200 lines
6.6 KiB
C
200 lines
6.6 KiB
C
|
#ifndef SMOOTH_H
|
||
|
|
#define SMOOTH_H
|
||
|
|
|
||
|
|
#include <iostream>
|
||
|
|
#include "../../include/config.h"
|
||
|
|
#include "../../include/utils.h"
|
||
|
|
|
||
|
|
|
||
|
|
void calc_inversed_laplacian(CUSTOMREAL* d, CUSTOMREAL* Ap,
|
||
|
|
const int i, const int j, const int k,
|
||
|
|
const CUSTOMREAL lr, const CUSTOMREAL lt, const CUSTOMREAL lp,
|
||
|
|
const CUSTOMREAL dr, const CUSTOMREAL dt, const CUSTOMREAL dp) {
|
||
|
|
// calculate inversed laplacian operator
|
||
|
|
CUSTOMREAL termx = _0_CR, termy = _0_CR, termz = _0_CR;
|
||
|
|
|
||
|
|
if (i==0) {
|
||
|
|
termx = dp*lp/3.0 * (1/(lp*dp)*(d[I2V(i,j,k)]) - lp/(dp*dp*dp)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i+1,j,k)]));
|
||
|
|
} else if (i==loc_I-1) {
|
||
|
|
termx = dp*lp/3.0 * (1/(lp*dp)*(d[I2V(i,j,k)]) - lp/(dp*dp*dp)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i-1,j,k)]));
|
||
|
|
} else {
|
||
|
|
termx = dp*lp/3.0 * (1/(lp*dp)*(d[I2V(i,j,k)]) - lp/(dp*dp*dp)*(-2.0*d[I2V(i,j,k)]+d[I2V(i-1,j,k)]+d[I2V(i+1,j,k)]));
|
||
|
|
}
|
||
|
|
|
||
|
|
if (j==0) {
|
||
|
|
termy = dt*lt/3.0 * (1/(lt*dt)*(d[I2V(i,j,k)]) - lt/(dt*dt*dt)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i,j+1,k)]));
|
||
|
|
} else if (j==loc_J-1) {
|
||
|
|
termy = dt*lt/3.0 * (1/(lt*dt)*(d[I2V(i,j,k)]) - lt/(dt*dt*dt)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i,j-1,k)]));
|
||
|
|
} else {
|
||
|
|
termy = dt*lt/3.0 * (1/(lt*dt)*(d[I2V(i,j,k)]) - lt/(dt*dt*dt)*(-2.0*d[I2V(i,j,k)]+d[I2V(i,j-1,k)]+d[I2V(i,j+1,k)]));
|
||
|
|
}
|
||
|
|
|
||
|
|
if (k==0) {
|
||
|
|
termz = dr*lr/3.0 * (1/(lr*dr)*(d[I2V(i,j,k)]) - lr/(dr*dr*dr)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i,j,k+1)]));
|
||
|
|
} else if (k==loc_K-1) {
|
||
|
|
termz = dr*lr/3.0 * (1/(lr*dr)*(d[I2V(i,j,k)]) - lr/(dr*dr*dr)*(-2.0*d[I2V(i,j,k)]+2*d[I2V(i,j,k-1)]));
|
||
|
|
} else {
|
||
|
|
termz = dr*lr/3.0 * (1/(lr*dr)*(d[I2V(i,j,k)]) - lr/(dr*dr*dr)*(-2.0*d[I2V(i,j,k)]+d[I2V(i,j,k-1)]+d[I2V(i,j,k+1)]));
|
||
|
|
}
|
||
|
|
|
||
|
|
Ap[I2V(i,j,k)] = termx+termy+termz;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
void CG_smooth(CUSTOMREAL* arr_in, CUSTOMREAL* arr_out, CUSTOMREAL lr, CUSTOMREAL lt, CUSTOMREAL lp) {
|
||
|
|
// arr: array to be smoothed
|
||
|
|
// lr: smooth length on r
|
||
|
|
// lt: smooth length on theta
|
||
|
|
// lp: smooth length on phi
|
||
|
|
|
||
|
|
// arrays :
|
||
|
|
// x_array model
|
||
|
|
// g_array gradient
|
||
|
|
// d_array descent direction
|
||
|
|
// Ap stiffness scalar (laplacian)
|
||
|
|
// pAp = d_array*Ap
|
||
|
|
// rr = g_array * g_array (dot product)
|
||
|
|
|
||
|
|
const int max_iter_cg = 1000;
|
||
|
|
const CUSTOMREAL xtol = 0.001;
|
||
|
|
const bool use_scaling = true;
|
||
|
|
|
||
|
|
CUSTOMREAL dr=1, dt=100, dp=100;
|
||
|
|
|
||
|
|
// allocate memory
|
||
|
|
CUSTOMREAL* x_array = new CUSTOMREAL[loc_I*loc_J*loc_K];
|
||
|
|
CUSTOMREAL* r_array = new CUSTOMREAL[loc_I*loc_J*loc_K];
|
||
|
|
CUSTOMREAL* p_array = new CUSTOMREAL[loc_I*loc_J*loc_K];
|
||
|
|
CUSTOMREAL* Ap = new CUSTOMREAL[loc_I*loc_J*loc_K];
|
||
|
|
CUSTOMREAL pAp=_0_CR, rr_0=_0_CR, rr=_0_CR, rr_new=_0_CR, aa=_0_CR, bb=_0_CR, tmp=_0_CR;
|
||
|
|
|
||
|
|
CUSTOMREAL scaling_A=_1_CR, scaling_coeff = _1_CR;
|
||
|
|
|
||
|
|
if (use_scaling) {
|
||
|
|
// calculate scaling factor
|
||
|
|
//scaling_A = std::sqrt(_1_CR / (_8_CR * PI * lr * lt * lp));
|
||
|
|
// scaling coefficient for gradient
|
||
|
|
scaling_coeff = find_absmax(arr_in, loc_I*loc_J*loc_K);
|
||
|
|
//tmp = scaling_coeff;
|
||
|
|
//allreduce_cr_single_max(tmp, scaling_coeff);
|
||
|
|
//if (scaling_coeff == _0_CR)
|
||
|
|
if (isZero(scaling_coeff))
|
||
|
|
scaling_coeff = _1_CR;
|
||
|
|
}
|
||
|
|
// std out scaling factors
|
||
|
|
//if (myrank == 0) {
|
||
|
|
std::cout << "scaling_A = " << scaling_A << std::endl;
|
||
|
|
std::cout << "scaling_coeff = " << scaling_coeff << std::endl;
|
||
|
|
//}
|
||
|
|
|
||
|
|
|
||
|
|
// array initialization
|
||
|
|
for (int i=0; i<loc_I*loc_J*loc_K; i++) {
|
||
|
|
x_array[i] = _0_CR; ///x
|
||
|
|
r_array[i] = arr_in[i]/scaling_coeff; // r
|
||
|
|
p_array[i] = r_array[i]; // p
|
||
|
|
Ap[i] = _0_CR;
|
||
|
|
}
|
||
|
|
|
||
|
|
// initial rr
|
||
|
|
rr = dot_product(r_array, r_array, loc_I*loc_J*loc_K);
|
||
|
|
//tmp = rr;
|
||
|
|
// sum all rr among all processors
|
||
|
|
//allreduce_cr_single(tmp, rr);
|
||
|
|
rr_0 = rr; // record initial rr
|
||
|
|
|
||
|
|
// CG loop
|
||
|
|
for (int iter=0; iter<max_iter_cg; iter++) {
|
||
|
|
|
||
|
|
// calculate laplacian
|
||
|
|
for (int k = 0; k < loc_K; k++) {
|
||
|
|
for (int j = 0; j < loc_J; j++) {
|
||
|
|
for (int i = 0; i < loc_I; i++) {
|
||
|
|
//scaling_coeff = std::max(scaling_coeff, std::abs(arr[I2V(i,j,k)]));
|
||
|
|
|
||
|
|
// calculate inversed laplacian operator
|
||
|
|
calc_inversed_laplacian(p_array,Ap,i,j,k,lr,lt,lp,dr,dt,dp);
|
||
|
|
|
||
|
|
|
||
|
|
//Ap[I2V(i,j,k)] = (
|
||
|
|
// - _2_CR * (lr+lt+lp) * d_array[I2V(i,j,k)] \
|
||
|
|
// + lr * (d_array[I2V(i,j,k+1)] + d_array[I2V(i,j,k-1)]) \
|
||
|
|
// + lt * (d_array[I2V(i,j+1,k)] + d_array[I2V(i,j-1,k)]) \
|
||
|
|
// + lp * (d_array[I2V(i+1,j,k)] + d_array[I2V(i-1,j,k)]) \
|
||
|
|
//);
|
||
|
|
|
||
|
|
// scaling
|
||
|
|
Ap[I2V(i,j,k)] = Ap[I2V(i,j,k)]*scaling_A;
|
||
|
|
//Ap[I2V(i,j,k)] = (p_array[I2V(i,j,k)]*Ap[I2V(i,j,k)])*scaling_A;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
// get the values on the boundaries
|
||
|
|
//grid.send_recev_boundary_data(Ap);
|
||
|
|
|
||
|
|
// calculate pAp
|
||
|
|
pAp = dot_product(p_array, Ap, loc_I*loc_J*loc_K);
|
||
|
|
//tmp = pAp;
|
||
|
|
// sum all pAp among all processors
|
||
|
|
//allreduce_cr_single(tmp, pAp);
|
||
|
|
|
||
|
|
// compute step length
|
||
|
|
aa = rr / pAp;
|
||
|
|
|
||
|
|
// update x_array (model)
|
||
|
|
for (int i=0; i<loc_I*loc_J*loc_K; i++) {
|
||
|
|
x_array[i] += aa * p_array[i];
|
||
|
|
}
|
||
|
|
|
||
|
|
// update r_array (gradient)
|
||
|
|
for (int i=0; i<loc_I*loc_J*loc_K; i++) {
|
||
|
|
r_array[i] -= aa * Ap[i];
|
||
|
|
}
|
||
|
|
|
||
|
|
// update rr
|
||
|
|
rr_new = dot_product(r_array, r_array, loc_I*loc_J*loc_K);
|
||
|
|
//tmp = rr_new;
|
||
|
|
// sum all rr among all processors
|
||
|
|
//allreduce_cr_single(tmp, rr_new);
|
||
|
|
|
||
|
|
// update d_array (descent direction)
|
||
|
|
bb = rr_new / rr;
|
||
|
|
for (int i=0; i<loc_I*loc_J*loc_K; i++) {
|
||
|
|
p_array[i] = r_array[i] + bb * p_array[i];
|
||
|
|
}
|
||
|
|
|
||
|
|
//if (myrank == 0 && iter%100==0){//} && if_verbose) {
|
||
|
|
if (iter%1==0){//} && if_verbose) {
|
||
|
|
std::cout << "iter: " << iter << " rr: " << rr << " rr_new: " << rr_new << " rr/rr_0: " << rr/rr_0 << " pAp: " << pAp << " aa: " << aa << " bb: " << bb << std::endl;
|
||
|
|
}
|
||
|
|
|
||
|
|
// update rr
|
||
|
|
rr = rr_new;
|
||
|
|
|
||
|
|
// check convergence
|
||
|
|
if (rr / rr_0 < xtol) {
|
||
|
|
std::cout << "CG converged in " << iter << " iterations." << std::endl;
|
||
|
|
break;
|
||
|
|
}
|
||
|
|
|
||
|
|
} // end of CG loop
|
||
|
|
|
||
|
|
|
||
|
|
// copy x_array to arr_out
|
||
|
|
for (int i=0; i<loc_I*loc_J*loc_K; i++) {
|
||
|
|
arr_out[i] = x_array[i]*scaling_coeff;
|
||
|
|
}
|
||
|
|
|
||
|
|
// deallocate
|
||
|
|
delete[] x_array;
|
||
|
|
delete[] r_array;
|
||
|
|
delete[] p_array;
|
||
|
|
delete[] Ap;
|
||
|
|
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
#endif
|