485 lines
21 KiB
C
485 lines
21 KiB
C
|
#ifndef MODEL_UPDATE_H
|
||
|
|
#define MODEL_UPDATE_H
|
||
|
|
|
||
|
|
#include <cmath>
|
||
|
|
#include "config.h"
|
||
|
|
#include "grid.h"
|
||
|
|
#include "smooth.h"
|
||
|
|
#include "smooth_descent_dir.h"
|
||
|
|
#include "smooth_grad_regul.h"
|
||
|
|
#include "lbfgs.h"
|
||
|
|
|
||
|
|
|
||
|
|
// generate smoothed kernels (K*_update_loc) from the kernels (K*_loc)
|
||
|
|
// before doing this, K*_loc should be summed up among all simultaneous runs (by calling sumup_kernels)
|
||
|
|
// before doing this, K*_update_loc has no meaning (unavailable)
|
||
|
|
void smooth_kernels(Grid& grid, InputParams& IP) {
|
||
|
|
|
||
|
|
if (subdom_main){ // parallel level 3
|
||
|
|
|
||
|
|
if (id_sim==0){ // parallel level 1
|
||
|
|
// initiaize update params
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
grid.Ks_update_loc_previous[I2V(i,j,k)] = grid.Ks_update_loc[I2V(i,j,k)];
|
||
|
|
grid.Keta_update_loc_previous[I2V(i,j,k)] = grid.Keta_update_loc[I2V(i,j,k)];
|
||
|
|
grid.Kxi_update_loc_previous[I2V(i,j,k)] = grid.Kxi_update_loc[I2V(i,j,k)];
|
||
|
|
grid.Ks_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
// grid.Ks_density_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
// grid.Kxi_density_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
// grid.Keta_density_update_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// check kernel
|
||
|
|
CUSTOMREAL max_kernel = _0_CR;
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
max_kernel = std::max(max_kernel, std::abs(grid.Ks_loc[I2V(i,j,k)]));
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
CUSTOMREAL tmp = 0;
|
||
|
|
allreduce_cr_single_max(max_kernel, tmp);
|
||
|
|
max_kernel = tmp;
|
||
|
|
|
||
|
|
if (max_kernel <= eps) {
|
||
|
|
std::cout << "Error: max_kernel is near zero (less than 10^-12), check data residual and whether no data is used" << std::endl;
|
||
|
|
exit(1);
|
||
|
|
}
|
||
|
|
|
||
|
|
if (smooth_method == 0) {
|
||
|
|
// Ks_loc, Keta_loc, Kxi_loc to be 0 for ghost layers to eliminate the effect of overlapping boundary
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
|
||
|
|
if (i == 0 && !grid.i_first()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
if (i == loc_I-1 && !grid.i_last()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
if (j == 0 && !grid.j_first()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
if (j == loc_J-1 && !grid.j_last()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
if (k == 0 && !grid.k_first()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
if (k == loc_K-1 && !grid.k_last()) {
|
||
|
|
grid.Ks_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Ks_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Kxi_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
grid.Keta_density_loc[I2V(i,j,k)] = _0_CR;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// grid based smoothing
|
||
|
|
smooth_inv_kernels_orig(grid, IP);
|
||
|
|
} else if (smooth_method == 1) {
|
||
|
|
// CG smoothing
|
||
|
|
smooth_inv_kernels_CG(grid, smooth_lr, smooth_lt, smooth_lp);
|
||
|
|
}
|
||
|
|
|
||
|
|
// shared values on the boundary
|
||
|
|
grid.send_recev_boundary_data(grid.Ks_update_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Keta_update_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Kxi_update_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Ks_density_update_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Kxi_density_update_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Keta_density_update_loc);
|
||
|
|
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Ks_update_loc);
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Keta_update_loc);
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Kxi_update_loc);
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Ks_density_update_loc);
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Kxi_density_update_loc);
|
||
|
|
grid.send_recev_boundary_data_kosumi(grid.Keta_density_update_loc);
|
||
|
|
|
||
|
|
} // end if id_sim == 0
|
||
|
|
|
||
|
|
// send the updated model to all the simultaneous run
|
||
|
|
broadcast_cr_inter_sim(grid.Ks_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Kxi_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Keta_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Ks_density_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Kxi_density_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Keta_density_update_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
|
||
|
|
// send the previous updated model to all the simultaneous run
|
||
|
|
broadcast_cr_inter_sim(grid.Ks_update_loc_previous, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Kxi_update_loc_previous, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Keta_update_loc_previous, loc_I*loc_J*loc_K, 0);
|
||
|
|
|
||
|
|
} // end if subdom_main
|
||
|
|
|
||
|
|
synchronize_all_world();
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
// smooth gradient regularization term
|
||
|
|
void smooth_gradient_regularization(Grid& grid) {
|
||
|
|
|
||
|
|
if (subdom_main && id_sim==0){ // only id_sim==0 has values for these arrays
|
||
|
|
if (smooth_method == 0) {
|
||
|
|
// grid based smoothing
|
||
|
|
smooth_gradient_regularization_orig(grid);
|
||
|
|
} else if (smooth_method == 1) {
|
||
|
|
// CG smoothing
|
||
|
|
smooth_gradient_regularization_CG(grid, smooth_lr, smooth_lt, smooth_lp);
|
||
|
|
}
|
||
|
|
|
||
|
|
grid.send_recev_boundary_data(grid.fun_gradient_regularization_penalty_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.eta_gradient_regularization_penalty_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.xi_gradient_regularization_penalty_loc);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
void smooth_descent_direction(Grid& grid){
|
||
|
|
if (subdom_main){
|
||
|
|
if (id_sim == 0) { // calculation of the update model is only done in the main simultaneous run
|
||
|
|
if (smooth_method == 0) {
|
||
|
|
// grid based smoothing
|
||
|
|
smooth_descent_dir(grid);
|
||
|
|
} else if (smooth_method == 1) {
|
||
|
|
// CG smoothing
|
||
|
|
smooth_descent_dir_CG(grid, smooth_lr, smooth_lt, smooth_lp);
|
||
|
|
}
|
||
|
|
|
||
|
|
// shared values on the boundary
|
||
|
|
grid.send_recev_boundary_data( grid.Ks_descent_dir_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.Keta_descent_dir_loc);
|
||
|
|
grid.send_recev_boundary_data( grid.Kxi_descent_dir_loc);
|
||
|
|
|
||
|
|
} // end if id_sim == 0
|
||
|
|
|
||
|
|
// send the updated model to all the simultaneous run
|
||
|
|
broadcast_cr_inter_sim( grid.Ks_descent_dir_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim( grid.Kxi_descent_dir_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
broadcast_cr_inter_sim(grid.Keta_descent_dir_loc, loc_I*loc_J*loc_K, 0);
|
||
|
|
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
void calc_descent_direction(Grid& grid, int i_inv, InputParams& IP) {
|
||
|
|
|
||
|
|
if (subdom_main) {
|
||
|
|
|
||
|
|
// routines for LBFGS
|
||
|
|
if (optim_method == LBFGS_MODE) {
|
||
|
|
|
||
|
|
// calculate the descent direction
|
||
|
|
if (i_inv > 0) {
|
||
|
|
calculate_descent_direction_lbfgs(grid, i_inv);
|
||
|
|
// use gradient for the first iteration
|
||
|
|
} else {
|
||
|
|
int n_grid = loc_I*loc_J*loc_K;
|
||
|
|
|
||
|
|
// first time, descent direction = - precond * gradient
|
||
|
|
// inverse the gradient to fit the update scheme for LBFGS
|
||
|
|
for (int i = 0; i < n_grid; i++){
|
||
|
|
grid.Ks_descent_dir_loc[i] = - _1_CR* grid.Ks_update_loc[i];
|
||
|
|
grid.Keta_descent_dir_loc[i] = - _1_CR* grid.Keta_update_loc[i];
|
||
|
|
grid.Kxi_descent_dir_loc[i] = - _1_CR* grid.Kxi_update_loc[i];
|
||
|
|
//grid.Ks_descent_dir_loc[i] = - _1_CR* grid.Ks_loc[i];
|
||
|
|
//grid.Keta_descent_dir_loc[i] = - _1_CR* grid.Keta_loc[i];
|
||
|
|
//grid.Kxi_descent_dir_loc[i] = - _1_CR* grid.Kxi_loc[i];
|
||
|
|
|
||
|
|
}
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
// return error
|
||
|
|
std::cout << "Error: optim_method is not set to LBFGS_MODE (=2)" << std::endl;
|
||
|
|
exit(1);
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
} // end if subdom_main
|
||
|
|
|
||
|
|
synchronize_all_world();
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
CUSTOMREAL direction_change_of_model_update(Grid& grid){
|
||
|
|
CUSTOMREAL norm_grad = _0_CR;
|
||
|
|
CUSTOMREAL norm_grad_previous = _0_CR;
|
||
|
|
CUSTOMREAL inner_product = _0_CR;
|
||
|
|
CUSTOMREAL cos_angle = _0_CR;
|
||
|
|
CUSTOMREAL angle = _0_CR;
|
||
|
|
if (subdom_main) {
|
||
|
|
// initiaize update params
|
||
|
|
inner_product = dot_product(grid.Ks_update_loc_previous, grid.Ks_update_loc, loc_I*loc_J*loc_K);
|
||
|
|
norm_grad = dot_product(grid.Ks_update_loc, grid.Ks_update_loc, loc_I*loc_J*loc_K);
|
||
|
|
norm_grad_previous = dot_product(grid.Ks_update_loc_previous, grid.Ks_update_loc_previous, loc_I*loc_J*loc_K);
|
||
|
|
|
||
|
|
CUSTOMREAL tmp;
|
||
|
|
allreduce_cr_single(inner_product,tmp);
|
||
|
|
inner_product = tmp;
|
||
|
|
|
||
|
|
allreduce_cr_single(norm_grad,tmp);
|
||
|
|
norm_grad = tmp;
|
||
|
|
|
||
|
|
allreduce_cr_single(norm_grad_previous,tmp);
|
||
|
|
norm_grad_previous = tmp;
|
||
|
|
|
||
|
|
cos_angle = inner_product / (std::sqrt(norm_grad) * std::sqrt(norm_grad_previous));
|
||
|
|
angle = acos(cos_angle) * RAD2DEG;
|
||
|
|
}
|
||
|
|
|
||
|
|
return angle;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
void set_new_model(Grid& grid, CUSTOMREAL step_length_new, bool init_bfgs=false) {
|
||
|
|
|
||
|
|
if (subdom_main) {
|
||
|
|
|
||
|
|
// for LBFGS mode. K*_update_loc is not directly the descent direction but smoothed gradient
|
||
|
|
// so for non LBFGS_mode, nextstep will be calculated with K*_update_loc
|
||
|
|
if (optim_method != LBFGS_MODE) {
|
||
|
|
// // get the scaling factor
|
||
|
|
// CUSTOMREAL Linf_Ks = _0_CR, Linf_Keta = _0_CR, Linf_Kxi = _0_CR;
|
||
|
|
// for (int k = 1; k < loc_K-1; k++) {
|
||
|
|
// for (int j = 1; j < loc_J-1; j++) {
|
||
|
|
// for (int i = 1; i < loc_I-1; i++) {
|
||
|
|
// Linf_Ks = std::max(Linf_Ks, std::abs(grid.Ks_update_loc[I2V(i,j,k)]));
|
||
|
|
// Linf_Keta = std::max(Linf_Keta, std::abs(grid.Keta_update_loc[I2V(i,j,k)]));
|
||
|
|
// Linf_Kxi = std::max(Linf_Kxi, std::abs(grid.Kxi_update_loc[I2V(i,j,k)]));
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
|
||
|
|
// // get the maximum scaling factor among subdomains
|
||
|
|
// CUSTOMREAL Linf_tmp;
|
||
|
|
// allreduce_cr_single_max(Linf_Ks, Linf_tmp); Linf_Ks = Linf_tmp;
|
||
|
|
// allreduce_cr_single_max(Linf_Keta, Linf_tmp); Linf_Keta = Linf_tmp;
|
||
|
|
// allreduce_cr_single_max(Linf_Kxi, Linf_tmp); Linf_Kxi = Linf_tmp;
|
||
|
|
|
||
|
|
// CUSTOMREAL Linf_all = _0_CR;
|
||
|
|
// Linf_all = std::max(Linf_Ks, std::max(Linf_Keta, Linf_Kxi));
|
||
|
|
|
||
|
|
// // if (myrank == 0 && id_sim == 0)
|
||
|
|
// // std::cout << "Scaling factor for all kernels: " << Linf_all << std::endl;
|
||
|
|
// // std::cout << "Scaling factor for model update for Ks, Keta, Kx, stepsize: " << Linf_Ks << ", " << Linf_Keta << ", " << Linf_Kxi << ", " << step_length_new << std::endl;
|
||
|
|
|
||
|
|
|
||
|
|
// Linf_Ks = Linf_all;
|
||
|
|
// Linf_Keta = Linf_all;
|
||
|
|
// Linf_Kxi = Linf_all;
|
||
|
|
|
||
|
|
// kernel update has been rescaled in "smooth_inv_kernels_orig"
|
||
|
|
|
||
|
|
// update the model
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
|
||
|
|
// // update
|
||
|
|
// grid.fun_loc[I2V(i,j,k)] *= (_1_CR - grid.Ks_update_loc[I2V(i,j,k) ] / (Linf_Ks /step_length_new) );
|
||
|
|
// grid.xi_loc[I2V(i,j,k)] -= grid.Kxi_update_loc[I2V(i,j,k) ] / (Linf_Kxi /step_length_new) ;
|
||
|
|
// grid.eta_loc[I2V(i,j,k)] -= grid.Keta_update_loc[I2V(i,j,k)] / (Linf_Keta /step_length_new) ;
|
||
|
|
|
||
|
|
// update
|
||
|
|
grid.fun_loc[I2V(i,j,k)] *= (_1_CR - grid.Ks_update_loc[I2V(i,j,k) ] * step_length_new);
|
||
|
|
grid.xi_loc[I2V(i,j,k)] -= grid.Kxi_update_loc[I2V(i,j,k) ] * step_length_new;
|
||
|
|
grid.eta_loc[I2V(i,j,k)] -= grid.Keta_update_loc[I2V(i,j,k)] * step_length_new;
|
||
|
|
|
||
|
|
|
||
|
|
// grid.fac_b_loc[I2V(i,j,k)] = _1_CR - _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
// grid.fac_c_loc[I2V(i,j,k)] = _1_CR + _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
// grid.fac_f_loc[I2V(i,j,k)] = - _2_CR * grid.eta_loc[I2V(i,j,k)];
|
||
|
|
|
||
|
|
// grid.fun_loc[I2V(i,j,k)] *= (_1_CR);
|
||
|
|
// grid.xi_loc[I2V(i,j,k)] -= (_0_CR);
|
||
|
|
// grid.eta_loc[I2V(i,j,k)] -= (_0_CR);
|
||
|
|
|
||
|
|
// grid.fac_b_loc[I2V(i,j,k)] = _1_CR - _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
// grid.fac_c_loc[I2V(i,j,k)] = _1_CR + _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
// grid.fac_f_loc[I2V(i,j,k)] = - _2_CR * grid.eta_loc[I2V(i,j,k)];
|
||
|
|
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
grid.rejuvenate_abcf();
|
||
|
|
|
||
|
|
|
||
|
|
} else { // for LBFGS routine
|
||
|
|
|
||
|
|
// here all the simultaneous runs have the same values used in this routine.
|
||
|
|
// thus we don't need to if(id_sim==0)
|
||
|
|
|
||
|
|
CUSTOMREAL step_length = step_length_new;
|
||
|
|
const CUSTOMREAL factor = - _1_CR;
|
||
|
|
|
||
|
|
// update the model
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
// update
|
||
|
|
grid.fun_loc[I2V(i,j,k)] *= (_1_CR - factor * grid.Ks_descent_dir_loc[I2V(i,j,k)] * step_length);
|
||
|
|
grid.xi_loc[I2V(i,j,k)] -= factor * grid.Kxi_descent_dir_loc[I2V(i,j,k) ] * step_length;
|
||
|
|
grid.eta_loc[I2V(i,j,k)] -= factor * grid.Keta_descent_dir_loc[I2V(i,j,k)] * step_length;
|
||
|
|
//grid.fun_loc[I2V(i,j,k)] += grid.Ks_descent_dir_loc[I2V(i,j,k)] *step_length_new;
|
||
|
|
//grid.xi_loc[I2V(i,j,k)] += grid.Kxi_descent_dir_loc[I2V(i,j,k)] *step_length_new;
|
||
|
|
//grid.eta_loc[I2V(i,j,k)] += grid.Keta_descent_dir_loc[I2V(i,j,k)]*step_length_new;
|
||
|
|
|
||
|
|
grid.fac_b_loc[I2V(i,j,k)] = _1_CR - _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
grid.fac_c_loc[I2V(i,j,k)] = _1_CR + _2_CR * grid.xi_loc[I2V(i,j,k)];
|
||
|
|
grid.fac_f_loc[I2V(i,j,k)] = - _2_CR * grid.eta_loc[I2V(i,j,k)];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
} // end LBFGS model update
|
||
|
|
|
||
|
|
// shared values on the boundary
|
||
|
|
grid.send_recev_boundary_data(grid.fun_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.xi_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.eta_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.fac_b_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.fac_c_loc);
|
||
|
|
grid.send_recev_boundary_data(grid.fac_f_loc);
|
||
|
|
|
||
|
|
} // end if subdom_main
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
// compute p_k * grad(f_k)
|
||
|
|
CUSTOMREAL compute_q_k(Grid& grid) {
|
||
|
|
|
||
|
|
CUSTOMREAL tmp_qk = _0_CR;
|
||
|
|
|
||
|
|
if (subdom_main && id_sim == 0) {
|
||
|
|
|
||
|
|
// grad * descent_direction
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
// we can add here a precondition (here use *_update_loc for storing descent direction)
|
||
|
|
tmp_qk += grid.Ks_update_loc[I2V(i,j,k)] * grid.Ks_descent_dir_loc[I2V(i,j,k)]
|
||
|
|
+ grid.Keta_update_loc[I2V(i,j,k)] * grid.Keta_descent_dir_loc[I2V(i,j,k)]
|
||
|
|
+ grid.Kxi_update_loc[I2V(i,j,k)] * grid.Kxi_descent_dir_loc[I2V(i,j,k)];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// add tmp_qk among all subdomain
|
||
|
|
allreduce_cr_single(tmp_qk, tmp_qk);
|
||
|
|
}
|
||
|
|
|
||
|
|
// share tmp_qk among all simultaneous runs
|
||
|
|
if (subdom_main)
|
||
|
|
broadcast_cr_single_inter_sim(tmp_qk,0);
|
||
|
|
|
||
|
|
return tmp_qk;
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
// check if the wolfe conditions are satisfied
|
||
|
|
bool check_wolfe_cond(Grid& grid, \
|
||
|
|
CUSTOMREAL q_0, CUSTOMREAL q_t, \
|
||
|
|
CUSTOMREAL qp_0, CUSTOMREAL qp_t, \
|
||
|
|
CUSTOMREAL& td, CUSTOMREAL& tg, CUSTOMREAL& step_length_sub) {
|
||
|
|
/*
|
||
|
|
q_0 : initial cost function
|
||
|
|
q_t : current cost function
|
||
|
|
q_p0 : initial grad * descent_dir
|
||
|
|
q_pt : current grad * descent_dir
|
||
|
|
td : right step size
|
||
|
|
tg : left step size
|
||
|
|
step_length_sub : current step size
|
||
|
|
*/
|
||
|
|
|
||
|
|
bool step_accepted = false;
|
||
|
|
|
||
|
|
// check if the wolfe conditions are satisfied and update the step_length_sub
|
||
|
|
CUSTOMREAL qpt = (q_t - q_0) / step_length_sub;
|
||
|
|
|
||
|
|
if (subdom_main) {
|
||
|
|
// good step size
|
||
|
|
if (qpt <= wolfe_c1*qp_0 \
|
||
|
|
&& wolfe_c2*qp_0 <= qp_t \
|
||
|
|
){
|
||
|
|
if (myrank==0)
|
||
|
|
std::cout << "Wolfe rules: step accepted" << std::endl;
|
||
|
|
step_accepted = true;
|
||
|
|
} else {
|
||
|
|
// modify the stepsize
|
||
|
|
if (wolfe_c1*qp_0 < qpt) {
|
||
|
|
td = step_length_sub;
|
||
|
|
if (myrank==0)
|
||
|
|
std::cout << "Wolfe rules: right step size updated." << std::endl;
|
||
|
|
}
|
||
|
|
if (qpt <= wolfe_c1*qp_0 && qp_t < wolfe_c2*qp_0) {
|
||
|
|
tg = step_length_sub;
|
||
|
|
if (myrank==0)
|
||
|
|
std::cout << "Wolfe rules: left step size updated." << std::endl;
|
||
|
|
}
|
||
|
|
//if (isZero(td)) {
|
||
|
|
if (td == _0_CR) {
|
||
|
|
step_length_sub = _2_CR * step_length_sub;
|
||
|
|
if (myrank==0)
|
||
|
|
std::cout << "Wolfe rules: step size too small. Increased to " << step_length_sub << std::endl;
|
||
|
|
} else {
|
||
|
|
step_length_sub = _0_5_CR * (td+tg);
|
||
|
|
if (myrank==0)
|
||
|
|
std::cout << "Wolfe rules: step size too large. Decreased to " << step_length_sub << std::endl;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// share step_accepted among all subdomains
|
||
|
|
broadcast_bool_single(step_accepted,0);
|
||
|
|
broadcast_cr_single(step_length_sub,0);
|
||
|
|
} // end if subdom_main
|
||
|
|
|
||
|
|
broadcast_bool_single_sub(step_accepted,0);
|
||
|
|
broadcast_cr_single_sub(step_length_sub,0);
|
||
|
|
|
||
|
|
return step_accepted;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
#endif // MODEL_UPDATE_H
|