, &jj, 1, MPIU_INT, MPIU_MIN, wcomm ); */ 935e78576d6SMark F. Adams /* PetscPrintf(wcomm," **** [%d]%s %d total done, %d nodes (%d local done), min agg. size = %d\n",mype,__FUNCT__,ii,kk/bs,ndone,jj); */ 9362e68589bSMark F. Adams 9379057884aSMark F. Adams #ifdef OUT_AGGS 938b2a4f308SMark F. Adams if(llev==1) fclose(file); 9399057884aSMark F. Adams #endif 9400cbbd2e1SMark F. Adams ierr = GAMGTableDestroy( &fgid_flid ); CHKERRQ(ierr); 9412e68589bSMark F. Adams 9422e68589bSMark F. Adams PetscFunctionReturn(0); 9432e68589bSMark F. Adams } 9442e68589bSMark F. Adams 9452e68589bSMark F. Adams /* -------------------------------------------------------------------------- */ 9462e68589bSMark F. Adams /* 947c8b0795cSMark F. Adams PCGAMGgraph_AGG 9482e68589bSMark F. Adams 9492e68589bSMark F. Adams Input Parameter: 9502e68589bSMark F. Adams . pc - this 9512e68589bSMark F. Adams . Amat - matrix on this fine level 9522e68589bSMark F. Adams Output Parameter: 953c8b0795cSMark F. Adams . a_Gmat - 9542e68589bSMark F. Adams */ 9552e68589bSMark F. Adams #undef __FUNCT__ 956c8b0795cSMark F. Adams #define __FUNCT__ "PCGAMGgraph_AGG" 957c8b0795cSMark F. Adams PetscErrorCode PCGAMGgraph_AGG( PC pc, 9582e68589bSMark F. Adams const Mat Amat, 959c8b0795cSMark F. Adams Mat *a_Gmat 960c8b0795cSMark F. Adams ) 961c8b0795cSMark F. Adams { 962c8b0795cSMark F. Adams PetscErrorCode ierr; 963c8b0795cSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 964c8b0795cSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 965c8b0795cSMark F. Adams const PetscInt verbose = pc_gamg->verbose; 966c8b0795cSMark F. Adams const PetscReal vfilter = pc_gamg->threshold; 967c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 968c8b0795cSMark F. Adams PetscMPIInt mype,npe; 969e0940f08SMark F. Adams Mat Gmat; 970c8b0795cSMark F. Adams MPI_Comm wcomm = ((PetscObject)Amat)->comm; 9710cbbd2e1SMark F. Adams PetscBool set,flg,symm; 972c8b0795cSMark F. Adams 973c8b0795cSMark F. Adams PetscFunctionBegin; 9740cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 9750cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(PC_GAMGGgraph_AGG,0,0,0,0);CHKERRQ(ierr); 9760cbbd2e1SMark F. Adams #endif 977c8b0795cSMark F. Adams ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 978c8b0795cSMark F. Adams ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 979c8b0795cSMark F. Adams 9800cbbd2e1SMark F. Adams ierr = MatIsSymmetricKnown(Amat, &set, &flg); CHKERRQ(ierr); 981263489e9SJed Brown symm = (PetscBool)(pc_gamg_agg->sym_graph || !(set && flg)); 9820cbbd2e1SMark F. Adams 9832d7fac45SMark F. Adams ierr = PCGAMGCreateGraph( Amat, &Gmat ); CHKERRQ( ierr ); 9842d7fac45SMark F. Adams ierr = PCGAMGFilterGraph( &Gmat, vfilter, symm, verbose ); CHKERRQ( ierr ); 985c8b0795cSMark F. Adams 986e0940f08SMark F. Adams *a_Gmat = Gmat; 987c8b0795cSMark F. Adams 9880cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 9890cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(PC_GAMGGgraph_AGG,0,0,0,0);CHKERRQ(ierr); 9900cbbd2e1SMark F. Adams #endif 991c8b0795cSMark F. Adams PetscFunctionReturn(0); 992c8b0795cSMark F. Adams } 993c8b0795cSMark F. Adams 994c8b0795cSMark F. Adams /* -------------------------------------------------------------------------- */ 995c8b0795cSMark F. Adams /* 996b43b03e9SMark F. Adams PCGAMGCoarsen_AGG 997c8b0795cSMark F. Adams 998c8b0795cSMark F. Adams Input Parameter: 999e0940f08SMark F. Adams . a_pc - this 1000e0940f08SMark F. Adams Input/Output Parameter: 10010cbbd2e1SMark F. Adams . a_Gmat1 - graph on this fine level - coarsening can change this (squares it) 1002c8b0795cSMark F. Adams Output Parameter: 10030cbbd2e1SMark F. Adams . agg_lists - list of aggregates 1004c8b0795cSMark F. Adams */ 1005c8b0795cSMark F. Adams #undef __FUNCT__ 1006b43b03e9SMark F. Adams #define __FUNCT__ "PCGAMGCoarsen_AGG" 1007e0940f08SMark F. Adams PetscErrorCode PCGAMGCoarsen_AGG( PC a_pc, 1008e0940f08SMark F. Adams Mat *a_Gmat1, 10090cbbd2e1SMark F. Adams PetscCoarsenData **agg_lists 1010c8b0795cSMark F. Adams ) 1011c8b0795cSMark F. Adams { 1012c8b0795cSMark F. Adams PetscErrorCode ierr; 1013e0940f08SMark F. Adams PC_MG *mg = (PC_MG*)a_pc->data; 1014c8b0795cSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 1015c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 10160cbbd2e1SMark F. Adams Mat mat,Gmat2, Gmat1 = *a_Gmat1; /* squared graph */ 10170cbbd2e1SMark F. Adams IS perm; 1018c8b0795cSMark F. Adams PetscInt Ii,nloc,bs,n,m; 1019c8b0795cSMark F. Adams PetscInt *permute; 1020c8b0795cSMark F. Adams PetscBool *bIndexSet; 1021b43b03e9SMark F. Adams MatCoarsen crs; 1022e0940f08SMark F. Adams MPI_Comm wcomm = ((PetscObject)Gmat1)->comm; 10230cbbd2e1SMark F. Adams PetscMPIInt mype,npe; 1024c8b0795cSMark F. Adams 1025c8b0795cSMark F. Adams PetscFunctionBegin; 10260cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 10270cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(PC_GAMGCoarsen_AGG,0,0,0,0);CHKERRQ(ierr); 10280cbbd2e1SMark F. Adams #endif 10290cbbd2e1SMark F. Adams ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 10300cbbd2e1SMark F. Adams ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 1031e0940f08SMark F. Adams ierr = MatGetLocalSize( Gmat1, &n, &m ); CHKERRQ(ierr); 1032e0940f08SMark F. Adams ierr = MatGetBlockSize( Gmat1, &bs ); CHKERRQ(ierr); assert(bs==1); 1033c8b0795cSMark F. Adams nloc = n/bs; 1034c8b0795cSMark F. Adams 1035e0940f08SMark F. Adams if( pc_gamg_agg->square_graph ) { 1036e0940f08SMark F. Adams ierr = MatTransposeMatMult( Gmat1, Gmat1, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Gmat2 ); 1037e0940f08SMark F. Adams CHKERRQ(ierr); 1038e0940f08SMark F. Adams } 1039e0940f08SMark F. Adams else Gmat2 = Gmat1; 1040c8b0795cSMark F. Adams 1041c8b0795cSMark F. Adams /* get MIS aggs */ 1042c8b0795cSMark F. Adams /* randomize */ 1043c8b0795cSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscInt), &permute ); CHKERRQ(ierr); 1044c8b0795cSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscBool), &bIndexSet ); CHKERRQ(ierr); 1045c8b0795cSMark F. Adams for ( Ii = 0; Ii < nloc ; Ii++ ){ 1046c8b0795cSMark F. Adams bIndexSet[Ii] = PETSC_FALSE; 1047c8b0795cSMark F. Adams permute[Ii] = Ii; 1048c8b0795cSMark F. Adams } 1049c8b0795cSMark F. Adams srand(1); /* make deterministic */ 1050c8b0795cSMark F. Adams for ( Ii = 0; Ii < nloc ; Ii++ ) { 1051c8b0795cSMark F. Adams PetscInt iSwapIndex = rand()%nloc; 1052c8b0795cSMark F. Adams if (!bIndexSet[iSwapIndex] && iSwapIndex != Ii) { 1053c8b0795cSMark F. Adams PetscInt iTemp = permute[iSwapIndex]; 1054c8b0795cSMark F. Adams permute[iSwapIndex] = permute[Ii]; 1055c8b0795cSMark F. Adams permute[Ii] = iTemp; 1056c8b0795cSMark F. Adams bIndexSet[iSwapIndex] = PETSC_TRUE; 1057c8b0795cSMark F. Adams } 1058c8b0795cSMark F. Adams } 1059c8b0795cSMark F. Adams ierr = PetscFree( bIndexSet ); CHKERRQ(ierr); 1060c8b0795cSMark F. Adams 1061c8b0795cSMark F. Adams ierr = ISCreateGeneral(PETSC_COMM_SELF, nloc, permute, PETSC_USE_POINTER, &perm); 1062c8b0795cSMark F. Adams CHKERRQ(ierr); 10630cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 10640cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET4],0,0,0,0);CHKERRQ(ierr); 1065b43b03e9SMark F. Adams #endif 1066b43b03e9SMark F. Adams ierr = MatCoarsenCreate( wcomm, &crs ); CHKERRQ(ierr); 10679057884aSMark F. Adams /* ierr = MatCoarsenSetType( crs, MATCOARSENMIS ); CHKERRQ(ierr); */ 10689057884aSMark F. Adams ierr = MatCoarsenSetFromOptions( crs ); CHKERRQ(ierr); 1069b43b03e9SMark F. Adams ierr = MatCoarsenSetGreedyOrdering( crs, perm ); CHKERRQ(ierr); 1070b43b03e9SMark F. Adams ierr = MatCoarsenSetAdjacency( crs, Gmat2 ); CHKERRQ(ierr); 1071b43b03e9SMark F. Adams ierr = MatCoarsenSetVerbose( crs, pc_gamg->verbose ); CHKERRQ(ierr); 1072b43b03e9SMark F. Adams ierr = MatCoarsenSetStrictAggs( crs, PETSC_TRUE ); CHKERRQ(ierr); 1073b43b03e9SMark F. Adams ierr = MatCoarsenApply( crs ); CHKERRQ(ierr); 10740cbbd2e1SMark F. Adams ierr = MatCoarsenGetData( crs, agg_lists ); CHKERRQ(ierr); /* output */ 1075b43b03e9SMark F. Adams ierr = MatCoarsenDestroy( &crs ); CHKERRQ(ierr); 1076b43b03e9SMark F. Adams 1077c8b0795cSMark F. Adams ierr = ISDestroy( &perm ); CHKERRQ(ierr); 1078c8b0795cSMark F. Adams ierr = PetscFree( permute ); CHKERRQ(ierr); 10790cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 10800cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET4],0,0,0,0);CHKERRQ(ierr); 1081b43b03e9SMark F. Adams #endif 1082c8b0795cSMark F. Adams /* smooth aggs */ 1083e0940f08SMark F. Adams if( Gmat2 != Gmat1 ) { 10840cbbd2e1SMark F. Adams const PetscCoarsenData *llist = *agg_lists; 10850cbbd2e1SMark F. Adams ierr = smoothAggs( Gmat2, Gmat1, *agg_lists ); CHKERRQ(ierr); 1086c8b0795cSMark F. Adams ierr = MatDestroy( &Gmat1 ); CHKERRQ(ierr); 1087e0940f08SMark F. Adams *a_Gmat1 = Gmat2; /* output */ 108841b27cdeSMark F. Adams ierr = PetscCDGetMat( llist, &mat ); CHKERRQ(ierr); 10890cbbd2e1SMark F. Adams if(mat) SETERRQ(wcomm,PETSC_ERR_ARG_WRONG, "Auxilary matrix with squared graph????"); 1090ef4ad70eSMark F. Adams } 10910cbbd2e1SMark F. Adams else { 10920cbbd2e1SMark F. Adams const PetscCoarsenData *llist = *agg_lists; 10930cbbd2e1SMark F. Adams /* see if we have a matrix that takes pecedence (returned from MatCoarsenAppply) */ 109441b27cdeSMark F. Adams ierr = PetscCDGetMat( llist, &mat ); CHKERRQ(ierr); 10950cbbd2e1SMark F. Adams if( mat ) { 10960cbbd2e1SMark F. Adams ierr = MatDestroy( &Gmat1 ); CHKERRQ(ierr); 10970cbbd2e1SMark F. Adams *a_Gmat1 = mat; /* output */ 10980cbbd2e1SMark F. Adams } 10990cbbd2e1SMark F. Adams } 11000cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 11010cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(PC_GAMGCoarsen_AGG,0,0,0,0);CHKERRQ(ierr); 11020cbbd2e1SMark F. Adams #endif 1103c8b0795cSMark F. Adams PetscFunctionReturn(0); 1104c8b0795cSMark F. Adams } 1105c8b0795cSMark F. Adams 1106c8b0795cSMark F. Adams /* -------------------------------------------------------------------------- */ 1107c8b0795cSMark F. Adams /* 11080cbbd2e1SMark F. Adams PCGAMGProlongator_AGG 1109c8b0795cSMark F. Adams 1110c8b0795cSMark F. Adams Input Parameter: 1111c8b0795cSMark F. Adams . pc - this 1112c8b0795cSMark F. Adams . Amat - matrix on this fine level 1113c8b0795cSMark F. Adams . Graph - used to get ghost data for nodes in 11140cbbd2e1SMark F. Adams . agg_lists - list of aggregates 1115c8b0795cSMark F. Adams Output Parameter: 1116c8b0795cSMark F. Adams . a_P_out - prolongation operator to the next level 1117c8b0795cSMark F. Adams */ 1118c8b0795cSMark F. Adams #undef __FUNCT__ 11190cbbd2e1SMark F. Adams #define __FUNCT__ "PCGAMGProlongator_AGG" 11200cbbd2e1SMark F. Adams PetscErrorCode PCGAMGProlongator_AGG( PC pc, 1121c8b0795cSMark F. Adams const Mat Amat, 1122c8b0795cSMark F. Adams const Mat Gmat, 11230cbbd2e1SMark F. Adams PetscCoarsenData *agg_lists, 1124c8b0795cSMark F. Adams Mat *a_P_out 11252e68589bSMark F. Adams ) 11262e68589bSMark F. Adams { 11272e68589bSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 11282e68589bSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 11292e68589bSMark F. Adams const PetscInt verbose = pc_gamg->verbose; 1130c8b0795cSMark F. Adams const PetscInt data_cols = pc_gamg->data_cell_cols; 11312e68589bSMark F. Adams PetscErrorCode ierr; 1132c8b0795cSMark F. Adams PetscInt Istart,Iend,nloc,ii,jj,kk,my0,nLocalSelected,bs; 1133c8b0795cSMark F. Adams Mat Prol; 11342e68589bSMark F. Adams PetscMPIInt mype, npe; 11352e68589bSMark F. Adams MPI_Comm wcomm = ((PetscObject)Amat)->comm; 11360cbbd2e1SMark F. Adams const PetscInt col_bs=data_cols; 1137c8b0795cSMark F. Adams PetscReal *data_w_ghost; 1138c8b0795cSMark F. Adams PetscInt myCrs0, nbnodes=0, *flid_fgid; 11392e68589bSMark F. Adams 11402e68589bSMark F. Adams PetscFunctionBegin; 11410cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 11420cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(PC_GAMGProlongator_AGG,0,0,0,0);CHKERRQ(ierr); 11430cbbd2e1SMark F. Adams #endif 11442e68589bSMark F. Adams ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 11452e68589bSMark F. Adams ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 11462e68589bSMark F. Adams ierr = MatGetOwnershipRange( Amat, &Istart, &Iend ); CHKERRQ(ierr); 1147c8b0795cSMark F. Adams ierr = MatGetBlockSize( Amat, &bs ); CHKERRQ( ierr ); 1148c8b0795cSMark F. Adams nloc = (Iend-Istart)/bs; my0 = Istart/bs; assert((Iend-Istart)%bs==0); 11492e68589bSMark F. Adams 11502e68589bSMark F. Adams /* get 'nLocalSelected' */ 11510cbbd2e1SMark F. Adams for( ii=0, nLocalSelected = 0 ; ii < nloc ; ii++ ){ 1152e78576d6SMark F. Adams PetscBool ise; 11530cbbd2e1SMark F. Adams /* filter out singletons 0 or 1? */ 1154e78576d6SMark F. Adams ierr = PetscCDEmptyAt( agg_lists, ii, &ise ); CHKERRQ(ierr); 1155e78576d6SMark F. Adams if( !ise ) nLocalSelected++; 11562e68589bSMark F. Adams } 11572e68589bSMark F. Adams 11582e68589bSMark F. Adams /* create prolongator, create P matrix */ 115969b1f4b7SBarry Smith ierr = MatCreateAIJ( wcomm, 1160c8b0795cSMark F. Adams nloc*bs, nLocalSelected*col_bs, 11612e68589bSMark F. Adams PETSC_DETERMINE, PETSC_DETERMINE, 11622e68589bSMark F. Adams data_cols, PETSC_NULL, data_cols, PETSC_NULL, 11632e68589bSMark F. Adams &Prol ); 11642e68589bSMark F. Adams CHKERRQ(ierr); 11652e68589bSMark F. Adams 11662e68589bSMark F. Adams /* can get all points "removed" */ 1167c8b0795cSMark F. Adams ierr = MatGetSize( Prol, &kk, &ii ); CHKERRQ(ierr); 1168c8b0795cSMark F. Adams if( ii==0 ) { 11692e68589bSMark F. Adams if( verbose ) { 1170c8b0795cSMark F. Adams PetscPrintf(wcomm,"[%d]%s no selected points on coarse grid\n",mype,__FUNCT__); 11712e68589bSMark F. Adams } 11722e68589bSMark F. Adams ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 11732e68589bSMark F. Adams *a_P_out = PETSC_NULL; /* out */ 11742e68589bSMark F. Adams PetscFunctionReturn(0); 11752e68589bSMark F. Adams } 1176c8b0795cSMark F. Adams if( verbose ) { 1177e78576d6SMark F. Adams PetscPrintf(wcomm,"\t\t[%d]%s New grid %d nodes\n",mype,__FUNCT__,ii/col_bs); 1178c8b0795cSMark F. Adams } 1179c8b0795cSMark F. Adams ierr = MatGetOwnershipRangeColumn( Prol, &myCrs0, &kk ); CHKERRQ(ierr); 11800cbbd2e1SMark F. Adams 11810cbbd2e1SMark F. Adams assert((kk-myCrs0)%col_bs==0); 1182c8b0795cSMark F. Adams myCrs0 = myCrs0/col_bs; 11830cbbd2e1SMark F. Adams assert((kk/col_bs-myCrs0)==nLocalSelected); 11842e68589bSMark F. Adams 11852e68589bSMark F. Adams /* create global vector of data in 'data_w_ghost' */ 11860cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 11870cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET7],0,0,0,0);CHKERRQ(ierr); 11882e68589bSMark F. Adams #endif 1189c8b0795cSMark F. Adams if (npe > 1) { /* */ 11902e68589bSMark F. Adams PetscReal *tmp_gdata,*tmp_ldata,*tp2; 11912e68589bSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscReal), &tmp_ldata ); CHKERRQ(ierr); 11922e68589bSMark F. Adams for( jj = 0 ; jj < data_cols ; jj++ ){ 1193c8b0795cSMark F. Adams for( kk = 0 ; kk < bs ; kk++) { 11942e68589bSMark F. Adams PetscInt ii,nnodes; 1195c8b0795cSMark F. Adams const PetscReal *tp = pc_gamg->data + jj*bs*nloc + kk; 1196c8b0795cSMark F. Adams for( ii = 0 ; ii < nloc ; ii++, tp += bs ){ 11972e68589bSMark F. Adams tmp_ldata[ii] = *tp; 11982e68589bSMark F. Adams } 11990cbbd2e1SMark F. Adams ierr = PCGAMGGetDataWithGhosts( Gmat, 1, tmp_ldata, &nnodes, &tmp_gdata ); 12002e68589bSMark F. Adams CHKERRQ(ierr); 12012e68589bSMark F. Adams if(jj==0 && kk==0) { /* now I know how many todal nodes - allocate */ 1202c8b0795cSMark F. Adams ierr = PetscMalloc( nnodes*bs*data_cols*sizeof(PetscReal), &data_w_ghost ); CHKERRQ(ierr); 1203c8b0795cSMark F. Adams nbnodes = bs*nnodes; 12042e68589bSMark F. Adams } 1205c8b0795cSMark F. Adams tp2 = data_w_ghost + jj*bs*nnodes + kk; 1206c8b0795cSMark F. Adams for( ii = 0 ; ii < nnodes ; ii++, tp2 += bs ){ 12072e68589bSMark F. Adams *tp2 = tmp_gdata[ii]; 12082e68589bSMark F. Adams } 12092e68589bSMark F. Adams ierr = PetscFree( tmp_gdata ); CHKERRQ(ierr); 12102e68589bSMark F. Adams } 12112e68589bSMark F. Adams } 12122e68589bSMark F. Adams ierr = PetscFree( tmp_ldata ); CHKERRQ(ierr); 12132e68589bSMark F. Adams } 12142e68589bSMark F. Adams else { 1215c8b0795cSMark F. Adams nbnodes = bs*nloc; 1216c8b0795cSMark F. Adams data_w_ghost = (PetscReal*)pc_gamg->data; 12172e68589bSMark F. Adams } 12182e68589bSMark F. Adams 12192e68589bSMark F. Adams /* get P0 */ 12202e68589bSMark F. Adams if( npe > 1 ){ 12212e68589bSMark F. Adams PetscReal *fid_glid_loc,*fiddata; 12222e68589bSMark F. Adams PetscInt nnodes; 12232e68589bSMark F. Adams 12242e68589bSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscReal), &fid_glid_loc ); CHKERRQ(ierr); 12252e68589bSMark F. Adams for(kk=0;kk<nloc;kk++) fid_glid_loc[kk] = (PetscReal)(my0+kk); 12260cbbd2e1SMark F. Adams ierr = PCGAMGGetDataWithGhosts( Gmat, 1, fid_glid_loc, &nnodes, &fiddata ); 12272e68589bSMark F. Adams CHKERRQ(ierr); 12282e68589bSMark F. Adams ierr = PetscMalloc( nnodes*sizeof(PetscInt), &flid_fgid ); CHKERRQ(ierr); 12292e68589bSMark F. Adams for(kk=0;kk<nnodes;kk++) flid_fgid[kk] = (PetscInt)fiddata[kk]; 12302e68589bSMark F. Adams ierr = PetscFree( fiddata ); CHKERRQ(ierr); 1231c8b0795cSMark F. Adams assert(nnodes==nbnodes/bs); 12322e68589bSMark F. Adams ierr = PetscFree( fid_glid_loc ); CHKERRQ(ierr); 12332e68589bSMark F. Adams } 12342e68589bSMark F. Adams else { 12352e68589bSMark F. Adams ierr = PetscMalloc( nloc*sizeof(PetscInt), &flid_fgid ); CHKERRQ(ierr); 12362e68589bSMark F. Adams for(kk=0;kk<nloc;kk++) flid_fgid[kk] = my0 + kk; 12372e68589bSMark F. Adams } 12380cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 12390cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET7],0,0,0,0);CHKERRQ(ierr); 12400cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET8],0,0,0,0);CHKERRQ(ierr); 12412e68589bSMark F. Adams #endif 1242c8b0795cSMark F. Adams { 1243ffc955d6SMark F. Adams PetscReal *data_out = PETSC_NULL; 12440cbbd2e1SMark F. Adams ierr = formProl0( agg_lists, bs, data_cols, myCrs0, nbnodes, 1245c8b0795cSMark F. Adams data_w_ghost, flid_fgid, &data_out, Prol ); 12462e68589bSMark F. Adams CHKERRQ(ierr); 1247c8b0795cSMark F. Adams ierr = PetscFree( pc_gamg->data ); CHKERRQ( ierr ); 1248c8b0795cSMark F. Adams pc_gamg->data = data_out; 1249c8b0795cSMark F. Adams pc_gamg->data_cell_rows = data_cols; 1250c8b0795cSMark F. Adams pc_gamg->data_sz = data_cols*data_cols*nLocalSelected; 1251c8b0795cSMark F. Adams } 12520cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 12530cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET8],0,0,0,0);CHKERRQ(ierr); 1254c8b0795cSMark F. Adams #endif 12552e68589bSMark F. Adams if (npe > 1) ierr = PetscFree( data_w_ghost ); CHKERRQ(ierr); 12562e68589bSMark F. Adams ierr = PetscFree( flid_fgid ); CHKERRQ(ierr); 12572e68589bSMark F. Adams 1258c8b0795cSMark F. Adams /* attach block size of columns */ 1259c8b0795cSMark F. Adams if( pc_gamg->col_bs_id == -1 ) { 1260c8b0795cSMark F. Adams ierr = PetscObjectComposedDataRegister( &pc_gamg->col_bs_id ); assert(pc_gamg->col_bs_id != -1 ); 1261c8b0795cSMark F. Adams } 1262c8b0795cSMark F. Adams ierr = PetscObjectComposedDataSetInt( (PetscObject)Prol, pc_gamg->col_bs_id, data_cols ); CHKERRQ(ierr); 1263c8b0795cSMark F. Adams 1264c8b0795cSMark F. Adams *a_P_out = Prol; /* out */ 12650cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 12660cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(PC_GAMGProlongator_AGG,0,0,0,0);CHKERRQ(ierr); 12670cbbd2e1SMark F. Adams #endif 1268c8b0795cSMark F. Adams PetscFunctionReturn(0); 1269c8b0795cSMark F. Adams } 1270c8b0795cSMark F. Adams 1271c8b0795cSMark F. Adams /* -------------------------------------------------------------------------- */ 1272c8b0795cSMark F. Adams /* 12730cbbd2e1SMark F. Adams PCGAMGOptprol_AGG 1274c8b0795cSMark F. Adams 1275c8b0795cSMark F. Adams Input Parameter: 1276c8b0795cSMark F. Adams . pc - this 1277c8b0795cSMark F. Adams . Amat - matrix on this fine level 1278c8b0795cSMark F. Adams In/Output Parameter: 1279c8b0795cSMark F. Adams . a_P_out - prolongation operator to the next level 1280c8b0795cSMark F. Adams */ 1281c8b0795cSMark F. Adams #undef __FUNCT__ 12820cbbd2e1SMark F. Adams #define __FUNCT__ "PCGAMGOptprol_AGG" 12830cbbd2e1SMark F. Adams PetscErrorCode PCGAMGOptprol_AGG( PC pc, 1284c8b0795cSMark F. Adams const Mat Amat, 1285c8b0795cSMark F. Adams Mat *a_P 1286c8b0795cSMark F. Adams ) 1287c8b0795cSMark F. Adams { 1288c8b0795cSMark F. Adams PetscErrorCode ierr; 1289c8b0795cSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 1290c8b0795cSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 1291c8b0795cSMark F. Adams const PetscInt verbose = pc_gamg->verbose; 1292c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg = (PC_GAMG_AGG*)pc_gamg->subctx; 1293c8b0795cSMark F. Adams PetscInt jj; 1294c8b0795cSMark F. Adams PetscMPIInt mype,npe; 1295c8b0795cSMark F. Adams Mat Prol = *a_P; 1296c8b0795cSMark F. Adams MPI_Comm wcomm = ((PetscObject)Amat)->comm; 1297c8b0795cSMark F. Adams 1298c8b0795cSMark F. Adams PetscFunctionBegin; 12990cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 13000cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(PC_GAMGOptprol_AGG,0,0,0,0);CHKERRQ(ierr); 13010cbbd2e1SMark F. Adams #endif 1302c8b0795cSMark F. Adams ierr = MPI_Comm_rank( wcomm, &mype); CHKERRQ(ierr); 1303c8b0795cSMark F. Adams ierr = MPI_Comm_size( wcomm, &npe); CHKERRQ(ierr); 1304c8b0795cSMark F. Adams 13052e68589bSMark F. Adams /* smooth P0 */ 1306c8b0795cSMark F. Adams for( jj = 0 ; jj < pc_gamg_agg->nsmooths ; jj++ ){ 13072e68589bSMark F. Adams Mat tMat; 13082e68589bSMark F. Adams Vec diag; 13092e68589bSMark F. Adams PetscReal alpha, emax, emin; 13100cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 13110cbbd2e1SMark F. Adams ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET9],0,0,0,0);CHKERRQ(ierr); 13122e68589bSMark F. Adams #endif 13132e68589bSMark F. Adams if( jj == 0 ) { 13142e68589bSMark F. Adams KSP eksp; 13152e68589bSMark F. Adams Vec bb, xx; 13162e68589bSMark F. Adams PC pc; 13172e68589bSMark F. Adams ierr = MatGetVecs( Amat, &bb, 0 ); CHKERRQ(ierr); 13182e68589bSMark F. Adams ierr = MatGetVecs( Amat, &xx, 0 ); CHKERRQ(ierr); 13192e68589bSMark F. Adams { 13202e68589bSMark F. Adams PetscRandom rctx; 13212e68589bSMark F. Adams ierr = PetscRandomCreate(wcomm,&rctx);CHKERRQ(ierr); 13222e68589bSMark F. Adams ierr = PetscRandomSetFromOptions(rctx);CHKERRQ(ierr); 13232e68589bSMark F. Adams ierr = VecSetRandom(bb,rctx);CHKERRQ(ierr); 13242e68589bSMark F. Adams ierr = PetscRandomDestroy( &rctx ); CHKERRQ(ierr); 13252e68589bSMark F. Adams } 13262e68589bSMark F. Adams ierr = KSPCreate(wcomm,&eksp); CHKERRQ(ierr); 1327db36e5aeSMark F. Adams ierr = KSPAppendOptionsPrefix( eksp, "gamg_est_"); CHKERRQ(ierr); 13282e68589bSMark F. Adams ierr = KSPSetFromOptions( eksp ); CHKERRQ(ierr); 13292e68589bSMark F. Adams ierr = KSPSetInitialGuessNonzero( eksp, PETSC_FALSE ); CHKERRQ(ierr); 13302e68589bSMark F. Adams ierr = KSPSetOperators( eksp, Amat, Amat, SAME_NONZERO_PATTERN ); 13312e68589bSMark F. Adams CHKERRQ( ierr ); 13322e68589bSMark F. Adams ierr = KSPGetPC( eksp, &pc ); CHKERRQ( ierr ); 13332e68589bSMark F. Adams ierr = PCSetType( pc, PCJACOBI ); CHKERRQ(ierr); /* smoother */ 13342e68589bSMark F. Adams ierr = KSPSetTolerances(eksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,10); 13352e68589bSMark F. Adams CHKERRQ(ierr); 13362e68589bSMark F. Adams ierr = KSPSetNormType( eksp, KSP_NORM_NONE ); CHKERRQ(ierr); 13372e68589bSMark F. Adams ierr = KSPSetComputeSingularValues( eksp,PETSC_TRUE ); CHKERRQ(ierr); 13382e68589bSMark F. Adams 13392e68589bSMark F. Adams /* solve - keep stuff out of logging */ 13402e68589bSMark F. Adams ierr = PetscLogEventDeactivate(KSP_Solve);CHKERRQ(ierr); 13412e68589bSMark F. Adams ierr = PetscLogEventDeactivate(PC_Apply);CHKERRQ(ierr); 13422e68589bSMark F. Adams ierr = KSPSolve( eksp, bb, xx ); CHKERRQ(ierr); 13432e68589bSMark F. Adams ierr = PetscLogEventActivate(KSP_Solve);CHKERRQ(ierr); 13442e68589bSMark F. Adams ierr = PetscLogEventActivate(PC_Apply);CHKERRQ(ierr); 13452e68589bSMark F. Adams 13462e68589bSMark F. Adams ierr = KSPComputeExtremeSingularValues( eksp, &emax, &emin ); CHKERRQ(ierr); 13472e68589bSMark F. Adams if( verbose ) { 1348c8b0795cSMark F. Adams PetscPrintf(wcomm,"\t\t\t%s smooth P0: max eigen=%e min=%e PC=%s\n", 13492e68589bSMark F. Adams __FUNCT__,emax,emin,PCJACOBI); 13502e68589bSMark F. Adams } 13512e68589bSMark F. Adams ierr = VecDestroy( &xx ); CHKERRQ(ierr); 13522e68589bSMark F. Adams ierr = VecDestroy( &bb ); CHKERRQ(ierr); 13532e68589bSMark F. Adams ierr = KSPDestroy( &eksp ); CHKERRQ(ierr); 13542e68589bSMark F. Adams 13552e68589bSMark F. Adams if( pc_gamg->emax_id == -1 ) { 13562e68589bSMark F. Adams ierr = PetscObjectComposedDataRegister( &pc_gamg->emax_id ); 13572e68589bSMark F. Adams assert(pc_gamg->emax_id != -1 ); 13582e68589bSMark F. Adams } 13592e68589bSMark F. Adams ierr = PetscObjectComposedDataSetScalar( (PetscObject)Amat, pc_gamg->emax_id, emax ); CHKERRQ(ierr); 13602e68589bSMark F. Adams } 13612e68589bSMark F. Adams 13622e68589bSMark F. Adams /* smooth P1 := (I - omega/lam D^{-1}A)P0 */ 13632e68589bSMark F. Adams ierr = MatMatMult( Amat, Prol, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &tMat ); CHKERRQ(ierr); 13642e68589bSMark F. Adams ierr = MatGetVecs( Amat, &diag, 0 ); CHKERRQ(ierr); 13652e68589bSMark F. Adams ierr = MatGetDiagonal( Amat, diag ); CHKERRQ(ierr); /* effectively PCJACOBI */ 13662e68589bSMark F. Adams ierr = VecReciprocal( diag ); CHKERRQ(ierr); 13672e68589bSMark F. Adams ierr = MatDiagonalScale( tMat, diag, 0 ); CHKERRQ(ierr); 13682e68589bSMark F. Adams ierr = VecDestroy( &diag ); CHKERRQ(ierr); 13692e68589bSMark F. Adams alpha = -1.5/emax; 13702e68589bSMark F. Adams ierr = MatAYPX( tMat, alpha, Prol, SUBSET_NONZERO_PATTERN ); CHKERRQ(ierr); 13712e68589bSMark F. Adams ierr = MatDestroy( &Prol ); CHKERRQ(ierr); 13722e68589bSMark F. Adams Prol = tMat; 13730cbbd2e1SMark F. Adams #if defined PETSC_GAMG_USE_LOG 13740cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET9],0,0,0,0);CHKERRQ(ierr); 13752e68589bSMark F. Adams #endif 13762e68589bSMark F. Adams } 13770cbbd2e1SMark F. Adams #if defined PETSC_USE_LOG 13780cbbd2e1SMark F. Adams ierr = PetscLogEventEnd(PC_GAMGOptprol_AGG,0,0,0,0);CHKERRQ(ierr); 13790cbbd2e1SMark F. Adams #endif 1380c8b0795cSMark F. Adams *a_P = Prol; 1381c8b0795cSMark F. Adams 1382c8b0795cSMark F. Adams PetscFunctionReturn(0); 13832e68589bSMark F. Adams } 13842e68589bSMark F. Adams 1385c8b0795cSMark F. Adams /* -------------------------------------------------------------------------- */ 1386c8b0795cSMark F. Adams /* 1387c8b0795cSMark F. Adams PCCreateGAMG_AGG 13882e68589bSMark F. Adams 1389c8b0795cSMark F. Adams Input Parameter: 1390c8b0795cSMark F. Adams . pc - 1391c8b0795cSMark F. Adams */ 1392c8b0795cSMark F. Adams #undef __FUNCT__ 1393c8b0795cSMark F. Adams #define __FUNCT__ "PCCreateGAMG_AGG" 1394c8b0795cSMark F. Adams PetscErrorCode PCCreateGAMG_AGG( PC pc ) 1395c8b0795cSMark F. Adams { 1396c8b0795cSMark F. Adams PetscErrorCode ierr; 1397c8b0795cSMark F. Adams PC_MG *mg = (PC_MG*)pc->data; 1398c8b0795cSMark F. Adams PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx; 1399c8b0795cSMark F. Adams PC_GAMG_AGG *pc_gamg_agg; 14002e68589bSMark F. Adams 1401c8b0795cSMark F. Adams PetscFunctionBegin; 1402c8b0795cSMark F. Adams /* create sub context for SA */ 1403c8b0795cSMark F. Adams ierr = PetscNewLog( pc, PC_GAMG_AGG, &pc_gamg_agg ); CHKERRQ(ierr); 1404c8b0795cSMark F. Adams assert(!pc_gamg->subctx); 1405c8b0795cSMark F. Adams pc_gamg->subctx = pc_gamg_agg; 1406c8b0795cSMark F. Adams 1407c8b0795cSMark F. Adams pc->ops->setfromoptions = PCSetFromOptions_GAMG_AGG; 1408c8b0795cSMark F. Adams pc->ops->destroy = PCDestroy_AGG; 1409c8b0795cSMark F. Adams /* reset does not do anything; setup not virtual */ 1410c8b0795cSMark F. Adams 1411c8b0795cSMark F. Adams /* set internal function pointers */ 1412c8b0795cSMark F. Adams pc_gamg->graph = PCGAMGgraph_AGG; 1413b43b03e9SMark F. Adams pc_gamg->coarsen = PCGAMGCoarsen_AGG; 14140cbbd2e1SMark F. Adams pc_gamg->prolongator = PCGAMGProlongator_AGG; 14150cbbd2e1SMark F. Adams pc_gamg->optprol = PCGAMGOptprol_AGG; 1416c8b0795cSMark F. Adams 1417c8b0795cSMark F. Adams pc_gamg->createdefaultdata = PCSetData_AGG; 1418c8b0795cSMark F. Adams 1419c8b0795cSMark F. Adams ierr = PetscObjectComposeFunctionDynamic( (PetscObject)pc, 1420c8b0795cSMark F. Adams "PCSetCoordinates_C", 1421c8b0795cSMark F. Adams "PCSetCoordinates_AGG", 1422c8b0795cSMark F. Adams PCSetCoordinates_AGG); 14232e68589bSMark F. Adams PetscFunctionReturn(0); 14242e68589bSMark F. Adams } 1425