HYPRE_Int hypre_BoomerAMGCycle( void *amg_vdata, hypre_ParVector **F_array, hypre_ParVector **U_array ) { hypre_ParAMGData *amg_data = amg_vdata; HYPRE_Solver *smoother; /* Data Structure variables */ hypre_ParCSRMatrix **A_array; hypre_ParCSRMatrix **P_array; hypre_ParCSRMatrix **R_array; hypre_ParVector *Utemp; hypre_ParVector *Vtemp; hypre_ParVector *Rtemp; hypre_ParVector *Ptemp; hypre_ParVector *Ztemp; hypre_ParVector *Aux_U; hypre_ParVector *Aux_F; hypre_ParCSRBlockMatrix **A_block_array; hypre_ParCSRBlockMatrix **P_block_array; hypre_ParCSRBlockMatrix **R_block_array; HYPRE_Real *Ztemp_data; HYPRE_Real *Ptemp_data; HYPRE_Int **CF_marker_array; /* HYPRE_Int **unknown_map_array; HYPRE_Int **point_map_array; HYPRE_Int **v_at_point_array; */ HYPRE_Real cycle_op_count; HYPRE_Int cycle_type; HYPRE_Int num_levels; HYPRE_Int max_levels; HYPRE_Real *num_coeffs; HYPRE_Int *num_grid_sweeps; HYPRE_Int *grid_relax_type; HYPRE_Int **grid_relax_points; HYPRE_Int block_mode; HYPRE_Real *max_eig_est; HYPRE_Real *min_eig_est; HYPRE_Int cheby_order; HYPRE_Real cheby_fraction; /* Local variables */ HYPRE_Int *lev_counter; HYPRE_Int Solve_err_flag; HYPRE_Int k; HYPRE_Int i, j, jj; HYPRE_Int level; HYPRE_Int cycle_param; HYPRE_Int coarse_grid; HYPRE_Int fine_grid; HYPRE_Int Not_Finished; HYPRE_Int num_sweep; HYPRE_Int cg_num_sweep = 1; HYPRE_Int relax_type; HYPRE_Int relax_points; HYPRE_Int relax_order; HYPRE_Int relax_local; HYPRE_Int old_version = 0; HYPRE_Real *relax_weight; HYPRE_Real *omega; HYPRE_Real alfa, beta, gammaold; HYPRE_Real gamma = 1.0; HYPRE_Int local_size; /* HYPRE_Int *smooth_option; */ HYPRE_Int smooth_type; HYPRE_Int smooth_num_levels; HYPRE_Int num_threads, my_id; HYPRE_Real alpha; HYPRE_Real **l1_norms = NULL; HYPRE_Real *l1_norms_level; HYPRE_Int seq_cg = 0; MPI_Comm comm; #if 0 HYPRE_Real *D_mat; HYPRE_Real *S_vec; #endif /* Acquire data and allocate storage */ num_threads = hypre_NumThreads(); A_array = hypre_ParAMGDataAArray(amg_data); P_array = hypre_ParAMGDataPArray(amg_data); R_array = hypre_ParAMGDataRArray(amg_data); CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data); Vtemp = hypre_ParAMGDataVtemp(amg_data); Rtemp = hypre_ParAMGDataRtemp(amg_data); Ptemp = hypre_ParAMGDataPtemp(amg_data); Ztemp = hypre_ParAMGDataZtemp(amg_data); num_levels = hypre_ParAMGDataNumLevels(amg_data); max_levels = hypre_ParAMGDataMaxLevels(amg_data); cycle_type = hypre_ParAMGDataCycleType(amg_data); A_block_array = hypre_ParAMGDataABlockArray(amg_data); P_block_array = hypre_ParAMGDataPBlockArray(amg_data); R_block_array = hypre_ParAMGDataRBlockArray(amg_data); block_mode = hypre_ParAMGDataBlockMode(amg_data); num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data); grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data); grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data); relax_order = hypre_ParAMGDataRelaxOrder(amg_data); relax_weight = hypre_ParAMGDataRelaxWeight(amg_data); omega = hypre_ParAMGDataOmega(amg_data); smooth_type = hypre_ParAMGDataSmoothType(amg_data); smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data); l1_norms = hypre_ParAMGDataL1Norms(amg_data); /* smooth_option = hypre_ParAMGDataSmoothOption(amg_data); */ max_eig_est = hypre_ParAMGDataMaxEigEst(amg_data); min_eig_est = hypre_ParAMGDataMinEigEst(amg_data); cheby_order = hypre_ParAMGDataChebyOrder(amg_data); cheby_fraction = hypre_ParAMGDataChebyFraction(amg_data); cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data); lev_counter = hypre_CTAlloc(HYPRE_Int, num_levels); if (hypre_ParAMGDataParticipate(amg_data)) seq_cg = 1; /* Initialize */ Solve_err_flag = 0; if (grid_relax_points) old_version = 1; num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels); num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]); comm = hypre_ParCSRMatrixComm(A_array[0]); hypre_MPI_Comm_rank(comm,&my_id); if (block_mode) { for (j = 1; j < num_levels; j++) num_coeffs[j] = hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]); } else { for (j = 1; j < num_levels; j++) num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]); } /*--------------------------------------------------------------------- * Initialize cycling control counter * * Cycling is controlled using a level counter: lev_counter[k] * * Each time relaxation is performed on level k, the * counter is decremented by 1. If the counter is then * negative, we go to the next finer level. If non- * negative, we go to the next coarser level. The * following actions control cycling: * * a. lev_counter[0] is initialized to 1. * b. lev_counter[k] is initialized to cycle_type for k>0. * * c. During cycling, when going down to level k, lev_counter[k] * is set to the max of (lev_counter[k],cycle_type) *---------------------------------------------------------------------*/ Not_Finished = 1; lev_counter[0] = 1; for (k = 1; k < num_levels; ++k) { lev_counter[k] = cycle_type; } level = 0; cycle_param = 1; smoother = hypre_ParAMGDataSmoother(amg_data); if (smooth_num_levels > 0) { if (smooth_type == 7 || smooth_type == 8 || smooth_type == 17 || smooth_type == 18 || smooth_type == 9 || smooth_type == 19) { HYPRE_Int actual_local_size = hypre_ParVectorActualLocalSize(Vtemp); Utemp = hypre_ParVectorCreate(comm,hypre_ParVectorGlobalSize(Vtemp), hypre_ParVectorPartitioning(Vtemp)); hypre_ParVectorOwnsPartitioning(Utemp) = 0; local_size = hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp)); if (local_size < actual_local_size) { hypre_VectorData(hypre_ParVectorLocalVector(Utemp)) = hypre_CTAlloc(HYPRE_Complex, actual_local_size); hypre_ParVectorActualLocalSize(Utemp) = actual_local_size; } else hypre_ParVectorInitialize(Utemp); } } /*--------------------------------------------------------------------- * Main loop of cycling *--------------------------------------------------------------------*/ while (Not_Finished) { if (num_levels > 1) { local_size = hypre_VectorSize(hypre_ParVectorLocalVector(F_array[level])); hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp)) = local_size; if (smooth_num_levels <= level) { cg_num_sweep = 1; num_sweep = num_grid_sweeps[cycle_param]; Aux_U = U_array[level]; Aux_F = F_array[level]; } else if (smooth_type > 9) { hypre_VectorSize(hypre_ParVectorLocalVector(Ztemp)) = local_size; hypre_VectorSize(hypre_ParVectorLocalVector(Rtemp)) = local_size; hypre_VectorSize(hypre_ParVectorLocalVector(Ptemp)) = local_size; Ztemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ztemp)); Ptemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ptemp)); hypre_ParVectorSetConstantValues(Ztemp,0); alpha = -1.0; beta = 1.0; hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level], U_array[level], beta, F_array[level], Rtemp); cg_num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data); num_sweep = num_grid_sweeps[cycle_param]; Aux_U = Ztemp; Aux_F = Rtemp; } else { cg_num_sweep = 1; num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data); Aux_U = U_array[level]; Aux_F = F_array[level]; } relax_type = grid_relax_type[cycle_param]; } else /* AB: 4/08: removed the max_levels > 1 check - should do this when max-levels = 1 also */ { /* If no coarsening occurred, apply a simple smoother once */ Aux_U = U_array[level]; Aux_F = F_array[level]; num_sweep = 1; /* TK: Use the user relax type (instead of 0) to allow for setting a convergent smoother (e.g. in the solution of singular problems). */ relax_type = hypre_ParAMGDataUserRelaxType(amg_data); } if (l1_norms != NULL) l1_norms_level = l1_norms[level]; else l1_norms_level = NULL; if (cycle_param == 3 && seq_cg) { hypre_seqAMGCycle(amg_data, level, F_array, U_array); } else { /*------------------------------------------------------------------ * Do the relaxation num_sweep times *-----------------------------------------------------------------*/ for (jj = 0; jj < cg_num_sweep; jj++) { if (smooth_num_levels > level && smooth_type > 9) hypre_ParVectorSetConstantValues(Aux_U,0); for (j = 0; j < num_sweep; j++) { if (num_levels == 1 && max_levels > 1) { relax_points = 0; relax_local = 0; } else { if (old_version) relax_points = grid_relax_points[cycle_param][j]; relax_local = relax_order; } /*----------------------------------------------- * VERY sloppy approximation to cycle complexity *-----------------------------------------------*/ if (old_version && level < num_levels -1) { switch (relax_points) { case 1: cycle_op_count += num_coeffs[level+1]; break; case -1: cycle_op_count += (num_coeffs[level]-num_coeffs[level+1]); break; } } else { cycle_op_count += num_coeffs[level]; } /*----------------------------------------------- Choose Smoother -----------------------------------------------*/ if (smooth_num_levels > level && (smooth_type == 7 || smooth_type == 8 || smooth_type == 9 || smooth_type == 19 || smooth_type == 17 || smooth_type == 18)) { hypre_VectorSize(hypre_ParVectorLocalVector(Utemp)) = local_size; alpha = -1.0; beta = 1.0; hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level], U_array[level], beta, Aux_F, Vtemp); if (smooth_type == 8 || smooth_type == 18) HYPRE_ParCSRParaSailsSolve(smoother[level], (HYPRE_ParCSRMatrix) A_array[level], (HYPRE_ParVector) Vtemp, (HYPRE_ParVector) Utemp); else if (smooth_type == 7 || smooth_type == 17) HYPRE_ParCSRPilutSolve(smoother[level], (HYPRE_ParCSRMatrix) A_array[level], (HYPRE_ParVector) Vtemp, (HYPRE_ParVector) Utemp); else if (smooth_type == 9 || smooth_type == 19) HYPRE_EuclidSolve(smoother[level], (HYPRE_ParCSRMatrix) A_array[level], (HYPRE_ParVector) Vtemp, (HYPRE_ParVector) Utemp); hypre_ParVectorAxpy(relax_weight[level],Utemp,Aux_U); } else if (smooth_num_levels > level && (smooth_type == 6 || smooth_type == 16)) { HYPRE_SchwarzSolve(smoother[level], (HYPRE_ParCSRMatrix) A_array[level], (HYPRE_ParVector) Aux_F, (HYPRE_ParVector) Aux_U); } /*else if (relax_type == 99)*/ else if (relax_type == 9 || relax_type == 99) { /* Gaussian elimination */ hypre_GaussElimSolve(amg_data, level, relax_type); } else if (relax_type == 18) { /* L1 - Jacobi*/ if (relax_order == 1 && cycle_param < 3) { /* need to do CF - so can't use the AMS one */ HYPRE_Int i; HYPRE_Int loc_relax_points[2]; if (cycle_type < 2) { loc_relax_points[0] = 1; loc_relax_points[1] = -1; } else { loc_relax_points[0] = -1; loc_relax_points[1] = 1; } for (i=0; i < 2; i++) hypre_ParCSRRelax_L1_Jacobi(A_array[level], Aux_F, CF_marker_array[level], loc_relax_points[i], relax_weight[level], l1_norms[level], Aux_U, Vtemp); } else /* not CF - so use through AMS */ { if (num_threads == 1) hypre_ParCSRRelax(A_array[level], Aux_F, 1, 1, l1_norms_level, relax_weight[level], omega[level],0,0,0,0, Aux_U, Vtemp, Ztemp); else hypre_ParCSRRelaxThreads(A_array[level], Aux_F, 1, 1, l1_norms_level, relax_weight[level], omega[level], Aux_U, Vtemp, Ztemp); } } else if (relax_type == 15) { /* CG */ if (j ==0) /* do num sweep iterations of CG */ hypre_ParCSRRelax_CG( smoother[level], A_array[level], Aux_F, Aux_U, num_sweep); } else if (relax_type == 16) { /* scaled Chebyshev */ HYPRE_Int scale = 1; HYPRE_Int variant = 0; hypre_ParCSRRelax_Cheby(A_array[level], Aux_F, max_eig_est[level], min_eig_est[level], cheby_fraction, cheby_order, scale, variant, Aux_U, Vtemp, Ztemp ); } else if (relax_type ==17) { hypre_BoomerAMGRelax_FCFJacobi(A_array[level], Aux_F, CF_marker_array[level], relax_weight[level], Aux_U, Vtemp); } else if (old_version) { Solve_err_flag = hypre_BoomerAMGRelax(A_array[level], Aux_F, CF_marker_array[level], relax_type, relax_points, relax_weight[level], omega[level], l1_norms_level, Aux_U, Vtemp, Ztemp); } else { /* smoother than can have CF ordering */ if (block_mode) { Solve_err_flag = hypre_BoomerAMGBlockRelaxIF(A_block_array[level], Aux_F, CF_marker_array[level], relax_type, relax_local, cycle_param, relax_weight[level], omega[level], Aux_U, Vtemp); } else { Solve_err_flag = hypre_BoomerAMGRelaxIF(A_array[level], Aux_F, CF_marker_array[level], relax_type, relax_local, cycle_param, relax_weight[level], omega[level], l1_norms_level, Aux_U, Vtemp, Ztemp); } } if (Solve_err_flag != 0) return(Solve_err_flag); } if (smooth_num_levels > level && smooth_type > 9) { gammaold = gamma; gamma = hypre_ParVectorInnerProd(Rtemp,Ztemp); if (jj == 0) hypre_ParVectorCopy(Ztemp,Ptemp); else { beta = gamma/gammaold; for (i=0; i < local_size; i++) Ptemp_data[i] = Ztemp_data[i] + beta*Ptemp_data[i]; } hypre_ParCSRMatrixMatvec(1.0,A_array[level],Ptemp,0.0,Vtemp); alfa = gamma /hypre_ParVectorInnerProd(Ptemp,Vtemp); hypre_ParVectorAxpy(alfa,Ptemp,U_array[level]); hypre_ParVectorAxpy(-alfa,Vtemp,Rtemp); } } } /*------------------------------------------------------------------ * Decrement the control counter and determine which grid to visit next *-----------------------------------------------------------------*/ --lev_counter[level]; if (lev_counter[level] >= 0 && level != num_levels-1) { /*--------------------------------------------------------------- * Visit coarser level next. * Compute residual using hypre_ParCSRMatrixMatvec. * Perform restriction using hypre_ParCSRMatrixMatvecT. * Reset counters and cycling parameters for coarse level *--------------------------------------------------------------*/ fine_grid = level; coarse_grid = level + 1; hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0); alpha = -1.0; beta = 1.0; if (block_mode) { hypre_ParVectorCopy(F_array[fine_grid],Vtemp); hypre_ParCSRBlockMatrixMatvec(alpha, A_block_array[fine_grid], U_array[fine_grid], beta, Vtemp); } else { // JSP: avoid unnecessary copy using out-of-place version of SpMV hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[fine_grid], U_array[fine_grid], beta, F_array[fine_grid], Vtemp); } alpha = 1.0; beta = 0.0; if (block_mode) { hypre_ParCSRBlockMatrixMatvecT(alpha,R_block_array[fine_grid],Vtemp, beta,F_array[coarse_grid]); } else { hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp, beta,F_array[coarse_grid]); } ++level; lev_counter[level] = hypre_max(lev_counter[level],cycle_type); cycle_param = 1; if (level == num_levels-1) cycle_param = 3; } else if (level != 0) { /*--------------------------------------------------------------- * Visit finer level next. * Interpolate and add correction using hypre_ParCSRMatrixMatvec. * Reset counters and cycling parameters for finer level. *--------------------------------------------------------------*/ fine_grid = level - 1; coarse_grid = level; alpha = 1.0; beta = 1.0; if (block_mode) { hypre_ParCSRBlockMatrixMatvec(alpha, P_block_array[fine_grid], U_array[coarse_grid], beta, U_array[fine_grid]); } else { hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid], U_array[coarse_grid], beta, U_array[fine_grid]); } --level; cycle_param = 2; } else { Not_Finished = 0; } } hypre_ParAMGDataCycleOpCount(amg_data) = cycle_op_count; hypre_TFree(lev_counter); hypre_TFree(num_coeffs); if (smooth_num_levels > 0) { if (smooth_type == 7 || smooth_type == 8 || smooth_type == 9 || smooth_type == 17 || smooth_type == 18 || smooth_type == 19 ) hypre_ParVectorDestroy(Utemp); } return(Solve_err_flag); }
HYPRE_Int hypre_MaxwellSolve2( void * maxwell_vdata, hypre_SStructMatrix * A_in, hypre_SStructVector * f, hypre_SStructVector * u ) { hypre_MaxwellData *maxwell_data = maxwell_vdata; hypre_ParVector *f_edge; hypre_ParVector *u_edge; HYPRE_Int max_iter = maxwell_data-> max_iter; double tol = maxwell_data-> tol; HYPRE_Int rel_change = maxwell_data-> rel_change; HYPRE_Int zero_guess = maxwell_data-> zero_guess; HYPRE_Int npre_relax = maxwell_data-> num_pre_relax; HYPRE_Int npost_relax = maxwell_data-> num_post_relax; hypre_ParCSRMatrix **Ann_l = maxwell_data-> Ann_l; hypre_ParCSRMatrix **Pn_l = maxwell_data-> Pn_l; hypre_ParCSRMatrix **RnT_l = maxwell_data-> RnT_l; hypre_ParVector **bn_l = maxwell_data-> bn_l; hypre_ParVector **xn_l = maxwell_data-> xn_l; hypre_ParVector **resn_l = maxwell_data-> resn_l; hypre_ParVector **en_l = maxwell_data-> en_l; hypre_ParVector **nVtemp2_l = maxwell_data-> nVtemp2_l; HYPRE_Int **nCF_marker_l = maxwell_data-> nCF_marker_l; double *nrelax_weight= maxwell_data-> nrelax_weight; double *nomega = maxwell_data-> nomega; HYPRE_Int nrelax_type = maxwell_data-> nrelax_type; HYPRE_Int node_numlevs = maxwell_data-> node_numlevels; hypre_ParCSRMatrix *Tgrad = maxwell_data-> Tgrad; hypre_ParCSRMatrix *T_transpose = maxwell_data-> T_transpose; hypre_ParCSRMatrix **Aee_l = maxwell_data-> Aee_l; hypre_IJMatrix **Pe_l = maxwell_data-> Pe_l; hypre_IJMatrix **ReT_l = maxwell_data-> ReT_l; hypre_ParVector **be_l = maxwell_data-> be_l; hypre_ParVector **xe_l = maxwell_data-> xe_l; hypre_ParVector **rese_l = maxwell_data-> rese_l; hypre_ParVector **ee_l = maxwell_data-> ee_l; hypre_ParVector **eVtemp2_l = maxwell_data-> eVtemp2_l; HYPRE_Int **eCF_marker_l = maxwell_data-> eCF_marker_l; double *erelax_weight= maxwell_data-> erelax_weight; double *eomega = maxwell_data-> eomega; HYPRE_Int erelax_type = maxwell_data-> erelax_type; HYPRE_Int edge_numlevs = maxwell_data-> edge_numlevels; HYPRE_Int **BdryRanks_l = maxwell_data-> BdryRanks_l; HYPRE_Int *BdryRanksCnts_l= maxwell_data-> BdryRanksCnts_l; HYPRE_Int logging = maxwell_data-> logging; double *norms = maxwell_data-> norms; double *rel_norms = maxwell_data-> rel_norms; HYPRE_Int Solve_err_flag; HYPRE_Int relax_local, cycle_param; double b_dot_b = 0, r_dot_r, eps = 0; double e_dot_e, x_dot_x; HYPRE_Int i, j; HYPRE_Int level; HYPRE_Int ierr= 0; /* added for the relaxation routines */ hypre_ParVector *ze = NULL; if (hypre_NumThreads() > 1) { /* Aee is always bigger than Ann */ ze = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(Aee_l[0]), hypre_ParCSRMatrixGlobalNumRows(Aee_l[0]), hypre_ParCSRMatrixRowStarts(Aee_l[0])); hypre_ParVectorInitialize(ze); hypre_ParVectorSetPartitioningOwner(ze,0); } hypre_BeginTiming(maxwell_data-> time_index); hypre_SStructVectorConvert(f, &f_edge); hypre_SStructVectorConvert(u, &u_edge); hypre_ParVectorZeroBCValues(f_edge, BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParVectorZeroBCValues(u_edge, BdryRanks_l[0], BdryRanksCnts_l[0]); be_l[0]= f_edge; xe_l[0]= u_edge; /* the nodal fine vectors: xn= 0. bn= T'*(be- Aee*xe) is updated in the cycle. */ hypre_ParVectorSetConstantValues(xn_l[0], 0.0); relax_local= 0; cycle_param= 0; (maxwell_data-> num_iterations) = 0; /* if max_iter is zero, return */ if (max_iter == 0) { /* if using a zero initial guess, return zero */ if (zero_guess) { hypre_ParVectorSetConstantValues(xe_l[0], 0.0); } hypre_EndTiming(maxwell_data -> time_index); return ierr; } /* part of convergence check */ if (tol > 0.0) { /* eps = (tol^2) */ b_dot_b= hypre_ParVectorInnerProd(be_l[0], be_l[0]); eps = tol*tol; /* if rhs is zero, return a zero solution */ if (b_dot_b == 0.0) { hypre_ParVectorSetConstantValues(xe_l[0], 0.0); if (logging > 0) { norms[0] = 0.0; rel_norms[0] = 0.0; } hypre_EndTiming(maxwell_data -> time_index); return ierr; } } /*----------------------------------------------------- * Do V-cycles: * For each index l, "fine" = l, "coarse" = (l-1) * * solution update: * edge_sol= edge_sol + T*node_sol *-----------------------------------------------------*/ for (i = 0; i < max_iter; i++) { /* compute fine grid residual & nodal rhs. */ hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); hypre_ParVectorZeroBCValues(rese_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParCSRMatrixMatvec(1.0, T_transpose, rese_l[0], 0.0, bn_l[0]); /* convergence check */ if (tol > 0.0) { r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]); if (logging > 0) { norms[i] = sqrt(r_dot_r); if (b_dot_b > 0) rel_norms[i] = sqrt(r_dot_r/b_dot_b); else rel_norms[i] = 0.0; } /* always do at least 1 V-cycle */ if ((r_dot_r/b_dot_b < eps) && (i > 0)) { if (rel_change) { if ((e_dot_e/x_dot_x) < eps) break; } else { break; } } } hypre_ParVectorCopy(bn_l[0], resn_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]); r_dot_r= hypre_ParVectorInnerProd(resn_l[0], resn_l[0]); for (level= 0; level<= node_numlevs-2; level++) { /*----------------------------------------------- * Down cycle *-----------------------------------------------*/ for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); } /*for (j= 0; j< npre_relax; j++) */ /* compute residuals */ hypre_ParVectorCopy(bn_l[level], resn_l[level]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[level], xn_l[level], 1.0, resn_l[level]); /* restrict residuals */ hypre_ParCSRMatrixMatvecT(1.0, RnT_l[level], resn_l[level], 0.0, bn_l[level+1]); /* zero off initial guess for the next level */ hypre_ParVectorSetConstantValues(xn_l[level+1], 0.0); } /* for (level= 0; level<= node_numlevs-2; level++) */ /* coarsest node solve */ level= node_numlevs-1; Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); /*--------------------------------------------------------------------- * Cycle up the levels. *---------------------------------------------------------------------*/ for (level= (node_numlevs - 2); level>= 1; level--) { hypre_ParCSRMatrixMatvec(1.0, Pn_l[level], xn_l[level+1], 0.0, en_l[level]); hypre_ParVectorAxpy(1.0, en_l[level], xn_l[level]); /* post smooth */ for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level], bn_l[level], nCF_marker_l[level], nrelax_type, relax_local, cycle_param, nrelax_weight[level], nomega[level], NULL, xn_l[level], nVtemp2_l[level], ze); } } /* for (level= (en_numlevs - 2); level>= 1; level--) */ /* interpolate error and correct on finest grids */ hypre_ParCSRMatrixMatvec(1.0, Pn_l[0], xn_l[1], 0.0, en_l[0]); hypre_ParVectorAxpy(1.0, en_l[0], xn_l[0]); for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[0], bn_l[0], nCF_marker_l[0], nrelax_type, relax_local, cycle_param, nrelax_weight[0], nomega[0], NULL, xn_l[0], nVtemp2_l[0], ze); } /* for (j= 0; j< npost_relax; j++) */ hypre_ParVectorCopy(bn_l[0], resn_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]); /* add the gradient solution component to xe_l[0] */ hypre_ParCSRMatrixMatvec(1.0, Tgrad, xn_l[0], 1.0, xe_l[0]); hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]); for (level= 0; level<= edge_numlevs-2; level++) { /*----------------------------------------------- * Down cycle *-----------------------------------------------*/ for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } /*for (j= 0; j< npre_relax; j++) */ /* compute residuals */ hypre_ParVectorCopy(be_l[level], rese_l[level]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[level], xe_l[level], 1.0, rese_l[level]); /* restrict residuals */ hypre_ParCSRMatrixMatvecT(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(ReT_l[level]), rese_l[level], 0.0, be_l[level+1]); hypre_ParVectorZeroBCValues(be_l[level+1], BdryRanks_l[level+1], BdryRanksCnts_l[level+1]); /* zero off initial guess for the next level */ hypre_ParVectorSetConstantValues(xe_l[level+1], 0.0); } /* for (level= 1; level<= edge_numlevels-2; level++) */ /* coarsest edge solve */ level= edge_numlevs-1; for (j= 0; j< npre_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } /*--------------------------------------------------------------------- * Up cycle. *---------------------------------------------------------------------*/ for (level= (edge_numlevs - 2); level>= 1; level--) { hypre_ParCSRMatrixMatvec(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[level]), xe_l[level+1], 0.0, ee_l[level]); hypre_ParVectorZeroBCValues(ee_l[level], BdryRanks_l[level], BdryRanksCnts_l[level]); hypre_ParVectorAxpy(1.0, ee_l[level], xe_l[level]); /* post smooth */ for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level], be_l[level], eCF_marker_l[level], erelax_type, relax_local, cycle_param, erelax_weight[level], eomega[level], NULL, xe_l[level], eVtemp2_l[level], ze); } } /* for (level= (edge_numlevs - 2); level>= 1; level--) */ /* interpolate error and correct on finest grids */ hypre_ParCSRMatrixMatvec(1.0, (hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[0]), xe_l[1], 0.0, ee_l[0]); hypre_ParVectorZeroBCValues(ee_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]); hypre_ParVectorAxpy(1.0, ee_l[0], xe_l[0]); for (j= 0; j< npost_relax; j++) { Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[0], be_l[0], eCF_marker_l[0], erelax_type, relax_local, cycle_param, erelax_weight[0], eomega[0], NULL, xe_l[0], eVtemp2_l[0], ze); } /* for (j= 0; j< npost_relax; j++) */ e_dot_e= hypre_ParVectorInnerProd(ee_l[0], ee_l[0]); x_dot_x= hypre_ParVectorInnerProd(xe_l[0], xe_l[0]); hypre_ParVectorCopy(be_l[0], rese_l[0]); hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]); (maxwell_data -> num_iterations) = (i + 1); } hypre_EndTiming(maxwell_data -> time_index); if (ze) hypre_ParVectorDestroy(ze); return ierr; }
HYPRE_Int hypre_BoomerAMGAdditiveCycle( void *amg_vdata) { hypre_ParAMGData *amg_data = amg_vdata; /* Data Structure variables */ hypre_ParCSRMatrix **A_array; hypre_ParCSRMatrix **P_array; hypre_ParCSRMatrix **R_array; hypre_ParCSRMatrix *Lambda; hypre_ParVector **F_array; hypre_ParVector **U_array; hypre_ParVector *Vtemp; hypre_ParVector *Ztemp; hypre_ParVector *Xtilde, *Rtilde; HYPRE_Int **CF_marker_array; HYPRE_Int num_levels; HYPRE_Int addlvl; HYPRE_Int additive; HYPRE_Int mult_additive; HYPRE_Int simple; HYPRE_Int i, num_rows; HYPRE_Int n_global; HYPRE_Int rlx_order; /* Local variables */ HYPRE_Int Solve_err_flag = 0; HYPRE_Int level; HYPRE_Int coarse_grid; HYPRE_Int fine_grid; HYPRE_Int relax_type; HYPRE_Int rlx_down; HYPRE_Int rlx_up; HYPRE_Int *grid_relax_type; HYPRE_Real **l1_norms; HYPRE_Real alpha, beta; HYPRE_Int num_threads; HYPRE_Real *u_data; HYPRE_Real *f_data; HYPRE_Real *v_data; HYPRE_Real *l1_norms_lvl; HYPRE_Real *D_inv; HYPRE_Real *x_global; HYPRE_Real *r_global; HYPRE_Real *relax_weight; HYPRE_Real *omega; #if 0 HYPRE_Real *D_mat; HYPRE_Real *S_vec; #endif /* Acquire data and allocate storage */ num_threads = hypre_NumThreads(); A_array = hypre_ParAMGDataAArray(amg_data); F_array = hypre_ParAMGDataFArray(amg_data); U_array = hypre_ParAMGDataUArray(amg_data); P_array = hypre_ParAMGDataPArray(amg_data); R_array = hypre_ParAMGDataRArray(amg_data); CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data); Vtemp = hypre_ParAMGDataVtemp(amg_data); Ztemp = hypre_ParAMGDataZtemp(amg_data); num_levels = hypre_ParAMGDataNumLevels(amg_data); additive = hypre_ParAMGDataAdditive(amg_data); mult_additive = hypre_ParAMGDataMultAdditive(amg_data); simple = hypre_ParAMGDataSimple(amg_data); grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data); Lambda = hypre_ParAMGDataLambda(amg_data); Xtilde = hypre_ParAMGDataXtilde(amg_data); Rtilde = hypre_ParAMGDataRtilde(amg_data); l1_norms = hypre_ParAMGDataL1Norms(amg_data); D_inv = hypre_ParAMGDataDinv(amg_data); grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data); relax_weight = hypre_ParAMGDataRelaxWeight(amg_data); omega = hypre_ParAMGDataOmega(amg_data); rlx_order = hypre_ParAMGDataRelaxOrder(amg_data); /* Initialize */ addlvl = hypre_max(additive, mult_additive); addlvl = hypre_max(addlvl, simple); Solve_err_flag = 0; /*--------------------------------------------------------------------- * Main loop of cycling --- multiplicative version --- V-cycle *--------------------------------------------------------------------*/ /* down cycle */ relax_type = grid_relax_type[1]; rlx_down = grid_relax_type[1]; rlx_up = grid_relax_type[2]; for (level = 0; level < num_levels-1; level++) { fine_grid = level; coarse_grid = level + 1; u_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[fine_grid])); f_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[fine_grid])); v_data = hypre_VectorData(hypre_ParVectorLocalVector(Vtemp)); l1_norms_lvl = l1_norms[level]; hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0); if (level < addlvl) /* multiplicative version */ { /* smoothing step */ if (rlx_down == 0) { HYPRE_Real *A_data = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(A_array[fine_grid])); HYPRE_Int *A_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(A_array[fine_grid])); hypre_ParVectorCopy(F_array[fine_grid],Vtemp); num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid])); #ifdef HYPRE_USING_OPENMP #pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE #endif for (i = 0; i < num_rows; i++) u_data[i] = relax_weight[level]*v_data[i] / A_data[A_i[i]]; } else if (rlx_down != 18) { /*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_down,0,*/ hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid], CF_marker_array[fine_grid], rlx_down,rlx_order,1, relax_weight[fine_grid], omega[fine_grid], l1_norms_lvl, U_array[fine_grid], Vtemp, Ztemp); hypre_ParVectorCopy(F_array[fine_grid],Vtemp); } else { hypre_ParVectorCopy(F_array[fine_grid],Vtemp); num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid])); #ifdef HYPRE_USING_OPENMP #pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE #endif for (i = 0; i < num_rows; i++) u_data[i] += v_data[i] / l1_norms_lvl[i]; } alpha = -1.0; beta = 1.0; hypre_ParCSRMatrixMatvec(alpha, A_array[fine_grid], U_array[fine_grid], beta, Vtemp); alpha = 1.0; beta = 0.0; hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp, beta,F_array[coarse_grid]); } else /* additive version */ { hypre_ParVectorCopy(F_array[fine_grid],Vtemp); if (level == 0) /* compute residual */ { hypre_ParVectorCopy(Vtemp, Rtilde); hypre_ParVectorCopy(U_array[fine_grid],Xtilde); } alpha = 1.0; beta = 0.0; hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp, beta,F_array[coarse_grid]); } } /* solve coarse grid */ if (addlvl < num_levels) { if (simple > -1) { x_global = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde)); r_global = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde)); n_global = hypre_VectorSize(hypre_ParVectorLocalVector(Xtilde)); #ifdef HYPRE_USING_OPENMP #pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE #endif for (i=0; i < n_global; i++) x_global[i] += D_inv[i]*r_global[i]; } else hypre_ParCSRMatrixMatvec(1.0, Lambda, Rtilde, 1.0, Xtilde); if (addlvl == 0) hypre_ParVectorCopy(Xtilde, U_array[0]); } else { fine_grid = num_levels -1; hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid], 1, 1, l1_norms[fine_grid], 1.0, 1.0 ,0,0,0,0, U_array[fine_grid], Vtemp, Ztemp); } /* up cycle */ relax_type = grid_relax_type[2]; for (level = num_levels-1; level > 0; level--) { fine_grid = level - 1; coarse_grid = level; if (level <= addlvl) /* multiplicative version */ { alpha = 1.0; beta = 1.0; hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid], U_array[coarse_grid], beta, U_array[fine_grid]); if (rlx_up != 18) /*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_up,0,*/ hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid], CF_marker_array[fine_grid], rlx_up,rlx_order,2, relax_weight[fine_grid], omega[fine_grid], l1_norms[fine_grid], U_array[fine_grid], Vtemp, Ztemp); else if (rlx_order) { HYPRE_Int loc_relax_points[2]; loc_relax_points[0] = -1; loc_relax_points[1] = 1; for (i=0; i < 2; i++) hypre_ParCSRRelax_L1_Jacobi(A_array[fine_grid],F_array[fine_grid], CF_marker_array[fine_grid], loc_relax_points[i], 1.0, l1_norms[fine_grid], U_array[fine_grid], Vtemp); } else hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid], 1, 1, l1_norms[fine_grid], 1.0, 1.0 ,0,0,0,0, U_array[fine_grid], Vtemp, Ztemp); } else /* additive version */ { alpha = 1.0; beta = 1.0; hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid], U_array[coarse_grid], beta, U_array[fine_grid]); } } return(Solve_err_flag); }