PetscErrorCode PCSetUp_ML(PC pc) { PetscErrorCode ierr; PetscMPIInt size; FineGridCtx *PetscMLdata; ML *ml_object; ML_Aggregate *agg_object; ML_Operator *mlmat; PetscInt nlocal_allcols,Nlevels,mllevel,level,level1,m,fine_level,bs; Mat A,Aloc; GridCtx *gridctx; PC_MG *mg = (PC_MG*)pc->data; PC_ML *pc_ml = (PC_ML*)mg->innerctx; PetscBool isSeq, isMPI; KSP smoother; PC subpc; PetscInt mesh_level, old_mesh_level; PetscFunctionBegin; A = pc->pmat; ierr = MPI_Comm_size(((PetscObject)A)->comm,&size);CHKERRQ(ierr); if (pc->setupcalled) { if (pc->flag == SAME_NONZERO_PATTERN && pc_ml->reuse_interpolation) { /* Reuse interpolaton instead of recomputing aggregates and updating the whole hierarchy. This is less expensive for multiple solves in which the matrix is not changing too quickly. */ ml_object = pc_ml->ml_object; gridctx = pc_ml->gridctx; Nlevels = pc_ml->Nlevels; fine_level = Nlevels - 1; gridctx[fine_level].A = A; ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr); ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr); if (isMPI){ ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr); } else if (isSeq) { Aloc = A; ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr); } else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name); ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr); PetscMLdata = pc_ml->PetscMLdata; ierr = MatDestroy(&PetscMLdata->Aloc);CHKERRQ(ierr); PetscMLdata->A = A; PetscMLdata->Aloc = Aloc; ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata); ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec); mesh_level = ml_object->ML_finest_level; while (ml_object->SingleLevel[mesh_level].Rmat->to) { old_mesh_level = mesh_level; mesh_level = ml_object->SingleLevel[mesh_level].Rmat->to->levelnum; /* clean and regenerate A */ mlmat = &(ml_object->Amat[mesh_level]); ML_Operator_Clean(mlmat); ML_Operator_Init(mlmat,ml_object->comm); ML_Gen_AmatrixRAP(ml_object, old_mesh_level, mesh_level); } level = fine_level - 1; if (size == 1) { /* convert ML P, R and A into seqaij format */ for (mllevel=1; mllevel<Nlevels; mllevel++){ mlmat = &(ml_object->Amat[mllevel]); ierr = MatWrapML_SeqAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr); level--; } } else { /* convert ML P and R into shell format, ML A into mpiaij format */ for (mllevel=1; mllevel<Nlevels; mllevel++){ mlmat = &(ml_object->Amat[mllevel]); ierr = MatWrapML_MPIAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr); level--; } } for (level=0; level<fine_level; level++) { if (level > 0){ ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr); } ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr); } ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr); ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr); ierr = PCSetUp_MG(pc);CHKERRQ(ierr); PetscFunctionReturn(0); } else { /* since ML can change the size of vectors/matrices at any level we must destroy everything */ ierr = PCReset_ML(pc);CHKERRQ(ierr); ierr = PCReset_MG(pc);CHKERRQ(ierr); } } /* setup special features of PCML */ /*--------------------------------*/ /* covert A to Aloc to be used by ML at fine grid */ pc_ml->size = size; ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr); ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr); if (isMPI){ ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr); } else if (isSeq) { Aloc = A; ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr); } else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name); /* create and initialize struct 'PetscMLdata' */ ierr = PetscNewLog(pc,FineGridCtx,&PetscMLdata);CHKERRQ(ierr); pc_ml->PetscMLdata = PetscMLdata; ierr = PetscMalloc((Aloc->cmap->n+1)*sizeof(PetscScalar),&PetscMLdata->pwork);CHKERRQ(ierr); ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->x);CHKERRQ(ierr); ierr = VecSetSizes(PetscMLdata->x,Aloc->cmap->n,Aloc->cmap->n);CHKERRQ(ierr); ierr = VecSetType(PetscMLdata->x,VECSEQ);CHKERRQ(ierr); ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->y);CHKERRQ(ierr); ierr = VecSetSizes(PetscMLdata->y,A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr); ierr = VecSetType(PetscMLdata->y,VECSEQ);CHKERRQ(ierr); PetscMLdata->A = A; PetscMLdata->Aloc = Aloc; /* create ML discretization matrix at fine grid */ /* ML requires input of fine-grid matrix. It determines nlevels. */ ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr); ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr); ML_Create(&ml_object,pc_ml->MaxNlevels); ML_Comm_Set_UsrComm(ml_object->comm,((PetscObject)A)->comm); pc_ml->ml_object = ml_object; ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata); ML_Set_Amatrix_Getrow(ml_object,0,PetscML_getrow,PetscML_comm,nlocal_allcols); ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec); ML_Set_Symmetrize(ml_object,pc_ml->Symmetrize ? ML_YES : ML_NO); /* aggregation */ ML_Aggregate_Create(&agg_object); pc_ml->agg_object = agg_object; { MatNullSpace mnull; ierr = MatGetNearNullSpace(A,&mnull);CHKERRQ(ierr); if (pc_ml->nulltype == PCML_NULLSPACE_AUTO) { if (mnull) pc_ml->nulltype = PCML_NULLSPACE_USER; else if (bs > 1) pc_ml->nulltype = PCML_NULLSPACE_BLOCK; else pc_ml->nulltype = PCML_NULLSPACE_SCALAR; } switch (pc_ml->nulltype) { case PCML_NULLSPACE_USER: { PetscScalar *nullvec; const PetscScalar *v; PetscBool has_const; PetscInt i,j,mlocal,nvec,M; const Vec *vecs; if (!mnull) SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_USER,"Must provide explicit null space using MatSetNearNullSpace() to use user-specified null space"); ierr = MatGetSize(A,&M,PETSC_NULL);CHKERRQ(ierr); ierr = MatGetLocalSize(Aloc,&mlocal,PETSC_NULL);CHKERRQ(ierr); ierr = MatNullSpaceGetVecs(mnull,&has_const,&nvec,&vecs);CHKERRQ(ierr); ierr = PetscMalloc((nvec+!!has_const)*mlocal*sizeof *nullvec,&nullvec);CHKERRQ(ierr); if (has_const) for (i=0; i<mlocal; i++) nullvec[i] = 1.0/M; for (i=0; i<nvec; i++) { ierr = VecGetArrayRead(vecs[i],&v);CHKERRQ(ierr); for (j=0; j<mlocal; j++) nullvec[(i+!!has_const)*mlocal + j] = v[j]; ierr = VecRestoreArrayRead(vecs[i],&v);CHKERRQ(ierr); } ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,nvec+!!has_const,nullvec,mlocal);CHKERRQ(ierr); ierr = PetscFree(nullvec);CHKERRQ(ierr); } break; case PCML_NULLSPACE_BLOCK: ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,bs,0,0);CHKERRQ(ierr); break; case PCML_NULLSPACE_SCALAR: break; default: SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_SUP,"Unknown null space type"); } } ML_Aggregate_Set_MaxCoarseSize(agg_object,pc_ml->MaxCoarseSize); /* set options */ switch (pc_ml->CoarsenScheme) { case 1: ML_Aggregate_Set_CoarsenScheme_Coupled(agg_object);break; case 2: ML_Aggregate_Set_CoarsenScheme_MIS(agg_object);break; case 3: ML_Aggregate_Set_CoarsenScheme_METIS(agg_object);break; } ML_Aggregate_Set_Threshold(agg_object,pc_ml->Threshold); ML_Aggregate_Set_DampingFactor(agg_object,pc_ml->DampingFactor); if (pc_ml->SpectralNormScheme_Anorm){ ML_Set_SpectralNormScheme_Anorm(ml_object); } agg_object->keep_agg_information = (int)pc_ml->KeepAggInfo; agg_object->keep_P_tentative = (int)pc_ml->Reusable; agg_object->block_scaled_SA = (int)pc_ml->BlockScaling; agg_object->minimizing_energy = (int)pc_ml->EnergyMinimization; agg_object->minimizing_energy_droptol = (double)pc_ml->EnergyMinimizationDropTol; agg_object->cheap_minimizing_energy = (int)pc_ml->EnergyMinimizationCheap; if (pc_ml->OldHierarchy) { Nlevels = ML_Gen_MGHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object); } else { Nlevels = ML_Gen_MultiLevelHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object); } if (Nlevels<=0) SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_OUTOFRANGE,"Nlevels %d must > 0",Nlevels); pc_ml->Nlevels = Nlevels; fine_level = Nlevels - 1; ierr = PCMGSetLevels(pc,Nlevels,PETSC_NULL);CHKERRQ(ierr); /* set default smoothers */ for (level=1; level<=fine_level; level++){ if (size == 1){ ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr); ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr); ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr); ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr); } else { ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr); ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr); ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr); ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr); } } ierr = PetscObjectOptionsBegin((PetscObject)pc);CHKERRQ(ierr); ierr = PCSetFromOptions_MG(pc);CHKERRQ(ierr); /* should be called in PCSetFromOptions_ML(), but cannot be called prior to PCMGSetLevels() */ ierr = PetscOptionsEnd();CHKERRQ(ierr); ierr = PetscMalloc(Nlevels*sizeof(GridCtx),&gridctx);CHKERRQ(ierr); pc_ml->gridctx = gridctx; /* wrap ML matrices by PETSc shell matrices at coarsened grids. Level 0 is the finest grid for ML, but coarsest for PETSc! */ gridctx[fine_level].A = A; level = fine_level - 1; if (size == 1){ /* convert ML P, R and A into seqaij format */ for (mllevel=1; mllevel<Nlevels; mllevel++){ mlmat = &(ml_object->Pmat[mllevel]); ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr); mlmat = &(ml_object->Rmat[mllevel-1]); ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr); mlmat = &(ml_object->Amat[mllevel]); ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr); level--; } } else { /* convert ML P and R into shell format, ML A into mpiaij format */ for (mllevel=1; mllevel<Nlevels; mllevel++){ mlmat = &(ml_object->Pmat[mllevel]); ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr); mlmat = &(ml_object->Rmat[mllevel-1]); ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr); mlmat = &(ml_object->Amat[mllevel]); ierr = MatWrapML_MPIAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr); level--; } } /* create vectors and ksp at all levels */ for (level=0; level<fine_level; level++){ level1 = level + 1; ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].x);CHKERRQ(ierr); ierr = VecSetSizes(gridctx[level].x,gridctx[level].A->cmap->n,PETSC_DECIDE);CHKERRQ(ierr); ierr = VecSetType(gridctx[level].x,VECMPI);CHKERRQ(ierr); ierr = PCMGSetX(pc,level,gridctx[level].x);CHKERRQ(ierr); ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].b);CHKERRQ(ierr); ierr = VecSetSizes(gridctx[level].b,gridctx[level].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr); ierr = VecSetType(gridctx[level].b,VECMPI);CHKERRQ(ierr); ierr = PCMGSetRhs(pc,level,gridctx[level].b);CHKERRQ(ierr); ierr = VecCreate(((PetscObject)gridctx[level1].A)->comm,&gridctx[level1].r);CHKERRQ(ierr); ierr = VecSetSizes(gridctx[level1].r,gridctx[level1].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr); ierr = VecSetType(gridctx[level1].r,VECMPI);CHKERRQ(ierr); ierr = PCMGSetR(pc,level1,gridctx[level1].r);CHKERRQ(ierr); if (level == 0){ ierr = PCMGGetCoarseSolve(pc,&gridctx[level].ksp);CHKERRQ(ierr); } else { ierr = PCMGGetSmoother(pc,level,&gridctx[level].ksp);CHKERRQ(ierr); } } ierr = PCMGGetSmoother(pc,fine_level,&gridctx[fine_level].ksp);CHKERRQ(ierr); /* create coarse level and the interpolation between the levels */ for (level=0; level<fine_level; level++){ level1 = level + 1; ierr = PCMGSetInterpolation(pc,level1,gridctx[level].P);CHKERRQ(ierr); ierr = PCMGSetRestriction(pc,level1,gridctx[level].R);CHKERRQ(ierr); if (level > 0){ ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr); } ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); } ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr); ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); /* setupcalled is set to 0 so that MG is setup from scratch */ pc->setupcalled = 0; ierr = PCSetUp_MG(pc);CHKERRQ(ierr); PetscFunctionReturn(0); }
int main(int argc, char *argv[]){ ML *ml_object; int i, N_grids = 3, N_levels; double sol[5], rhs[5]; ML_Aggregate *agg_object; int proc, nlocal, nlocal_allcolumns; MPI_Init(&argc,&argv); ML_Set_PrintLevel(15); for (i = 0; i < 5; i++) sol[i] = 0.; for (i = 0; i < 5; i++) rhs[i] = 2.; ML_Create (&ml_object, N_grids); proc = ml_object->comm->ML_mypid; if (ml_object->comm->ML_nprocs != 2) { if (proc == 0) printf("Must be run on two processors\n"); ML_Destroy(&ml_object); MPI_Finalize(); exit(1); } if (proc == 0) {nlocal = 2; nlocal_allcolumns = 4;} else if (proc == 1){nlocal = 3; nlocal_allcolumns = 5;} else {nlocal = 0; nlocal_allcolumns = 0;} ML_Init_Amatrix (ml_object, 0, nlocal, nlocal, &proc); ML_Set_Amatrix_Getrow(ml_object, 0, Poisson_getrow, Poisson_comm, nlocal_allcolumns); ML_Set_Amatrix_Matvec(ml_object, 0, Poisson_matvec); ML_Aggregate_Create(&agg_object); ML_Aggregate_Set_MaxCoarseSize(agg_object,1); N_levels = ML_Gen_MGHierarchy_UsingAggregation(ml_object, 0, ML_INCREASING, agg_object); ML_Gen_Smoother_Jacobi(ml_object, ML_ALL_LEVELS, ML_PRESMOOTHER, 1, ML_DEFAULT); ML_Gen_Solver (ml_object, ML_MGV, 0, N_levels-1); ML_Iterate(ml_object, sol, rhs); if (proc == 0) { printf("sol(0) = %e\n",sol[1]); fflush(stdout); } ML_Comm_GsumInt(ml_object->comm,1); /* just used for synchronization */ if (proc == 1) { printf("sol(1) = %e\n",sol[0]); printf("sol(2) = %e\n",sol[1]); printf("sol(3) = %e\n",sol[2]); fflush(stdout); } ML_Comm_GsumInt(ml_object->comm,1); /* just used for synchronization */ if (proc == 0) { printf("sol(4) = %e\n",sol[0]); fflush(stdout); } ML_Aggregate_Destroy(&agg_object); ML_Destroy(&ml_object); MPI_Finalize(); return 0; }
HYPRE_Int HYPRE_ParCSRMLSetup( HYPRE_Solver solver, HYPRE_ParCSRMatrix A, HYPRE_ParVector b, HYPRE_ParVector x ) { HYPRE_Int i, my_id, nprocs, coarsest_level, level, sweeps, nlevels; HYPRE_Int *row_partition, localEqns, length; HYPRE_Int Nblocks, *blockList; double wght; MH_Context *context; MH_Matrix *mh_mat; /* -------------------------------------------------------- */ /* fetch the ML pointer */ /* -------------------------------------------------------- */ MH_Link *link = (MH_Link *) solver; ML *ml = link->ml_ptr; nlevels = link->nlevels; /* -------------------------------------------------------- */ /* set up the parallel environment */ /* -------------------------------------------------------- */ hypre_MPI_Comm_rank(link->comm, &my_id); hypre_MPI_Comm_size(link->comm, &nprocs); /* -------------------------------------------------------- */ /* fetch the matrix row partition information and put it */ /* into the matrix data object (for matvec and getrow) */ /* -------------------------------------------------------- */ HYPRE_ParCSRMatrixGetRowPartitioning( A, &row_partition ); localEqns = row_partition[my_id+1] - row_partition[my_id]; context = (MH_Context *) malloc(sizeof(MH_Context)); link->contxt = context; context->comm = link->comm; context->globalEqns = row_partition[nprocs]; context->partition = (HYPRE_Int *) malloc(sizeof(HYPRE_Int)*(nprocs+1)); for (i=0; i<=nprocs; i++) context->partition[i] = row_partition[i]; hypre_TFree( row_partition ); mh_mat = ( MH_Matrix * ) malloc( sizeof( MH_Matrix) ); context->Amat = mh_mat; HYPRE_ParCSRMLConstructMHMatrix(A,mh_mat,link->comm, context->partition,context); /* -------------------------------------------------------- */ /* set up the ML communicator information */ /* -------------------------------------------------------- */ ML_Set_Comm_Communicator(ml, link->comm); ML_Set_Comm_MyRank(ml, my_id); ML_Set_Comm_Nprocs(ml, nprocs); ML_Set_Comm_Send(ml, MH_Send); ML_Set_Comm_Recv(ml, MH_Irecv); ML_Set_Comm_Wait(ml, MH_Wait); /* -------------------------------------------------------- */ /* set up the ML matrix information */ /* -------------------------------------------------------- */ ML_Init_Amatrix(ml, nlevels-1, localEqns, localEqns, (void *) context); ML_Set_Amatrix_Matvec(ml, nlevels-1, MH_MatVec); length = localEqns; for (i=0; i<mh_mat->recvProcCnt; i++ ) length += mh_mat->recvLeng[i]; ML_Set_Amatrix_Getrow(ml, nlevels-1, MH_GetRow, MH_ExchBdry, length); /* -------------------------------------------------------- */ /* create an aggregate context */ /* -------------------------------------------------------- */ ML_Aggregate_Create(&(link->ml_ag)); link->ml_ag->max_levels = link->nlevels; ML_Aggregate_Set_Threshold( link->ml_ag, link->ag_threshold ); /* -------------------------------------------------------- */ /* perform aggregation */ /* -------------------------------------------------------- */ coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, nlevels-1, ML_DECREASING, link->ml_ag); if ( my_id == 0 ) hypre_printf("ML : number of levels = %d\n", coarsest_level); coarsest_level = nlevels - coarsest_level; /* -------------------------------------------------------- */ /* set up smoother and coarse solver */ /* -------------------------------------------------------- */ for (level = nlevels-1; level > coarsest_level; level--) { sweeps = link->pre_sweeps; wght = link->jacobi_wt; switch ( link->pre ) { case 0 : ML_Gen_SmootherJacobi(ml, level, ML_PRESMOOTHER, sweeps, wght); break; case 1 : ML_Gen_SmootherGaussSeidel(ml, level, ML_PRESMOOTHER, sweeps); break; case 2 : ML_Gen_SmootherSymGaussSeidel(ml,level,ML_PRESMOOTHER,sweeps,1.0); break; case 3 : Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level ); ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList ); ML_Gen_SmootherVBlockGaussSeidel(ml,level,ML_PRESMOOTHER, sweeps, Nblocks, blockList); break; case 4 : Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level ); ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList ); ML_Gen_SmootherVBlockJacobi(ml,level,ML_PRESMOOTHER, sweeps, wght, Nblocks, blockList); break; } sweeps = link->post_sweeps; switch ( link->post ) { case 0 : ML_Gen_SmootherJacobi(ml, level, ML_POSTSMOOTHER, sweeps, wght); break; case 1 : ML_Gen_SmootherGaussSeidel(ml, level, ML_POSTSMOOTHER, sweeps); break; case 2 : ML_Gen_SmootherSymGaussSeidel(ml,level,ML_POSTSMOOTHER,sweeps,1.0); break; case 3 : Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level ); ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList ); ML_Gen_SmootherVBlockGaussSeidel(ml,level,ML_POSTSMOOTHER, sweeps, Nblocks, blockList); break; case 4 : Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level ); ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList ); ML_Gen_SmootherVBlockJacobi(ml,level,ML_POSTSMOOTHER, sweeps, wght, Nblocks, blockList); break; } } ML_Gen_CoarseSolverSuperLU(ml, coarsest_level); //ML_Gen_SmootherGaussSeidel(ml, coarsest_level, ML_PRESMOOTHER, 100); ML_Gen_Solver(ml, ML_MGV, nlevels-1, coarsest_level); return 0; }
int main(int argc, char *argv[]) { int num_PDE_eqns=1, N_levels=3, nsmooth=2; int leng, level, N_grid_pts, coarsest_level; int leng1,leng2; /* See Aztec User's Guide for more information on the */ /* variables that follow. */ int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE]; double params[AZ_PARAMS_SIZE], status[AZ_STATUS_SIZE]; /* data structure for matrix corresponding to the fine grid */ double *val = NULL, *xxx, *rhs, solve_time, setup_time, start_time; AZ_MATRIX *Amat; AZ_PRECOND *Pmat = NULL; ML *ml; FILE *fp; int i, j, Nrigid, *garbage, nblocks=0, *blocks = NULL, *block_pde=NULL; struct AZ_SCALING *scaling; ML_Aggregate *ag; double *mode, *rigid=NULL, alpha; char filename[80]; int one = 1; int proc,nprocs; char pathfilename[100]; #ifdef ML_MPI MPI_Init(&argc,&argv); /* get number of processors and the name of this processor */ AZ_set_proc_config(proc_config, MPI_COMM_WORLD); proc = proc_config[AZ_node]; nprocs = proc_config[AZ_N_procs]; #else AZ_set_proc_config(proc_config, AZ_NOT_MPI); proc = 0; nprocs = 1; #endif if (proc_config[AZ_node] == 0) { sprintf(pathfilename,"%s/inputfile",argv[1]); ML_Reader_ReadInput(pathfilename, &context); } else context = (struct reader_context *) ML_allocate(sizeof(struct reader_context)); AZ_broadcast((char *) context, sizeof(struct reader_context), proc_config, AZ_PACK); AZ_broadcast((char *) NULL , 0 , proc_config, AZ_SEND); N_levels = context->N_levels; printf("N_levels %d\n",N_levels); nsmooth = context->nsmooth; num_PDE_eqns = context->N_dofPerNode; printf("num_PDE_eqns %d\n",num_PDE_eqns); ML_Set_PrintLevel(context->output_level); /* read in the number of matrix equations */ leng = 0; if (proc_config[AZ_node] == 0) { sprintf(pathfilename,"%s/data_matrix.txt",argv[1]); fp=fopen(pathfilename,"r"); if (fp==NULL) { printf("**ERR** couldn't open file data_matrix.txt\n"); exit(1); } fscanf(fp,"%d",&leng); fclose(fp); } leng = AZ_gsum_int(leng, proc_config); N_grid_pts=leng/num_PDE_eqns; /* initialize the list of global indices. NOTE: the list of global */ /* indices must be in ascending order so that subsequent calls to */ /* AZ_find_index() will function properly. */ #if 0 if (proc_config[AZ_N_procs] == 1) i = AZ_linear; else i = AZ_file; #endif i = AZ_linear; /* cannot use AZ_input_update for variable blocks (forgot why, but debugged through it)*/ /* make a linear distribution of the matrix */ /* if the linear distribution does not align with the blocks, */ /* this is corrected in ML_AZ_Reader_ReadVariableBlocks */ leng1 = leng/nprocs; leng2 = leng-leng1*nprocs; if (proc >= leng2) { leng2 += (proc*leng1); } else { leng1++; leng2 = proc*leng1; } N_update = leng1; update = (int*)AZ_allocate((N_update+1)*sizeof(int)); if (update==NULL) { (void) fprintf (stderr, "Not enough space to allocate 'update'\n"); fflush(stderr); exit(EXIT_FAILURE); } for (i=0; i<N_update; i++) update[i] = i+leng2; #if 0 /* debug */ printf("proc %d N_update %d\n",proc_config[AZ_node],N_update); fflush(stdout); #endif sprintf(pathfilename,"%s/data_vblocks.txt",argv[1]); ML_AZ_Reader_ReadVariableBlocks(pathfilename,&nblocks,&blocks,&block_pde, &N_update,&update,proc_config); #if 0 /* debug */ printf("proc %d N_update %d\n",proc_config[AZ_node],N_update); fflush(stdout); #endif sprintf(pathfilename,"%s/data_matrix.txt",argv[1]); AZ_input_msr_matrix(pathfilename,update, &val, &bindx, N_update, proc_config); /* This code is to fix things up so that we are sure we have */ /* all blocks (including the ghost nodes) the same size. */ /* not sure, whether this is a good idea with variable blocks */ /* the examples inpufiles (see top of this file) don't need it */ /* anyway */ /* AZ_block_MSR(&bindx, &val, N_update, num_PDE_eqns, update); */ AZ_transform_norowreordering(proc_config, &external, bindx, val, update, &update_index, &extern_index, &data_org, N_update, 0, 0, 0, &cpntr, AZ_MSR_MATRIX); Amat = AZ_matrix_create( leng ); AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL); Amat->matrix_type = data_org[AZ_matrix_type]; data_org[AZ_N_rows] = data_org[AZ_N_internal] + data_org[AZ_N_border]; start_time = AZ_second(); options[AZ_scaling] = AZ_none; ML_Create(&ml, N_levels); /* set up discretization matrix and matrix vector function */ AZ_ML_Set_Amat(ml, 0, N_update, N_update, Amat, proc_config); ML_Set_ResidualOutputFrequency(ml, context->output); ML_Set_Tolerance(ml, context->tol); ML_Aggregate_Create( &ag ); if (ML_strcmp(context->agg_coarsen_scheme,"Mis") == 0) { ML_Aggregate_Set_CoarsenScheme_MIS(ag); } else if (ML_strcmp(context->agg_coarsen_scheme,"Uncoupled") == 0) { ML_Aggregate_Set_CoarsenScheme_Uncoupled(ag); } else if (ML_strcmp(context->agg_coarsen_scheme,"Coupled") == 0) { ML_Aggregate_Set_CoarsenScheme_Coupled(ag); } else if (ML_strcmp(context->agg_coarsen_scheme,"Metis") == 0) { ML_Aggregate_Set_CoarsenScheme_METIS(ag); for (i=0; i<N_levels; i++) ML_Aggregate_Set_NodesPerAggr(ml,ag,i,9); } else if (ML_strcmp(context->agg_coarsen_scheme,"VBMetis") == 0) { /* when no blocks read, use standard metis assuming constant block sizes */ if (!blocks) ML_Aggregate_Set_CoarsenScheme_METIS(ag); else { ML_Aggregate_Set_CoarsenScheme_VBMETIS(ag); ML_Aggregate_Set_Vblocks_CoarsenScheme_VBMETIS(ag,0,N_levels,nblocks, blocks,block_pde,N_update); } for (i=0; i<N_levels; i++) ML_Aggregate_Set_NodesPerAggr(ml,ag,i,9); } else { printf("**ERR** ML: Unknown aggregation scheme %s\n",context->agg_coarsen_scheme); exit(-1); } ML_Aggregate_Set_DampingFactor(ag, context->agg_damping); ML_Aggregate_Set_MaxCoarseSize( ag, context->maxcoarsesize); ML_Aggregate_Set_Threshold(ag, context->agg_thresh); if (ML_strcmp(context->agg_spectral_norm,"Calc") == 0) { ML_Set_SpectralNormScheme_Calc(ml); } else if (ML_strcmp(context->agg_spectral_norm,"Anorm") == 0) { ML_Set_SpectralNormScheme_Anorm(ml); } else { printf("**WRN** ML: Unknown spectral norm scheme %s\n",context->agg_spectral_norm); } /* read in the rigid body modes */ Nrigid = 0; if (proc_config[AZ_node] == 0) { sprintf(filename,"data_nullsp%d.txt",Nrigid); sprintf(pathfilename,"%s/%s",argv[1],filename); while( (fp = fopen(pathfilename,"r")) != NULL) { fclose(fp); Nrigid++; sprintf(filename,"data_nullsp%d.txt",Nrigid); sprintf(pathfilename,"%s/%s",argv[1],filename); } } Nrigid = AZ_gsum_int(Nrigid,proc_config); if (Nrigid != 0) { rigid = (double *) ML_allocate( sizeof(double)*Nrigid*(N_update+1) ); if (rigid == NULL) { printf("Error: Not enough space for rigid body modes\n"); } } /* Set rhs */ sprintf(pathfilename,"%s/data_rhs.txt",argv[1]); fp = fopen(pathfilename,"r"); if (fp == NULL) { rhs=(double *)ML_allocate(leng*sizeof(double)); if (proc_config[AZ_node] == 0) printf("taking linear vector for rhs\n"); for (i = 0; i < N_update; i++) rhs[i] = (double) update[i]; } else { fclose(fp); if (proc_config[AZ_node] == 0) printf("reading rhs from a file\n"); AZ_input_msr_matrix(pathfilename, update, &rhs, &garbage, N_update, proc_config); } AZ_reorder_vec(rhs, data_org, update_index, NULL); for (i = 0; i < Nrigid; i++) { sprintf(filename,"data_nullsp%d.txt",i); sprintf(pathfilename,"%s/%s",argv[1],filename); AZ_input_msr_matrix(pathfilename, update, &mode, &garbage, N_update, proc_config); AZ_reorder_vec(mode, data_org, update_index, NULL); #if 0 /* test the given rigid body mode, output-vector should be ~0 */ Amat->matvec(mode, rigid, Amat, proc_config); for (j = 0; j < N_update; j++) printf("this is %d %e\n",j,rigid[j]); #endif for (j = 0; j < i; j++) { alpha = -AZ_gdot(N_update, mode, &(rigid[j*N_update]), proc_config)/ AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); DAXPY_F77(&N_update, &alpha, &(rigid[j*N_update]), &one, mode, &one); } /* rhs orthogonalization */ alpha = -AZ_gdot(N_update, mode, rhs, proc_config)/ AZ_gdot(N_update, mode, mode, proc_config); DAXPY_F77(&N_update, &alpha, mode, &one, rhs, &one); for (j = 0; j < N_update; j++) rigid[i*N_update+j] = mode[j]; free(mode); free(garbage); } for (j = 0; j < Nrigid; j++) { alpha = -AZ_gdot(N_update, rhs, &(rigid[j*N_update]), proc_config)/ AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); DAXPY_F77(&N_update, &alpha, &(rigid[j*N_update]), &one, rhs, &one); } #if 0 /* for testing the default nullsp */ ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, 6, NULL, N_update); #else if (Nrigid != 0) { ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update); } #endif if (rigid) ML_free(rigid); ag->keep_agg_information = 1; coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, 0, ML_INCREASING, ag); coarsest_level--; if ( proc_config[AZ_node] == 0 ) printf("Coarse level = %d \n", coarsest_level); #if 0 /* set up smoothers */ if (!blocks) blocks = (int *) ML_allocate(sizeof(int)*N_update); #endif for (level = 0; level < coarsest_level; level++) { num_PDE_eqns = ml->Amat[level].num_PDEs; /* Sparse approximate inverse smoother that acutally does both */ /* pre and post smoothing. */ if (ML_strcmp(context->smoother,"Parasails") == 0) { ML_Gen_Smoother_ParaSails(ml , level, ML_PRESMOOTHER, nsmooth, parasails_sym, parasails_thresh, parasails_nlevels, parasails_filter, (int) parasails_loadbal, parasails_factorized); } /* This is the symmetric Gauss-Seidel smoothing that we usually use. */ /* In parallel, it is not a true Gauss-Seidel in that each processor */ /* does a Gauss-Seidel on its local submatrix independent of the */ /* other processors. */ else if (ML_strcmp(context->smoother,"GaussSeidel") == 0) { ML_Gen_Smoother_GaussSeidel(ml , level, ML_BOTH, nsmooth,1.); } else if (ML_strcmp(context->smoother,"SymGaussSeidel") == 0) { ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.); } else if (ML_strcmp(context->smoother,"Poly") == 0) { ML_Gen_Smoother_Cheby(ml, level, ML_BOTH, 30., nsmooth); } else if (ML_strcmp(context->smoother,"BlockGaussSeidel") == 0) { ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_BOTH, nsmooth,1., num_PDE_eqns); } else if (ML_strcmp(context->smoother,"VBSymGaussSeidel") == 0) { if (blocks) ML_free(blocks); if (block_pde) ML_free(block_pde); blocks = NULL; block_pde = NULL; nblocks = 0; ML_Aggregate_Get_Vblocks_CoarsenScheme_VBMETIS(ag,level,N_levels,&nblocks, &blocks,&block_pde); if (blocks==NULL) ML_Gen_Blocks_Aggregates(ag, level, &nblocks, &blocks); ML_Gen_Smoother_VBlockSymGaussSeidel(ml , level, ML_BOTH, nsmooth,1., nblocks, blocks); } /* This is a true Gauss Seidel in parallel. This seems to work for */ /* elasticity problems. However, I don't believe that this is very */ /* efficient in parallel. */ /* nblocks = ml->Amat[level].invec_leng; for (i =0; i < nblocks; i++) blocks[i] = i; ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml , level, ML_PRESMOOTHER, nsmooth, 1., nblocks, blocks); ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml, level, ML_POSTSMOOTHER, nsmooth, 1., nblocks, blocks); */ /* Jacobi Smoothing */ else if (ML_strcmp(context->smoother,"Jacobi") == 0) { ML_Gen_Smoother_Jacobi(ml , level, ML_PRESMOOTHER, nsmooth,.4); ML_Gen_Smoother_Jacobi(ml , level, ML_POSTSMOOTHER, nsmooth,.4); } /* This does a block Gauss-Seidel (not true GS in parallel) */ /* where each processor has 'nblocks' blocks. */ /* */ else if (ML_strcmp(context->smoother,"Metis") == 0) { if (blocks) ML_free(blocks); if (block_pde) ML_free(block_pde); nblocks = 250; ML_Gen_Blocks_Metis(ml, level, &nblocks, &blocks); ML_Gen_Smoother_VBlockSymGaussSeidel(ml , level, ML_BOTH, nsmooth,1., nblocks, blocks); } else { printf("unknown smoother %s\n",context->smoother); exit(1); } } /* set coarse level solver */ nsmooth = context->coarse_its; /* Sparse approximate inverse smoother that acutally does both */ /* pre and post smoothing. */ if (ML_strcmp(context->coarse_solve,"Parasails") == 0) { ML_Gen_Smoother_ParaSails(ml , coarsest_level, ML_PRESMOOTHER, nsmooth, parasails_sym, parasails_thresh, parasails_nlevels, parasails_filter, (int) parasails_loadbal, parasails_factorized); } else if (ML_strcmp(context->coarse_solve,"GaussSeidel") == 0) { ML_Gen_Smoother_GaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.); } else if (ML_strcmp(context->coarse_solve,"Poly") == 0) { ML_Gen_Smoother_Cheby(ml, coarsest_level, ML_BOTH, 30., nsmooth); } else if (ML_strcmp(context->coarse_solve,"SymGaussSeidel") == 0) { ML_Gen_Smoother_SymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.); } else if (ML_strcmp(context->coarse_solve,"BlockGaussSeidel") == 0) { ML_Gen_Smoother_BlockGaussSeidel(ml, coarsest_level, ML_BOTH, nsmooth,1., num_PDE_eqns); } else if (ML_strcmp(context->coarse_solve,"Aggregate") == 0) { if (blocks) ML_free(blocks); if (block_pde) ML_free(block_pde); ML_Gen_Blocks_Aggregates(ag, coarsest_level, &nblocks, &blocks); ML_Gen_Smoother_VBlockSymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1., nblocks, blocks); } else if (ML_strcmp(context->coarse_solve,"Jacobi") == 0) { ML_Gen_Smoother_Jacobi(ml , coarsest_level, ML_BOTH, nsmooth,.5); } else if (ML_strcmp(context->coarse_solve,"Metis") == 0) { if (blocks) ML_free(blocks); if (block_pde) ML_free(block_pde); nblocks = 250; ML_Gen_Blocks_Metis(ml, coarsest_level, &nblocks, &blocks); ML_Gen_Smoother_VBlockSymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1., nblocks, blocks); } else if (ML_strcmp(context->coarse_solve,"SuperLU") == 0) { ML_Gen_CoarseSolverSuperLU( ml, coarsest_level); } else if (ML_strcmp(context->coarse_solve,"Amesos") == 0) { ML_Gen_Smoother_Amesos(ml,coarsest_level,ML_AMESOS_KLU,-1, 0.0); } else { printf("unknown coarse grid solver %s\n",context->coarse_solve); exit(1); } ML_Gen_Solver(ml, ML_MGV, 0, coarsest_level); AZ_defaults(options, params); if (ML_strcmp(context->krylov,"Cg") == 0) { options[AZ_solver] = AZ_cg; } else if (ML_strcmp(context->krylov,"Bicgstab") == 0) { options[AZ_solver] = AZ_bicgstab; } else if (ML_strcmp(context->krylov,"Tfqmr") == 0) { options[AZ_solver] = AZ_tfqmr; } else if (ML_strcmp(context->krylov,"Gmres") == 0) { options[AZ_solver] = AZ_gmres; } else { printf("unknown krylov method %s\n",context->krylov); } if (blocks) ML_free(blocks); if (block_pde) ML_free(block_pde); options[AZ_scaling] = AZ_none; options[AZ_precond] = AZ_user_precond; options[AZ_conv] = AZ_r0; options[AZ_output] = 1; options[AZ_max_iter] = context->max_outer_its; options[AZ_poly_ord] = 5; options[AZ_kspace] = 130; params[AZ_tol] = context->tol; options[AZ_output] = context->output; ML_free(context); AZ_set_ML_preconditioner(&Pmat, Amat, ml, options); setup_time = AZ_second() - start_time; xxx = (double *) malloc( leng*sizeof(double)); for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0; /* Set x */ /* there is no initguess supplied with these examples for the moment.... */ fp = fopen("initguessfile","r"); if (fp != NULL) { fclose(fp); if (proc_config[AZ_node]== 0) printf("reading initial guess from file\n"); AZ_input_msr_matrix("data_initguess.txt", update, &xxx, &garbage, N_update, proc_config); options[AZ_conv] = AZ_expected_values; } else if (proc_config[AZ_node]== 0) printf("taking 0 initial guess \n"); AZ_reorder_vec(xxx, data_org, update_index, NULL); /* if Dirichlet BC ... put the answer in */ for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) { if ( (val[i] > .99999999) && (val[i] < 1.0000001)) xxx[i] = rhs[i]; } fp = fopen("AZ_no_multilevel.dat","r"); scaling = AZ_scaling_create(); start_time = AZ_second(); if (fp != NULL) { fclose(fp); options[AZ_precond] = AZ_none; options[AZ_scaling] = AZ_sym_diag; options[AZ_ignore_scaling] = AZ_TRUE; options[AZ_keep_info] = 1; AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); /* options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); */ } else { options[AZ_keep_info] = 1; AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; /* if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); */ } solve_time = AZ_second() - start_time; if (proc_config[AZ_node] == 0) printf("Solve time = %e, MG Setup time = %e\n", solve_time, setup_time); if (proc_config[AZ_node] == 0) printf("Printing out a few entries of the solution ...\n"); for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 7) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 23) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 47) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 101) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 171) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} ML_Aggregate_Destroy(&ag); ML_Destroy(&ml); AZ_free((void *) Amat->data_org); AZ_free((void *) Amat->val); AZ_free((void *) Amat->bindx); AZ_free((void *) update); AZ_free((void *) external); AZ_free((void *) extern_index); AZ_free((void *) update_index); AZ_scaling_destroy(&scaling); if (Amat != NULL) AZ_matrix_destroy(&Amat); if (Pmat != NULL) AZ_precond_destroy(&Pmat); free(xxx); free(rhs); #ifdef ML_MPI MPI_Finalize(); #endif return 0; }
int main(int argc, char *argv[]) { int num_PDE_eqns=3, N_levels=3, nsmooth=1; int leng, level, N_grid_pts, coarsest_level; /* See Aztec User's Guide for more information on the */ /* variables that follow. */ int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE]; double params[AZ_PARAMS_SIZE], status[AZ_STATUS_SIZE]; /* data structure for matrix corresponding to the fine grid */ int *data_org = NULL, *update = NULL, *external = NULL; int *update_index = NULL, *extern_index = NULL; int *cpntr = NULL; int *bindx = NULL, N_update, iii; double *val = NULL; double *xxx, *rhs; AZ_MATRIX *Amat; AZ_PRECOND *Pmat = NULL; ML *ml; FILE *fp; int ch,i,j, Nrigid, *garbage; struct AZ_SCALING *scaling; double solve_time, setup_time, start_time, *mode, *rigid; ML_Aggregate *ag; int nblocks, *blocks; char filename[80]; double alpha; int one = 1; #ifdef ML_MPI MPI_Init(&argc,&argv); /* get number of processors and the name of this processor */ AZ_set_proc_config(proc_config, MPI_COMM_WORLD); #else AZ_set_proc_config(proc_config, AZ_NOT_MPI); #endif leng = 0; if (proc_config[AZ_node] == 0) { #ifdef binary fp=fopen(".data","rb"); #else fp=fopen(".data","r"); #endif if (fp==NULL) { printf("couldn't open file .data\n"); exit(1); } #ifdef binary fread(&leng, sizeof(int), 1, fp); #else fscanf(fp,"%d",&leng); #endif fclose(fp); } leng = AZ_gsum_int(leng, proc_config); N_grid_pts=leng/num_PDE_eqns; /* initialize the list of global indices. NOTE: the list of global */ /* indices must be in ascending order so that subsequent calls to */ /* AZ_find_index() will function properly. */ AZ_read_update(&N_update, &update, proc_config, N_grid_pts, num_PDE_eqns, AZ_linear); AZ_read_msr_matrix(update, &val, &bindx, N_update, proc_config); AZ_transform(proc_config, &external, bindx, val, update, &update_index, &extern_index, &data_org, N_update, 0, 0, 0, &cpntr, AZ_MSR_MATRIX); Amat = AZ_matrix_create( leng ); AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL); Amat->matrix_type = data_org[AZ_matrix_type]; data_org[AZ_N_rows] = data_org[AZ_N_internal] + data_org[AZ_N_border]; start_time = AZ_second(); AZ_defaults(options, params); /* scaling = AZ_scaling_create(); xxx = (double *) calloc( leng,sizeof(double)); rhs=(double *)calloc(leng,sizeof(double)); options[AZ_scaling] = AZ_sym_diag; options[AZ_precond] = AZ_none; options[AZ_max_iter] = 30; options[AZ_keep_info] = 1; AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); don't forget vector rescaling ... free(xxx); free(rhs); */ options[AZ_scaling] = AZ_none; ML_Create(&ml, N_levels); /* set up discretization matrix and matrix vector function */ AZ_ML_Set_Amat(ml, N_levels-1, N_update, N_update, Amat, proc_config); ML_Aggregate_Create( &ag ); Nrigid = 0; if (proc_config[AZ_node] == 0) { sprintf(filename,"rigid_body_mode%d",Nrigid+1); while( (fp = fopen(filename,"r")) != NULL) { fclose(fp); Nrigid++; sprintf(filename,"rigid_body_mode%d",Nrigid+1); } } Nrigid = AZ_gsum_int(Nrigid,proc_config); if (Nrigid != 0) { rigid = (double *) ML_allocate( sizeof(double)*Nrigid*(N_update+1) ); if (rigid == NULL) { printf("Error: Not enough space for rigid body modes\n"); } } rhs=(double *)malloc(leng*sizeof(double)); AZ_random_vector(rhs, data_org, proc_config); for (i = 0; i < Nrigid; i++) { sprintf(filename,"rigid_body_mode%d",i+1); AZ_input_msr_matrix(filename, update, &mode, &garbage, N_update, proc_config); /* AZ_sym_rescale_sl(mode, Amat->data_org, options, proc_config, scaling); */ /* Amat->matvec(mode, rigid, Amat, proc_config); for (j = 0; j < N_update; j++) printf("this is %d %e\n",j,rigid[j]); */ for (j = 0; j < i; j++) { alpha = -AZ_gdot(N_update, mode, &(rigid[j*N_update]), proc_config)/AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); daxpy_(&N_update, &alpha, &(rigid[j*N_update]), &one, mode, &one); printf("alpha1 is %e\n",alpha); } alpha = -AZ_gdot(N_update, mode, rhs, proc_config)/AZ_gdot(N_update, mode, mode, proc_config); printf("alpha2 is %e\n",alpha); daxpy_(&N_update, &alpha, mode, &one, rhs, &one); for (j = 0; j < N_update; j++) rigid[i*N_update+j] = mode[j]; free(mode); free(garbage); } for (j = 0; j < Nrigid; j++) { alpha = -AZ_gdot(N_update, rhs, &(rigid[j*N_update]), proc_config)/AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); daxpy_(&N_update, &alpha, &(rigid[j*N_update]), &one, rhs, &one); printf("alpha4 is %e\n",alpha); } for (i = 0; i < Nrigid; i++) { alpha = -AZ_gdot(N_update, &(rigid[i*N_update]), rhs, proc_config); printf("alpha is %e\n",alpha); } if (Nrigid != 0) { ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update); /* free(rigid); */ } coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, N_levels-1, ML_DECREASING, ag); coarsest_level = N_levels - coarsest_level; /* ML_Operator_Print(&(ml->Pmat[N_levels-2]), "Pmat"); exit(1); */ if ( proc_config[AZ_node] == 0 ) printf("Coarse level = %d \n", coarsest_level); /* set up smoothers */ for (level = N_levels-1; level > coarsest_level; level--) { j = 10; if (level == N_levels-1) j = 10; options[AZ_solver] = AZ_cg; options[AZ_precond]=AZ_sym_GS; options[AZ_subdomain_solve]=AZ_icc; /* options[AZ_precond] = AZ_none; */ options[AZ_poly_ord] = 5; ML_Gen_SmootherAztec(ml, level, options, params, proc_config, status, j, ML_PRESMOOTHER,NULL); ML_Gen_SmootherAztec(ml, level, options, params, proc_config, status, j, ML_POSTSMOOTHER,NULL); /* ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth,1.0); ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth,1.0); */ /* nblocks = ML_Aggregate_Get_AggrCount( ag, level ); ML_Aggregate_Get_AggrMap( ag, level, &blocks); ML_Gen_Smoother_VBlockSymGaussSeidel( ml , level, ML_BOTH, nsmooth, 1.0, nblocks, blocks); ML_Gen_Smoother_VBlockSymGaussSeidel( ml , level, ML_POSTSMOOTHER, nsmooth, 1.0, nblocks, blocks); */ /* ML_Gen_Smoother_VBlockJacobi( ml , level, ML_PRESMOOTHER, nsmooth, .5, nblocks, blocks); ML_Gen_Smoother_VBlockJacobi( ml , level, ML_POSTSMOOTHER, nsmooth,.5, nblocks, blocks); */ /* ML_Gen_Smoother_GaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth); ML_Gen_Smoother_GaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth); */ /* need to change this when num_pdes is different on different levels */ /* if (level == N_levels-1) { ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth, 0.5, num_PDE_eqns); ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth, 0.5, num_PDE_eqns); } else { ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth, 0.5, 2*num_PDE_eqns); ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth, 0.5, 2*num_PDE_eqns); } */ /* */ /* ML_Gen_SmootherJacobi(ml , level, ML_PRESMOOTHER, nsmooth, .67); ML_Gen_SmootherJacobi(ml , level, ML_POSTSMOOTHER, nsmooth, .67 ); */ } /* ML_Gen_CoarseSolverSuperLU( ml, coarsest_level); */ /* ML_Gen_SmootherSymGaussSeidel(ml , coarsest_level, ML_PRESMOOTHER, 2*nsmooth,1.); */ /* ML_Gen_SmootherBlockGaussSeidel(ml , level, ML_PRESMOOTHER, 50*nsmooth, 1.0, 2*num_PDE_eqns); */ ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, 2*nsmooth, 1.0, num_PDE_eqns); ML_Gen_Solver(ml, ML_MGV, N_levels-1, coarsest_level); AZ_defaults(options, params); options[AZ_solver] = AZ_GMRESR; options[AZ_scaling] = AZ_none; options[AZ_precond] = AZ_user_precond; options[AZ_conv] = AZ_rhs; options[AZ_output] = 1; options[AZ_max_iter] = 1500; options[AZ_poly_ord] = 5; options[AZ_kspace] = 130; params[AZ_tol] = 1.0e-8; AZ_set_ML_preconditioner(&Pmat, Amat, ml, options); setup_time = AZ_second() - start_time; xxx = (double *) malloc( leng*sizeof(double)); /* Set rhs */ fp = fopen("AZ_capture_rhs.dat","r"); if (fp == NULL) { if (proc_config[AZ_node] == 0) printf("taking random vector for rhs\n"); /* AZ_random_vector(rhs, data_org, proc_config); AZ_reorder_vec(rhs, data_org, update_index, NULL); AZ_random_vector(xxx, data_org, proc_config); AZ_reorder_vec(xxx, data_org, update_index, NULL); Amat->matvec(xxx, rhs, Amat, proc_config); */ } else { ch = getc(fp); if (ch == 'S') { while ( (ch = getc(fp)) != '\n') ; } else ungetc(ch,fp); for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) fscanf(fp,"%lf",&(rhs[i])); fclose(fp); } for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0; /* Set x */ fp = fopen("AZ_capture_init_guess.dat","r"); if (fp != NULL) { ch = getc(fp); if (ch == 'S') { while ( (ch = getc(fp)) != '\n') ; } else ungetc(ch,fp); for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) fscanf(fp,"%lf",&(xxx[i])); fclose(fp); options[AZ_conv] = AZ_expected_values; } /* if Dirichlet BC ... put the answer in */ for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) { if ( (val[i] > .99999999) && (val[i] < 1.0000001)) xxx[i] = rhs[i]; } fp = fopen("AZ_no_multilevel.dat","r"); scaling = AZ_scaling_create(); start_time = AZ_second(); if (fp != NULL) { fclose(fp); options[AZ_precond] = AZ_none; options[AZ_scaling] = AZ_sym_diag; options[AZ_ignore_scaling] = AZ_TRUE; options[AZ_keep_info] = 1; AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); /* options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); */ } else { options[AZ_keep_info] = 1; /* options[AZ_max_iter] = 40; */ AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); for (j = 0; j < Nrigid; j++) { alpha = -AZ_gdot(N_update, xxx, &(rigid[j*N_update]), proc_config)/AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); daxpy_(&N_update, &alpha, &(rigid[j*N_update]), &one, xxx, &one); printf("alpha5 is %e\n",alpha); } AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; /* if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); */ } solve_time = AZ_second() - start_time; if (proc_config[AZ_node] == 0) printf("Solve time = %e, MG Setup time = %e\n", solve_time, setup_time); ML_Aggregate_Destroy(&ag); ML_Destroy(&ml); AZ_free((void *) Amat->data_org); AZ_free((void *) Amat->val); AZ_free((void *) Amat->bindx); AZ_free((void *) update); AZ_free((void *) external); AZ_free((void *) extern_index); AZ_free((void *) update_index); if (Amat != NULL) AZ_matrix_destroy(&Amat); if (Pmat != NULL) AZ_precond_destroy(&Pmat); free(xxx); free(rhs); #ifdef ML_MPI MPI_Finalize(); #endif return 0; }
int main(int argc, char *argv[]) { int num_PDE_eqns=6, N_levels=4, nsmooth=2; int leng, level, N_grid_pts, coarsest_level; /* See Aztec User's Guide for more information on the */ /* variables that follow. */ int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE]; double params[AZ_PARAMS_SIZE], status[AZ_STATUS_SIZE]; /* data structure for matrix corresponding to the fine grid */ double *val = NULL, *xxx, *rhs, solve_time, setup_time, start_time; AZ_MATRIX *Amat; AZ_PRECOND *Pmat = NULL; ML *ml; FILE *fp; int i, j, Nrigid, *garbage = NULL; #ifdef ML_partition int nblocks; int *block_list = NULL; int k; #endif struct AZ_SCALING *scaling; ML_Aggregate *ag; double *mode, *rigid; char filename[80]; double alpha; int allocated = 0; int old_prec, old_sol; double old_tol; /* double *Amode, beta, biggest; int big_ind = -1, ii; */ ML_Operator *Amatrix; int *rowi_col = NULL, rowi_N, count2, ccc; double *rowi_val = NULL; double max_diag, min_diag, max_sum, sum; int nBlocks, *blockIndices, Ndof; #ifdef ML_partition FILE *fp2; int count; if (argc != 2) { printf("Usage: ml_read_elas num_processors\n"); exit(1); } else sscanf(argv[1],"%d",&nblocks); #endif #ifdef HAVE_MPI MPI_Init(&argc,&argv); /* get number of processors and the name of this processor */ AZ_set_proc_config(proc_config, MPI_COMM_WORLD); #else AZ_set_proc_config(proc_config, AZ_NOT_MPI); #endif /* read in the number of matrix equations */ leng = 0; if (proc_config[AZ_node] == 0) { # ifdef binary fp=fopen(".data","rb"); # else fp=fopen(".data","r"); # endif if (fp==NULL) { printf("couldn't open file .data\n"); exit(1); } # ifdef binary fread(&leng, sizeof(int), 1, fp); # else fscanf(fp,"%d",&leng); # endif fclose(fp); } leng = AZ_gsum_int(leng, proc_config); N_grid_pts=leng/num_PDE_eqns; /* initialize the list of global indices. NOTE: the list of global */ /* indices must be in ascending order so that subsequent calls to */ /* AZ_find_index() will function properly. */ if (proc_config[AZ_N_procs] == 1) i = AZ_linear; else i = AZ_file; AZ_read_update(&N_update, &update, proc_config, N_grid_pts, num_PDE_eqns,i); AZ_read_msr_matrix(update, &val, &bindx, N_update, proc_config); /* This code is to fix things up so that we are sure we have */ /* all block (including the ghost nodes the same size. */ AZ_block_MSR(&bindx, &val, N_update, num_PDE_eqns, update); AZ_transform_norowreordering(proc_config, &external, bindx, val, update, &update_index, &extern_index, &data_org, N_update, 0, 0, 0, &cpntr, AZ_MSR_MATRIX); Amat = AZ_matrix_create( leng ); AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL); Amat->matrix_type = data_org[AZ_matrix_type]; data_org[AZ_N_rows] = data_org[AZ_N_internal] + data_org[AZ_N_border]; #ifdef SCALE_ME ML_MSR_sym_diagonal_scaling(Amat, proc_config, &scaling_vect); #endif start_time = AZ_second(); options[AZ_scaling] = AZ_none; ML_Create(&ml, N_levels); ML_Set_PrintLevel(10); /* set up discretization matrix and matrix vector function */ AZ_ML_Set_Amat(ml, N_levels-1, N_update, N_update, Amat, proc_config); #ifdef ML_partition /* this code is meant to partition the matrices so that things can be */ /* run in parallel later. */ /* It is meant to be run on only one processor. */ #ifdef MB_MODIF fp2 = fopen(".update","w"); #else fp2 = fopen("partition_file","w"); #endif ML_Operator_AmalgamateAndDropWeak(&(ml->Amat[N_levels-1]), num_PDE_eqns, 0.0); ML_Gen_Blocks_Metis(ml, N_levels-1, &nblocks, &block_list); for (i = 0; i < nblocks; i++) { count = 0; for (j = 0; j < ml->Amat[N_levels-1].outvec_leng; j++) { if (block_list[j] == i) count++; } fprintf(fp2," %d\n",count*num_PDE_eqns); for (j = 0; j < ml->Amat[N_levels-1].outvec_leng; j++) { if (block_list[j] == i) { for (k = 0; k < num_PDE_eqns; k++) fprintf(fp2,"%d\n",j*num_PDE_eqns+k); } } } fclose(fp2); ML_Operator_UnAmalgamateAndDropWeak(&(ml->Amat[N_levels-1]),num_PDE_eqns,0.0); #ifdef MB_MODIF printf(" partition file dumped in .update\n"); #endif exit(1); #endif ML_Aggregate_Create( &ag ); /* ML_Aggregate_Set_CoarsenScheme_MIS(ag); */ #ifdef MB_MODIF ML_Aggregate_Set_DampingFactor(ag,1.50); #else ML_Aggregate_Set_DampingFactor(ag,1.5); #endif ML_Aggregate_Set_CoarsenScheme_METIS(ag); ML_Aggregate_Set_NodesPerAggr( ml, ag, -1, 35); /* ML_Aggregate_Set_Phase3AggregateCreationAggressiveness(ag, 10.001); */ ML_Aggregate_Set_Threshold(ag, 0.0); ML_Aggregate_Set_MaxCoarseSize( ag, 300); /* read in the rigid body modes */ Nrigid = 0; /* to ensure compatibility with RBM dumping software */ if (proc_config[AZ_node] == 0) { sprintf(filename,"rigid_body_mode%02d",Nrigid+1); while( (fp = fopen(filename,"r")) != NULL) { which_filename = 1; fclose(fp); Nrigid++; sprintf(filename,"rigid_body_mode%02d",Nrigid+1); } sprintf(filename,"rigid_body_mode%d",Nrigid+1); while( (fp = fopen(filename,"r")) != NULL) { fclose(fp); Nrigid++; sprintf(filename,"rigid_body_mode%d",Nrigid+1); } } Nrigid = AZ_gsum_int(Nrigid,proc_config); if (Nrigid != 0) { rigid = (double *) ML_allocate( sizeof(double)*Nrigid*(N_update+1) ); if (rigid == NULL) { printf("Error: Not enough space for rigid body modes\n"); } } rhs = (double *) malloc(leng*sizeof(double)); xxx = (double *) malloc(leng*sizeof(double)); for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0; for (i = 0; i < Nrigid; i++) { if (which_filename == 1) sprintf(filename,"rigid_body_mode%02d",i+1); else sprintf(filename,"rigid_body_mode%d",i+1); AZ_input_msr_matrix(filename,update,&mode,&garbage,N_update,proc_config); AZ_reorder_vec(mode, data_org, update_index, NULL); /* here is something to stick a rigid body mode as the initial */ /* The idea is to solve A x = 0 without smoothing with a two */ /* level method. If everything is done properly, we should */ /* converge in 2 iterations. */ /* Note: we must also zero out components of the rigid body */ /* mode that correspond to Dirichlet bcs. */ if (i == -4) { for (iii = 0; iii < leng; iii++) xxx[iii] = mode[iii]; ccc = 0; Amatrix = &(ml->Amat[N_levels-1]); for (iii = 0; iii < Amatrix->outvec_leng; iii++) { ML_get_matrix_row(Amatrix,1,&iii,&allocated,&rowi_col,&rowi_val, &rowi_N, 0); count2 = 0; for (j = 0; j < rowi_N; j++) if (rowi_val[j] != 0.) count2++; if (count2 <= 1) { xxx[iii] = 0.; ccc++; } } free(rowi_col); free(rowi_val); allocated = 0; rowi_col = NULL; rowi_val = NULL; } /* * Rescale matrix/rigid body modes and checking * AZ_sym_rescale_sl(mode, Amat->data_org, options, proc_config, scaling); Amat->matvec(mode, rigid, Amat, proc_config); for (j = 0; j < N_update; j++) printf("this is %d %e\n",j,rigid[j]); */ /* Here is some code to check that the rigid body modes are */ /* really rigid body modes. The idea is to multiply by A and */ /* then to zero out things that we "think" are boundaries. */ /* In this hardwired example, things near boundaries */ /* correspond to matrix rows that do not have 81 nonzeros. */ /* Amode = (double *) malloc(leng*sizeof(double)); Amat->matvec(mode, Amode, Amat, proc_config); j = 0; biggest = 0.0; for (ii = 0; ii < N_update; ii++) { if ( Amat->bindx[ii+1] - Amat->bindx[ii] != 80) { Amode[ii] = 0.; j++; } else { if ( fabs(Amode[ii]) > biggest) { biggest=fabs(Amode[ii]); big_ind = ii; } } } printf("%d entries zeroed out of %d elements\n",j,N_update); alpha = AZ_gdot(N_update, Amode, Amode, proc_config); beta = AZ_gdot(N_update, mode, mode, proc_config); printf("||A r||^2 =%e, ||r||^2 = %e, ratio = %e\n", alpha,beta,alpha/beta); printf("the biggest is %e at row %d\n",biggest,big_ind); free(Amode); */ /* orthogonalize mode with respect to previous modes. */ for (j = 0; j < i; j++) { alpha = -AZ_gdot(N_update, mode, &(rigid[j*N_update]), proc_config)/ AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]), proc_config); /* daxpy_(&N_update,&alpha,&(rigid[j*N_update]), &one, mode, &one); */ } #ifndef MB_MODIF printf(" after mb %e %e %e\n",mode[0],mode[1],mode[2]); #endif for (j = 0; j < N_update; j++) rigid[i*N_update+j] = mode[j]; free(mode); free(garbage); garbage = NULL; } if (Nrigid != 0) { ML_Aggregate_Set_BlockDiagScaling(ag); ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update); free(rigid); } #ifdef SCALE_ME ML_Aggregate_Scale_NullSpace(ag, scaling_vect, N_update); #endif coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, N_levels-1, ML_DECREASING, ag); AZ_defaults(options, params); coarsest_level = N_levels - coarsest_level; if ( proc_config[AZ_node] == 0 ) printf("Coarse level = %d \n", coarsest_level); /* set up smoothers */ for (level = N_levels-1; level > coarsest_level; level--) { /* ML_Gen_Smoother_BlockGaussSeidel(ml, level,ML_BOTH, 1, 1., num_PDE_eqns); */ /* Sparse approximate inverse smoother that acutally does both */ /* pre and post smoothing. */ /* ML_Gen_Smoother_ParaSails(ml , level, ML_PRESMOOTHER, nsmooth, parasails_sym, parasails_thresh, parasails_nlevels, parasails_filter, parasails_loadbal, parasails_factorized); */ /* This is the symmetric Gauss-Seidel smoothing that we usually use. */ /* In parallel, it is not a true Gauss-Seidel in that each processor */ /* does a Gauss-Seidel on its local submatrix independent of the */ /* other processors. */ /* ML_Gen_Smoother_Cheby(ml, level, ML_BOTH, 30., nsmooth); */ Ndof = ml->Amat[level].invec_leng; ML_Gen_Blocks_Aggregates(ag, level, &nBlocks, &blockIndices); ML_Gen_Smoother_BlockDiagScaledCheby(ml, level, ML_BOTH, 30.,nsmooth, nBlocks, blockIndices); /* ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.); */ /* This is a true Gauss Seidel in parallel. This seems to work for */ /* elasticity problems. However, I don't believe that this is very */ /* efficient in parallel. */ /* nblocks = ml->Amat[level].invec_leng/num_PDE_eqns; blocks = (int *) ML_allocate(sizeof(int)*N_update); for (i =0; i < ml->Amat[level].invec_leng; i++) blocks[i] = i/num_PDE_eqns; ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml , level, ML_PRESMOOTHER, nsmooth, 1., nblocks, blocks); ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml, level, ML_POSTSMOOTHER, nsmooth, 1., nblocks, blocks); free(blocks); */ /* Block Jacobi Smoothing */ /* nblocks = ml->Amat[level].invec_leng/num_PDE_eqns; blocks = (int *) ML_allocate(sizeof(int)*N_update); for (i =0; i < ml->Amat[level].invec_leng; i++) blocks[i] = i/num_PDE_eqns; ML_Gen_Smoother_VBlockJacobi(ml , level, ML_BOTH, nsmooth, ML_ONE_STEP_CG, nblocks, blocks); free(blocks); */ /* Jacobi Smoothing */ /* ML_Gen_Smoother_Jacobi(ml , level, ML_PRESMOOTHER, nsmooth, ML_ONE_STEP_CG); ML_Gen_Smoother_Jacobi(ml , level, ML_POSTSMOOTHER, nsmooth,ML_ONE_STEP_CG); */ /* This does a block Gauss-Seidel (not true GS in parallel) */ /* where each processor has 'nblocks' blocks. */ /* nblocks = 250; ML_Gen_Blocks_Metis(ml, level, &nblocks, &blocks); ML_Gen_Smoother_VBlockJacobi(ml , level, ML_BOTH, nsmooth,ML_ONE_STEP_CG, nblocks, blocks); free(blocks); */ num_PDE_eqns = 6; } /* Choose coarse grid solver: mls, superlu, symGS, or Aztec */ /* ML_Gen_Smoother_Cheby(ml, coarsest_level, ML_BOTH, 30., nsmooth); ML_Gen_CoarseSolverSuperLU( ml, coarsest_level); */ /* ML_Gen_Smoother_SymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.); */ old_prec = options[AZ_precond]; old_sol = options[AZ_solver]; old_tol = params[AZ_tol]; params[AZ_tol] = 1.0e-9; params[AZ_tol] = 1.0e-5; options[AZ_precond] = AZ_Jacobi; options[AZ_solver] = AZ_cg; options[AZ_poly_ord] = 1; options[AZ_conv] = AZ_r0; options[AZ_orth_kvecs] = AZ_TRUE; j = AZ_gsum_int(ml->Amat[coarsest_level].outvec_leng, proc_config); options[AZ_keep_kvecs] = j - 6; options[AZ_max_iter] = options[AZ_keep_kvecs]; ML_Gen_SmootherAztec(ml, coarsest_level, options, params, proc_config, status, options[AZ_keep_kvecs], ML_PRESMOOTHER, NULL); options[AZ_conv] = AZ_noscaled; options[AZ_keep_kvecs] = 0; options[AZ_orth_kvecs] = 0; options[AZ_precond] = old_prec; options[AZ_solver] = old_sol; params[AZ_tol] = old_tol; /* */ #ifdef RST_MODIF ML_Gen_Solver(ml, ML_MGV, N_levels-1, coarsest_level); #else #ifdef MB_MODIF ML_Gen_Solver(ml, ML_SAAMG, N_levels-1, coarsest_level); #else ML_Gen_Solver(ml, ML_MGFULLV, N_levels-1, coarsest_level); #endif #endif options[AZ_solver] = AZ_GMRESR; options[AZ_solver] = AZ_cg; options[AZ_scaling] = AZ_none; options[AZ_precond] = AZ_user_precond; options[AZ_conv] = AZ_r0; options[AZ_conv] = AZ_noscaled; options[AZ_output] = 1; options[AZ_max_iter] = 500; options[AZ_poly_ord] = 5; options[AZ_kspace] = 40; params[AZ_tol] = 4.8e-6; AZ_set_ML_preconditioner(&Pmat, Amat, ml, options); setup_time = AZ_second() - start_time; /* Set rhs */ fp = fopen("AZ_capture_rhs.dat","r"); if (fp == NULL) { AZ_random_vector(rhs, data_org, proc_config); if (proc_config[AZ_node] == 0) printf("taking random vector for rhs\n"); for (i = 0; i < -N_update; i++) { rhs[i] = (double) update[i]; rhs[i] = 7.; } } else { if (proc_config[AZ_node]== 0) printf("reading rhs guess from file\n"); AZ_input_msr_matrix("AZ_capture_rhs.dat", update, &rhs, &garbage, N_update, proc_config); free(garbage); } AZ_reorder_vec(rhs, data_org, update_index, NULL); printf("changing rhs by multiplying with A\n"); Amat->matvec(rhs, xxx, Amat, proc_config); for (i = 0; i < N_update; i++) rhs[i] = xxx[i]; fp = fopen("AZ_capture_init_guess.dat","r"); if (fp != NULL) { fclose(fp); if (proc_config[AZ_node]== 0) printf("reading initial guess from file\n"); AZ_input_msr_matrix("AZ_capture_init_guess.dat", update, &xxx, &garbage, N_update, proc_config); free(garbage); xxx = (double *) realloc(xxx, sizeof(double)*( Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border] + Amat->data_org[AZ_N_external])); } AZ_reorder_vec(xxx, data_org, update_index, NULL); /* if Dirichlet BC ... put the answer in */ /* for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) { if ( (val[i] > .99999999) && (val[i] < 1.0000001)) xxx[i] = rhs[i]; } */ fp = fopen("AZ_no_multilevel.dat","r"); scaling = AZ_scaling_create(); start_time = AZ_second(); if (fp != NULL) { fclose(fp); options[AZ_precond] = AZ_none; options[AZ_scaling] = AZ_sym_diag; options[AZ_ignore_scaling] = AZ_TRUE; options[AZ_keep_info] = 1; AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); /* options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling); */ } else { options[AZ_keep_info] = 1; options[AZ_conv] = AZ_noscaled; options[AZ_conv] = AZ_r0; params[AZ_tol] = 1.0e-7; /* ML_Iterate(ml, xxx, rhs); */ alpha = sqrt(AZ_gdot(N_update, xxx, xxx, proc_config)); printf("init guess = %e\n",alpha); alpha = sqrt(AZ_gdot(N_update, rhs, rhs, proc_config)); printf("rhs = %e\n",alpha); #ifdef SCALE_ME ML_MSR_scalerhs(rhs, scaling_vect, data_org[AZ_N_internal] + data_org[AZ_N_border]); ML_MSR_scalesol(xxx, scaling_vect, data_org[AZ_N_internal] + data_org[AZ_N_border]); #endif max_diag = 0.; min_diag = 1.e30; max_sum = 0.; for (i = 0; i < N_update; i++) { if (Amat->val[i] < 0.) printf("woops negative diagonal A(%d,%d) = %e\n", i,i,Amat->val[i]); if (Amat->val[i] > max_diag) max_diag = Amat->val[i]; if (Amat->val[i] < min_diag) min_diag = Amat->val[i]; sum = fabs(Amat->val[i]); for (j = Amat->bindx[i]; j < Amat->bindx[i+1]; j++) { sum += fabs(Amat->val[j]); } if (sum > max_sum) max_sum = sum; } printf("Largest diagonal = %e, min diag = %e large abs row sum = %e\n", max_diag, min_diag, max_sum); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); options[AZ_pre_calc] = AZ_reuse; options[AZ_conv] = AZ_expected_values; /* if (proc_config[AZ_node] == 0) printf("\n-------- Second solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); if (proc_config[AZ_node] == 0) printf("\n-------- Third solve with improved convergence test -----\n"); AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling); */ } solve_time = AZ_second() - start_time; if (proc_config[AZ_node] == 0) printf("Solve time = %e, MG Setup time = %e\n", solve_time, setup_time); if (proc_config[AZ_node] == 0) printf("Printing out a few entries of the solution ...\n"); for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 7) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 23) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 47) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 101) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} j = AZ_gsum_int(7, proc_config); /* sync processors */ for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++) if (update[j] == 171) {printf("solution(gid = %d) = %10.4e\n", update[j],xxx[update_index[j]]); fflush(stdout);} ML_Aggregate_Destroy(&ag); ML_Destroy(&ml); AZ_free((void *) Amat->data_org); AZ_free((void *) Amat->val); AZ_free((void *) Amat->bindx); AZ_free((void *) update); AZ_free((void *) external); AZ_free((void *) extern_index); AZ_free((void *) update_index); AZ_scaling_destroy(&scaling); if (Amat != NULL) AZ_matrix_destroy(&Amat); if (Pmat != NULL) AZ_precond_destroy(&Pmat); free(xxx); free(rhs); #ifdef HAVE_MPI MPI_Finalize(); #endif return 0; }
/*----------------------------------------------------------------------* | Constructor (public) m.gee 01/05| | IMPORTANT: | | No matter on which level we are here, the vector xfine is ALWAYS | | a fine grid vector here! | | this is the constructor for the ismatrixfree==false case *----------------------------------------------------------------------*/ ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel( int level, int nlevel, int printlevel, ML* ml, ML_Aggregate* ag,Epetra_CrsMatrix** P, ML_NOX::Ml_Nox_Fineinterface& interface, const Epetra_Comm& comm, const Epetra_Vector& xfine, bool ismatrixfree, bool matfreelev0, bool isnlnCG, int nitersCG, bool broyden, Epetra_CrsMatrix* Jac, string fsmoothertype, string smoothertype, string coarsesolvetype, int nsmooth_fine, int nsmooth, int nsmooth_coarse, double conv_normF, double conv_nupdate, int conv_maxiter,int numPDE, int nullspdim) : fineinterface_(interface), comm_(comm) { level_ = level; // this level nlevel_ = nlevel; // number of total levels ml_printlevel_ = printlevel; // printlevel ml_ = ml; // the global ML object ag_ = ag; // the global ML_Aggregate object thislevel_prec_ = 0; // this level's linear preconditioner thislevel_ml_ = 0; // this level's local ML object thislevel_ag_ = 0; // this level's local ML_Aggregate object coarseinterface_ = 0; // this level's coarse interface coarseprepost_ = 0; xthis_ = 0; // this level's current solution matching this level's map!!!! thislevel_A_ = 0; // this level's NOX Matrixfree operator SmootherA_ = 0; // this level's Epetra_CrsMatrix for thislevel_prec_ ismatrixfree_ = ismatrixfree; // matrixfree flag conv_normF_ = conv_normF; // NOX convergence test stuff conv_nupdate_ = conv_nupdate; conv_maxiter_ = conv_maxiter; absresid_ = 0; nupdate_ = 0; fv_ = 0; maxiters_ = 0; combo1_ = 0; combo2_ = 0; thislevel_linSys_ = 0; // this level's NOX linear system nlParams_ = 0; // NOX parameters initialGuess_ = 0; // NOX initial guess group_ = 0; // NOX group solver_ = 0; // NOX solver isnlnCG_ = isnlnCG; azlinSys_ = 0; clone_ = 0; nitersCG_ = nitersCG; broyden_ = broyden; Broyd_ = 0; if (ismatrixfree_==true) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ismatrixfree_==true on level " << level_ << "\n" << "**ERR**: in constructor for ismatrixfree_==false - case\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // ------------------------------------------------------------------------ // get the Jacobian of this level const Epetra_CrsGraph* graph = 0; // ------------------------------------------------------------------------ if (level_==0) { graph = fineinterface_.getGraph(); // On fine level this is the fineinterface's Jacobian if (matfreelev0==false) SmootherA_ = fineinterface_.getJacobian(); else if (matfreelev0==true && Jac) SmootherA_ = Jac; else { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: something weired happened\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } } // ------------------------------------------------------------------------ else { // On coarse levels get Jacobian from hierarchy // Note: On levels>0 SmootherA_ is a real copy of the Jacobian int maxnnz=0; double cputime=0.0; ML_Operator2EpetraCrsMatrix(&(ml_->Amat[level_]), SmootherA_, maxnnz, false, cputime); SmootherA_->OptimizeStorage(); graph = &(SmootherA_->Graph()); } // just to be save if (!SmootherA_ || !graph) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: Smoother==NULL on level " << level_ << "\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // ------------------------------------------------------------------------ // generate this level's coarse interface coarseinterface_ = new ML_NOX::Nox_CoarseProblem_Interface( fineinterface_,level_,ml_printlevel_, P,&(graph->RowMap()),nlevel_); // ------------------------------------------------------------------------ // generate this level's coarse prepostoperator if (level_==0) coarseprepost_ = new ML_NOX::Ml_Nox_CoarsePrePostOperator(*coarseinterface_, fineinterface_); // ------------------------------------------------------------------------ // get the current solution to this level xthis_ = coarseinterface_->restrict_fine_to_this(xfine); // ------------------------------------------------------------------------ // create this level's preconditioner // We use a 1-level ML-hierarchy for that ML_Aggregate_Create(&thislevel_ag_); ML_Create(&thislevel_ml_,1); // set the Jacobian on level 0 of the local ml EpetraMatrix2MLMatrix(thislevel_ml_,0, (dynamic_cast<Epetra_RowMatrix*>(SmootherA_))); // construct a 1-level ML-hierarchy on this level as a smoother ML_Set_PrintLevel(ml_printlevel_); ML_Aggregate_Set_CoarsenScheme_Uncoupled(thislevel_ag_); ML_Aggregate_Set_DampingFactor(thislevel_ag_, 0.0); ML_Aggregate_Set_Threshold(thislevel_ag_, 0.0); ML_Aggregate_Set_MaxCoarseSize(thislevel_ag_,1); ML_Aggregate_Set_NullSpace(thislevel_ag_,numPDE,nullspdim,NULL, SmootherA_->NumMyRows()); int thislevel_nlevel = ML_Gen_MGHierarchy_UsingAggregation(thislevel_ml_,0, ML_INCREASING,thislevel_ag_); if (thislevel_nlevel != 1) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ML generated a local hierarchy of " << thislevel_nlevel << " on level " << level_ << "\n" << "**ERR**: this is supposed to be 1 Level only!\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // set the smoother if (level_==0) Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,fsmoothertype,nsmooth_fine); else if (level_ != nlevel_-1) // set the smoother from the input Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,smoothertype,nsmooth); else // set the coarse solver from the input Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,coarsesolvetype,nsmooth_coarse); // create this level's preconditioner class ML_Epetra::MultiLevelOperator* ml_tmp = new ML_Epetra::MultiLevelOperator( thislevel_ml_,comm_, SmootherA_->OperatorDomainMap(), SmootherA_->OperatorRangeMap()); thislevel_prec_ = new ML_NOX::ML_Nox_ConstrainedMultiLevelOperator(ml_tmp,*coarseinterface_); if (!thislevel_prec_) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: thislevel_prec_==NULL on level " << level_ << "\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // ------------------------------------------------------------------------ // set up NOX on this level // ------------------------------------------------------------------------ nlParams_ = new Teuchos::ParameterList(); Teuchos::ParameterList& printParams = nlParams_->sublist("Printing"); printParams.setParameter("MyPID", comm_.MyPID()); printParams.setParameter("Output Precision", 14); printParams.setParameter("Output Processor", 0); if (ml_printlevel_>9) printParams.setParameter("Output Information", NOX::Utils::OuterIteration + NOX::Utils::Warning); else if (ml_printlevel_>8) printParams.setParameter("Output Information", NOX::Utils::Warning); else printParams.setParameter("Output Information",0); if (level_==0) nlParams_->sublist("Solver Options").setParameter("User Defined Pre/Post Operator", *coarseprepost_); nlParams_->setParameter("Nonlinear Solver", "Line Search Based"); Teuchos::ParameterList& searchParams = nlParams_->sublist("Line Search"); Teuchos::ParameterList* lsParamsptr = 0; if (isnlnCG_) { searchParams.setParameter("Method", "NonlinearCG"); Teuchos::ParameterList& dirParams = nlParams_->sublist("Direction"); dirParams.setParameter("Method", "NonlinearCG"); Teuchos::ParameterList& nlcgParams = dirParams.sublist("Nonlinear CG"); nlcgParams.setParameter("Restart Frequency", 10); nlcgParams.setParameter("Precondition", "On"); nlcgParams.setParameter("Orthogonalize", "Polak-Ribiere"); //nlcgParams.setParameter("Orthogonalize", "Fletcher-Reeves"); Teuchos::ParameterList& lsParams = nlcgParams.sublist("Linear Solver"); lsParams.setParameter("Aztec Solver", "CG"); lsParams.setParameter("Max Iterations", 1); lsParams.setParameter("Tolerance", 1e-11); lsParams.setParameter("Output Frequency", 0); lsParams.setParameter("Preconditioning", "User Supplied Preconditioner"); lsParams.setParameter("Preconditioner","User Defined"); } else // Newton's method using ML-preconditioned Aztec as linear solver { searchParams.setParameter("Method", "Full Step"); // Sublist for direction Teuchos::ParameterList& dirParams = nlParams_->sublist("Direction"); dirParams.setParameter("Method", "Newton"); Teuchos::ParameterList& newtonParams = dirParams.sublist("Newton"); newtonParams.setParameter("Forcing Term Method", "Constant"); //newtonParams.setParameter("Forcing Term Method", "Type 1"); //newtonParams.setParameter("Forcing Term Method", "Type 2"); newtonParams.setParameter("Forcing Term Minimum Tolerance", 1.0e-6); newtonParams.setParameter("Forcing Term Maximum Tolerance", 0.1); Teuchos::ParameterList& lsParams = newtonParams.sublist("Linear Solver"); lsParamsptr = &lsParams; lsParams.setParameter("Size of Krylov Subspace", 100); lsParams.setParameter("Aztec Solver", "GMRES"); lsParams.setParameter("Max Iterations", nitersCG_); lsParams.setParameter("Tolerance", conv_normF_); // FIXME? is this correct? if (ml_printlevel_>8) lsParams.setParameter("Output Frequency", 50); else lsParams.setParameter("Output Frequency", 0); lsParams.setParameter("Preconditioning", "User Supplied Preconditioner"); lsParams.setParameter("Preconditioner","User Defined"); } // create the initial guess initialGuess_ = new NOX::Epetra::Vector(*xthis_, NOX::DeepCopy, true); // NOTE: do not delete xthis_, it's used and destroyed by initialGuess_ // create the necessary interfaces NOX::EpetraNew::Interface::Preconditioner* iPrec = 0; NOX::EpetraNew::Interface::Required* iReq = 0; NOX::EpetraNew::Interface::Jacobian* iJac = 0; if (isnlnCG_) { // create the matrixfree operator used in the nlnCG thislevel_A_ = new NOX::EpetraNew::MatrixFree(*coarseinterface_,*xthis_,false); // create the necessary interfaces iPrec = 0; iReq = coarseinterface_; iJac = thislevel_A_; // create the linear system thislevel_linSys_ = new ML_NOX::Ml_Nox_LinearSystem( *iJac,*thislevel_A_,*iPrec, coarseinterface_,*thislevel_prec_, *xthis_,ismatrixfree_,level_,ml_printlevel_); // create the group group_ = new NOX::EpetraNew::Group(printParams,*iReq,*initialGuess_,*thislevel_linSys_); } else // Modified Newton's method { if (!broyden_) { // create the necessary interfaces iPrec = this; iReq = coarseinterface_; //iJac = this; thislevel_A_ = new NOX::EpetraNew::MatrixFree(*coarseinterface_,*xthis_,false); // create the initial guess vector //clone_ = new Epetra_Vector(*xthis_); // create the linear system //azlinSys_ = new NOX::EpetraNew::LinearSystemAztecOO( // printParams,*lsParamsptr, // *iJac,*SmootherA_,*iPrec, // *thislevel_prec_,*clone_); azlinSys_ = new NOX::EpetraNew::LinearSystemAztecOO( printParams,*lsParamsptr, *thislevel_A_,*thislevel_A_,*iPrec, *thislevel_prec_,*xthis_); } else // use a Broyden update for the Jacobian { // create the initial guess vector //clone_ = new Epetra_Vector(*xthis_); // create the necessary interfaces iPrec = this; iReq = coarseinterface_; Broyd_ = new NOX::EpetraNew::BroydenOperator(*nlParams_,*xthis_, *SmootherA_,false); // create the linear system azlinSys_ = new NOX::EpetraNew::LinearSystemAztecOO( printParams,*lsParamsptr, *Broyd_,*SmootherA_,*iPrec, *thislevel_prec_,*xthis_); } // create the group group_ = new NOX::EpetraNew::Group(printParams,*iReq,*initialGuess_, *azlinSys_); } // create convergence test create_Nox_Convergencetest(conv_normF_,conv_nupdate_,conv_maxiter_); // create the solver solver_ = new NOX::Solver::Manager(*group_,*combo2_,*nlParams_); return; }
/*----------------------------------------------------------------------* | Constructor (public) m.gee 01/05| | IMPORTANT: | | No matter on which level we are here, the vector xfine is ALWAYS | | a fine grid vector here! | | this is the constructor for the ismatrixfree==true case *----------------------------------------------------------------------*/ ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel( int level, int nlevel, int printlevel, ML* ml, ML_Aggregate* ag,Epetra_CrsMatrix** P, ML_NOX::Ml_Nox_Fineinterface& interface, const Epetra_Comm& comm, const Epetra_Vector& xfine, bool ismatrixfree, bool isnlnCG, int nitersCG, bool broyden, string fsmoothertype, string smoothertype, string coarsesolvetype, int nsmooth_fine, int nsmooth, int nsmooth_coarse, double conv_normF, double conv_nupdate, int conv_maxiter, int numPDE, int nullspdim, Epetra_CrsMatrix* Mat, ML_NOX::Nox_CoarseProblem_Interface* coarseinterface) : fineinterface_(interface), comm_(comm) { level_ = level; // this level nlevel_ = nlevel; // number of total levels ml_printlevel_ = printlevel; // printlevel ml_ = ml; // the global ML object ag_ = ag; // the global ML_Aggregate object thislevel_prec_ = 0; // this level's linear preconditioner thislevel_ml_ = 0; // this level's local ML object thislevel_ag_ = 0; // this level's local ML_Aggregate object coarseinterface_ = coarseinterface; // this level's coarse interface coarseprepost_ = 0; xthis_ = 0; // this level's current solution matching this level's map!!!! thislevel_A_ = 0; // this level's NOX Matrixfree operator SmootherA_ = 0; // this level's Epetra_CrsMatrix for thislevel_prec_ ismatrixfree_ = ismatrixfree; // matrixfree flag conv_normF_ = conv_normF; // NOX convergence test stuff conv_nupdate_ = conv_nupdate; conv_maxiter_ = conv_maxiter; absresid_ = 0; nupdate_ = 0; fv_ = 0; maxiters_ = 0; combo1_ = 0; combo2_ = 0; thislevel_linSys_ = 0; // this level's NOX linear system nlParams_ = 0; // NOX parameters initialGuess_ = 0; // NOX initial guess group_ = 0; // NOX group solver_ = 0; // NOX solver SmootherA_ = Mat; isnlnCG_ = isnlnCG; azlinSys_ = 0; clone_ = 0; nitersCG_ = nitersCG; broyden_ = broyden; Broyd_ = 0; #if 0 if (isnlnCG_==false && (fsmoothertype == "Jacobi" || smoothertype == "Jacobi" || coarsesolvetype == "Jacobi" )) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: Modified Newton's method not supported for \n" << "**ERR**: ismatrixfree_==true && smoothertype == Jacobi-Smoother\n" << "**ERR**: because no full Jacobian exists!\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } #endif if (ismatrixfree_==false) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ismatrixfree_==false on level " << level_ << "\n" << "**ERR**: in constructor for ismatrixfree_==true - case\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } if (!coarseinterface_) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ptr to coarseinterface=NULL on level " << level_ << "\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } if (!Mat) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ptr to Matrix Mat=NULL on level " << level_ << "\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // ------------------------------------------------------------------------ Mat->OptimizeStorage(); // ------------------------------------------------------------------------ // get the current solution to this level xthis_ = coarseinterface_->restrict_fine_to_this(xfine); // ------------------------------------------------------------------------ // create this level's preconditioner // We use a 1-level ML-hierarchy for that ML_Aggregate_Create(&thislevel_ag_); ML_Create(&thislevel_ml_,1); // ------------------------------------------------------------------------ // set the Jacobian on level 0 of the local ml EpetraMatrix2MLMatrix(thislevel_ml_,0, (dynamic_cast<Epetra_RowMatrix*>(Mat))); // ------------------------------------------------------------------------ // construct a 1-level ML-hierarchy on this level as a smoother // ------------------------------------------------------------------------ ML_Set_PrintLevel(ml_printlevel_); ML_Aggregate_Set_CoarsenScheme_Uncoupled(thislevel_ag_); ML_Aggregate_Set_DampingFactor(thislevel_ag_, 0.0); ML_Aggregate_Set_Threshold(thislevel_ag_, 0.0); ML_Aggregate_Set_MaxCoarseSize(thislevel_ag_,1); ML_Aggregate_Set_NullSpace(thislevel_ag_,numPDE,nullspdim,NULL,Mat->NumMyRows()); int thislevel_nlevel = ML_Gen_MGHierarchy_UsingAggregation(thislevel_ml_,0, ML_INCREASING,thislevel_ag_); if (thislevel_nlevel != 1) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: ML generated a local hierarchy of " << thislevel_nlevel << " on level " << level_ << "\n" << "**ERR**: this is supposed to be 1 Level only!\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // set the smoother if (level_==0) Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,fsmoothertype,nsmooth_fine); else if (level_ != nlevel_-1) // set the smoother from the input Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,smoothertype,nsmooth); else // set the coarse solver from the input Set_Smoother(ml,ag,level_,nlevel,thislevel_ml_,thislevel_ag_,coarsesolvetype,nsmooth_coarse); // create this level's preconditioner class ML_Epetra::MultiLevelOperator* ml_tmp = new ML_Epetra::MultiLevelOperator( thislevel_ml_,comm_, Mat->OperatorDomainMap(), Mat->OperatorRangeMap()); thislevel_prec_ = new ML_NOX::ML_Nox_ConstrainedMultiLevelOperator(ml_tmp,*coarseinterface_); if (!thislevel_prec_) { cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::ML_Nox_NonlinearLevel:\n" << "**ERR**: thislevel_prec_==NULL on level " << level_ << "\n" << "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1; } // intensive test of this level's ML-smoother #if 0 { cout << "Test of smoother on level " << level_ << endl; Epetra_Vector *out = new Epetra_Vector(Copy,*xthis_,0); out->PutScalar(0.0); cout << "Input\n"; xthis_->PutScalar(1.0); Mat->Multiply(false,*xthis_,*out); xthis_->PutScalar(3.0); cout << "rhs\n"; cout << *out; double norm = 0.0; out->Norm1(&norm); cout << "Norm of rhs = " << norm << endl; thislevel_prec_->ApplyInverse(*out,*xthis_); cout << "result after smoother\n"; cout << *xthis_; delete out; out = 0; } if (level_==2) exit(0); #endif // ------------------------------------------------------------------------ // generate this level's coarse prepostoperator if (level_==0) coarseprepost_ = new ML_NOX::Ml_Nox_CoarsePrePostOperator(*coarseinterface_, fineinterface_); // ------------------------------------------------------------------------ // set up NOX on this level // ------------------------------------------------------------------------ nlParams_ = new Teuchos::ParameterList(); Teuchos::ParameterList& printParams = nlParams_->sublist("Printing"); printParams.setParameter("MyPID", comm_.MyPID()); printParams.setParameter("Output Precision", 9); printParams.setParameter("Output Processor", 0); if (ml_printlevel_>9) printParams.setParameter("Output Information", NOX::Utils::OuterIteration + //NOX::Utils::OuterIterationStatusTest + //NOX::Utils::InnerIteration + //NOX::Utils::Parameters + //NOX::Utils::Details + NOX::Utils::Warning); else if (ml_printlevel_>8) printParams.setParameter("Output Information", NOX::Utils::Warning); else printParams.setParameter("Output Information",0); if (level_==0) nlParams_->sublist("Solver Options").setParameter("User Defined Pre/Post Operator", *coarseprepost_); nlParams_->setParameter("Nonlinear Solver", "Line Search Based"); Teuchos::ParameterList& searchParams = nlParams_->sublist("Line Search"); Teuchos::ParameterList* lsParamsptr = 0; if (isnlnCG_) { searchParams.setParameter("Method", "NonlinearCG"); Teuchos::ParameterList& dirParams = nlParams_->sublist("Direction"); dirParams.setParameter("Method", "NonlinearCG"); Teuchos::ParameterList& nlcgParams = dirParams.sublist("Nonlinear CG"); nlcgParams.setParameter("Restart Frequency", 10); nlcgParams.setParameter("Precondition", "On"); nlcgParams.setParameter("Orthogonalize", "Polak-Ribiere"); //nlcgParams.setParameter("Orthogonalize", "Fletcher-Reeves"); Teuchos::ParameterList& lsParams = nlcgParams.sublist("Linear Solver"); lsParams.setParameter("Aztec Solver", "CG"); lsParams.setParameter("Max Iterations", 1); lsParams.setParameter("Tolerance", 1e-11); lsParams.setParameter("Output Frequency", 0); lsParams.setParameter("Preconditioning", "User Supplied Preconditioner"); lsParams.setParameter("Preconditioner","User Defined"); } else // Newton's method using ML-preconditioned Aztec as linear solver { searchParams.setParameter("Method", "Full Step"); // Sublist for direction Teuchos::ParameterList& dirParams = nlParams_->sublist("Direction"); dirParams.setParameter("Method", "Newton"); Teuchos::ParameterList& newtonParams = dirParams.sublist("Newton"); newtonParams.setParameter("Forcing Term Method", "Constant"); //newtonParams.setParameter("Forcing Term Method", "Type 1"); //newtonParams.setParameter("Forcing Term Method", "Type 2"); newtonParams.setParameter("Forcing Term Minimum Tolerance", 1.0e-6); newtonParams.setParameter("Forcing Term Maximum Tolerance", 0.1); Teuchos::ParameterList& lsParams = newtonParams.sublist("Linear Solver"); lsParamsptr = &lsParams; lsParams.setParameter("Aztec Solver", "CG"); lsParams.setParameter("Max Iterations", nitersCG_); lsParams.setParameter("Tolerance", conv_normF_); // FIXME? is this correct? if (ml_printlevel_>8) lsParams.setParameter("Output Frequency", 50); else lsParams.setParameter("Output Frequency", 0); lsParams.setParameter("Preconditioning", "User Supplied Preconditioner"); lsParams.setParameter("Preconditioner","User Defined"); } // create the initial guess initialGuess_ = new NOX::Epetra::Vector(*xthis_, NOX::DeepCopy, true); // NOTE: do not delete xthis_, it's used and destroyed by initialGuess_ // create the necessary interfaces NOX::EpetraNew::Interface::Preconditioner* iPrec = 0; NOX::EpetraNew::Interface::Required* iReq = 0; NOX::EpetraNew::Interface::Jacobian* iJac = 0; if (isnlnCG_) { // create the matrixfree operator used in the nlnCG thislevel_A_ = new NOX::EpetraNew::MatrixFree(*coarseinterface_,*xthis_,false); // create the necessary interfaces iPrec = 0; iReq = coarseinterface_; iJac = thislevel_A_; // create the linear system thislevel_linSys_ = new ML_NOX::Ml_Nox_LinearSystem( *iJac,*thislevel_A_,*iPrec, coarseinterface_,*thislevel_prec_, *xthis_,ismatrixfree_,level_,ml_printlevel_); // create the group group_ = new NOX::EpetraNew::Group(printParams,*iReq,*initialGuess_,*thislevel_linSys_); } else // Modified Newton's method { if (!broyden_) { // create the necessary interfaces iPrec = this; iReq = coarseinterface_; iJac = this; // create the initial guess vector clone_ = new Epetra_Vector(*xthis_); // create the linear system azlinSys_ = new NOX::EpetraNew::LinearSystemAztecOO( printParams,*lsParamsptr, *iJac,*SmootherA_,*iPrec, *thislevel_prec_,*clone_); } else { // create the initial guess vector clone_ = new Epetra_Vector(*xthis_); // create the necessary interfaces iPrec = this; iReq = coarseinterface_; Broyd_ = new NOX::EpetraNew::BroydenOperator(*nlParams_,*clone_, *SmootherA_,false); // create the linear system azlinSys_ = new NOX::EpetraNew::LinearSystemAztecOO( printParams,*lsParamsptr, *Broyd_,*SmootherA_,*iPrec, *thislevel_prec_,*clone_); } // create the group group_ = new NOX::EpetraNew::Group(printParams,*iReq,*initialGuess_,*azlinSys_); } // create convergence test create_Nox_Convergencetest(conv_normF_,conv_nupdate_,conv_maxiter_); // create the solver solver_ = new NOX::Solver::Manager(*group_,*combo2_,*nlParams_); return; }
int main(int argc, char *argv[]){ ML *ml_object; int i, N_grids = 3, N_levels; double sol[129], rhs[129]; ML_Aggregate *agg_object; ML_Operator *data; ML_Krylov *kdata; #ifdef ML_MPI MPI_Init(&argc,&argv); #endif for (i = 0; i < 129; i++) sol[i] = 0.; for (i = 0; i < 129; i++) rhs[i] = 2.; ML_Create (&ml_object, N_grids); ML_Init_Amatrix (ml_object, 0, 129, 129, NULL); ML_Set_Amatrix_Getrow(ml_object, 0, Poisson_getrow, NULL, 129); ML_Set_Amatrix_Matvec(ml_object, 0, Poisson_matvec); ML_Set_PrintLevel(10); ML_Aggregate_Create(&agg_object); ML_Aggregate_Set_MaxCoarseSize(agg_object,1); N_levels = ML_Gen_MGHierarchy_UsingAggregation(ml_object, 0, ML_INCREASING, agg_object); /******** Begin code to set a Jacobi smoother ****** ML_Gen_Smoother_Jacobi(ml_object, ML_ALL_LEVELS, ML_PRESMOOTHER, 1, ML_DEFAULT); ******** End code to set a Jacobi smoother ******/ /******** Begin code to set a user-defined smoother ******/ ML_Get_Amatrix(ml_object, 0, &data); ML_Set_Smoother(ml_object, 0, ML_BOTH, data, user_smoothing,"mine"); ML_Get_Amatrix(ml_object, 1, &data); ML_Set_Smoother(ml_object, 1, ML_BOTH, data, user_smoothing,"mine"); ML_Get_Amatrix(ml_object, 2, &data); ML_Set_Smoother(ml_object, 2, ML_BOTH, data, user_smoothing,"mine"); ML_Gen_Solver (ml_object, ML_MGV, 0, N_levels-1); /* This example uses an internal CG solver within ML */ /* ML has limited Krylov methods support. It is intended */ /* that ML be used with another package that supplies */ /* more sophisticated Krylov solver options (such as those */ /* found in the Trilinos or Aztec packages. */ kdata = ML_Krylov_Create(ml_object->comm); ML_Krylov_Set_PrintFreq( kdata, 1 ); ML_Krylov_Set_Method(kdata, ML_CG); ML_Krylov_Set_Amatrix(kdata, &(ml_object->Amat[0])); ML_Krylov_Set_PreconFunc(kdata, ML_MGVSolve_Wrapper); ML_Krylov_Set_Precon(kdata, ml_object); ML_Krylov_Set_Tolerance(kdata, 1.e-7); ML_Krylov_Solve(kdata, 129, rhs, sol); ML_Krylov_Destroy( &kdata ); ML_Aggregate_Destroy(&agg_object); ML_Destroy(&ml_object); /******** End code to set a user-defined smoother ******/ printf("answer is %e %e %e %e %e\n",sol[0],sol[1],sol[2],sol[3],sol[4]); #ifdef ML_MPI MPI_Finalize(); #endif exit(EXIT_SUCCESS); }