Homec4science
Diffusion Lenstool-HPC (develop_chris) Browse

R1448/ChiBenchmark/lenstoolchi/src/d_seeing_omp.c
c99f2e455a51develop_chris

Last Change
Raw File

d_seeing_omp.c
View Options

This document is not UTF8. It was detected as Shift JIS and converted to UTF8 for display.
//d_seeing_omp.c
//
// name of main function: d_seeing_omp
// author: Sergey Rodionov (astroseger@gmail.com)
// date: 01/2012
// place: Marseille
//
//function for fast add seeing to images by means of fft
//principal function is the last function
//ATTENTION!!! WE USE GLOBAL VARIABLES FOR OPTIMISATION REASONS
//so don't remove check_not_in_parallel();
//because historical reason we have nx, ny swapped in respect to lenstool
//here -- im[0..nx][0..ny] ... in lenstool im[0..ny][0..nx]
#include <gsl/gsl_fft_complex.h>
#include <stdio.h>
#include <stdlib.h>
#include "structure.h"
#include "lt.h"
struct cpa2d; //predifiniton of cpa2d (see below)
//ATTENTION!!!! GLOBAL VARIABLES
//fourier transformation of psf
//psf_f->nx, psf_f->ny --- sizes of our image!
static struct cpa2d* psf_f = NULL;
//preallocated "padded" image
//(we only us it directly from d_seeing_omp)
static struct cpa2d* pim_global = NULL;
//structure which keep parameters for which psf_f as constructed
//(just to check that it haven't been changed)
static struct g_observ O_psf_f;
//nx, ny and scale for first run (in addition to O_psf_f)
static int nx_init, ny_init;
static double scale_init;
//wavetables and workspaces for fourier trasform
// *_x - for psf->nx
// *_y - for psf->ny
//for each row or column --- separeated workspace
static gsl_fft_complex_wavetable * wt_x;
static gsl_fft_complex_wavetable * wt_y;
static gsl_fft_complex_workspace **ws_x; //size psf_f->ny !!! (reverse!)
static gsl_fft_complex_workspace **ws_y; //size psf_f->nx !!! (reserse!)
//gsl "complex packed array" in 2D. For 2D fourier trasform
struct cpa2d
{
double* v;
int nx;
int ny;
};
//
//several basic function for the access to cpa2d, and allocate/free of it
static struct cpa2d* cpa2d_alloc(int nx, int ny)
{
struct cpa2d* c = (struct cpa2d*) malloc(sizeof(struct cpa2d*));
c->v = (double*) malloc(sizeof(double) * nx * ny * 2);
c->nx = nx;
c->ny = ny;
return c;
}
//
static void cpa2d_free(struct cpa2d* c)
{
free(c->v);
free(c);
}
//
inline void cpa2d_set(struct cpa2d* c, int i, int j, double re, double im)
{
c->v[(i * c->ny + j) * 2] = re;
c->v[(i * c->ny + j) * 2 + 1] = im;
}
//
inline void cpa2d_set_re(struct cpa2d* c, int i, int j, double re)
{
c->v[(i * c->ny + j) * 2] = re;
}
//
inline double cpa2d_get_re(struct cpa2d* c, int i, int j)
{
return c->v[(i * c->ny + j) * 2];
}
//
inline double cpa2d_get_im(struct cpa2d* c, int i, int j)
{
return c->v[(i * c->ny + j) * 2 + 1];
}
//
//from im (nx,ny) to pim (pim->nx,pim->ny) with zero padding
static void zero_padding(double **im, int nx, int ny,
struct cpa2d* pim)
{
if (pim->nx <= nx || pim->ny <= ny)
{
fprintf(stderr, "error | zero_paddinig | pim->nx <= nx || pim->ny <= ny\n");
exit(EXIT_FAILURE);
}
int px = (pim->nx - nx) / 2;
int py = (pim->ny - ny) / 2;
int i;
#pragma omp parallel for schedule(static)
for (i = 0 ; i < pim->nx ; i++)
{
int j;
for (j = 0 ; j < pim->ny ; j++)
{
if ((i >= px) && (i < px + nx) &&
(j >= py) && (j < py + ny))
{
cpa2d_set(pim, i, j, im[i - px][j - py], 0);
}
else
{
cpa2d_set(pim, i, j, 0, 0);
}
}
}
}
//from pim to im (unpadding and "unsplitting")
//We take into account "splitting" in pim
//if cntrx, cntry - is position of maximum in psf
//and it is center of "splitting"
static void split_unpadding(double **im, int nx, int ny,
struct cpa2d* pim)
{
int px = (pim->nx - nx) / 2;
int py = (pim->ny - ny) / 2;
//calculate position of center in psf
//for odd center in n/2
//for even center in (n - 0.5)/2
int cntrx = (int)floor((pim->nx - 0.5) / 2.0);
int cntry = (int) floor((pim->ny - 0.5) / 2.0);
int i;
#pragma omp parallel for schedule(static)
for (i = px ; i < px + nx ; i++)
{
int j;
for (j = py ; j < py + ny ; j++)
{
int ip, jp;
ip = i;
jp = j;
//splitting
if (i < pim->nx - cntrx)
ip = i + cntrx;
else
ip = i - (pim->nx - cntrx);
if (j < pim->ny - cntry)
jp = j + cntry;
else
jp = j - (pim->ny - cntry);
//unpadding
im[i - px][j - py] = cpa2d_get_re(pim, ip, jp);
}
}
}
//two simple function for 2D fast fourier trasform (direct and invert)
//ATENTIANON this two function use global variables
static void fft_2D_omp(struct cpa2d* data)
{
int i;
#pragma omp parallel for schedule(static)
for (i = 0 ; i < data->nx ; i++)
{
gsl_fft_complex_forward(data->v + data->ny * i * 2, 1, data->ny,
wt_y, ws_y[i]);
}
#pragma omp parallel for schedule(static)
for (i = 0 ; i < data->ny ; i++)
{
gsl_fft_complex_forward(data->v + i * 2, data->ny,
data->nx, wt_x, ws_x[i]);
}
}
//
static void fft_2D_inv_omp(struct cpa2d* data)
{
int i;
#pragma omp parallel for schedule(static)
for (i = 0 ; i < data->nx ; i++)
{
gsl_fft_complex_inverse(data->v + data->ny * i * 2, 1,
data->ny, wt_y, ws_y[i]);
}
#pragma omp parallel for schedule(static)
for (i = 0 ; i < data->ny ; i++)
{
gsl_fft_complex_inverse(data->v + i * 2, data->ny,
data->nx, wt_x, ws_x[i]);
}
}
// ﾑ
//we choose smallest number greater than n for which gsl fft is fast
//(biggest factor should be equal to 7. see gsl documentation)
static int choose_fast_fft_n(int n)
{
const int primes[4] = {2, 3, 5, 7};
const int primes_n = 4;
int det = 0;
do
{
int tmp = n;
int i;
for (i = 0 ; i < primes_n ; i++)
while (tmp % primes[i] == 0)
tmp /= primes[i];
if (tmp == 1)
det = 1;
else
n++;
}
while (!det);
return n;
}
//
static void create_psf_psffile(double***spsf, int* nx, int* ny)
{
char* header;
const extern struct g_observ O;
//again nx and ny should be swapped here
*spsf = rdf_fits((char*)O.psffile, ny, nx, &header);
free(header);
//check maximum (it should be at nx/2 ny/2)
int max_i = *nx / 2;
int max_j = *ny / 2;
int i,j;
for (i = 0 ; i < *nx ; i++)
for (j = 0 ; j < *ny; j++)
{
if ((*spsf)[i][j] > (*spsf)[max_i][max_j])
{
fprintf(stderr, "Error | d_seeing_omp/create_psf_psffile | maximum is not located in center of psf\n");
exit(EXIT_FAILURE);
}
}
//normalization
long double sum = 0;
for (i = 0 ; i < *nx ; i++)
for (j = 0 ; j < *ny ; j++)
sum += (*spsf)[i][j];
for (i = 0 ; i < *nx ; i++)
for (j = 0 ; j < *ny ; j++)
{
(*spsf)[i][j] /= sum;
}
}
//
static void create_psf_create_filtre(double scale, int big_n, double***spsf, int* nx, int*ny)
{
int i,j;
const double PSF_ZERO = 1e-20;
const extern struct g_observ O;
//create_filtre work correctly only for even n
if (big_n % 2 != 0)
big_n++;
double** psf = (double **) alloc_square_double(big_n, big_n);
if (O.filtre)
{
fprintf(stderr, "Error | d_seeing_omp.c/create_psf | O.filtre != 0 we don't know what it means");
exit(EXIT_FAILURE);
}
if (O.setseeing == 1)
{
crea_filtre(O.seeing, scale, psf, big_n);
}
else if (O.setseeing == 2)
{
crea_filtre_e(O.seeing_a, O.seeing_b, O.seeing_angle, scale, psf, big_n);
}
else
{
fprintf(stderr, "Error | d_seeing_omp.c/create_psf | unknown O.setseeing = %d", O.setseeing);
exit(EXIT_FAILURE);
}
//just in case we check maximum value (it should be in big_n/2)
int max_i = big_n / 2;
for (i = 0 ; i < big_n ; i++)
for (j = 0 ; j < big_n; j++)
if (psf[i][j] > psf[max_i][max_i])
{
fprintf(stderr, "Unexpected behavior of crea_filtre\n");
exit(EXIT_FAILURE);
}
//now we can try to shrink psf
int si = 1;
int sj = 1;
int det = 1;
do
{
for (j = 0 ; j < big_n ; j++)
{
if (psf[si][j] > PSF_ZERO || psf[2 * max_i - si][j] > PSF_ZERO)
det = 0;
}
if (det)
si++;
if (si == max_i)
{
fprintf(stderr, "Unexpected behavior 3 of crea_filtre\n");
exit(EXIT_FAILURE);
}
}
while (det);
det = 1;
do
{
for (i = 0 ; i < big_n ; i++)
{
if (psf[i][sj] > PSF_ZERO || psf[i][2 * max_i - sj] > PSF_ZERO)
det = 0;
}
if (det)
sj++;
if (sj == max_i)
{
fprintf(stderr, "Unexpected behavior 3 of crea_filtre\n");
exit(EXIT_FAILURE);
}
}
while (det);
*nx = (max_i - si) * 2 + 1;
*ny = (max_i - sj) * 2 + 1;
(*spsf) = (double **) alloc_square_double(*nx, *ny);
long double sum = 0;
for (i = 0 ; i < *nx ; i++)
for (j = 0 ; j < *ny ; j++)
{
(*spsf)[i][j] = psf[si + i][sj + j];
sum += (*spsf)[i][j];
}
for (i = 0 ; i < *nx ; i++)
for (j = 0 ; j < *ny ; j++)
{
(*spsf)[i][j] /= sum;
}
free_square_double(psf, big_n);
}
//
static void create_psf(double scale, int big_n, double***spsf, int* nx, int*ny)
{
const extern struct g_observ O;
//we also save parameters for which psf was constructed
O_psf_f = O; //we cannot use O_psf_f.psfile anymore!x
if (O.setseeing == 1 || O.setseeing == 2)
{
create_psf_create_filtre(scale, big_n, spsf, nx, ny);
}
else if (O.setseeing == 3)
{
create_psf_psffile(spsf, nx, ny);
}
else
{
fprintf(stderr, "Error | d_seeing_omp.c/create_psf | unknown O.setseeing = %d", O.setseeing);
exit(EXIT_FAILURE);
}
}
//
//because historical reason we have nx, ny swapped in respect to lenstool
//here -- im[0..nx][0..ny] ... in lenstool im[0..ny][0..nx]
//ATTENTIAON!!! because of these nx ny was swapped here !!!
void d_seeing_omp(double **im, int ny, int nx, double scale)
{
//we must not be in parallel now
check_not_in_parallel("d_seeing_omp");
int i;
if (psf_f == NULL) //first run
{
double** psf;
int nx_psf,ny_psf;
//create psf and save parameters for which it was constructed
create_psf(scale, nx + 1, &psf, &nx_psf, &ny_psf);
//save value of nx and ny of first run
nx_init = nx;
ny_init = ny;
//save scale parameter of first run
scale_init = scale;
//we should add "zero" padding of size at least nx_psf/2 (ny_psf/2)
//this value is exactly equal to size of wings of psf...
//and after we search appropriative size of array for which fft
//could be calculated fast
int nx1 = choose_fast_fft_n(nx + nx_psf / 2);
int ny1 = choose_fast_fft_n(ny + ny_psf / 2);
// printf("nx=%d ny=%d nx_psf=%d ny_psf=%d nx1=%d ny1=%d\n",
// nx, ny, nx_psf, ny_psf, nx1, ny1);
//now we can allocated wavetables and workspaces for fft
wt_x = gsl_fft_complex_wavetable_alloc (nx1);
wt_y = gsl_fft_complex_wavetable_alloc (ny1);
ws_x = (gsl_fft_complex_workspace**)
malloc(sizeof(gsl_fft_complex_workspace*) * ny1);
ws_y = (gsl_fft_complex_workspace**)
malloc(sizeof(gsl_fft_complex_workspace*) * nx1);
for (i = 0 ; i < ny1 ; i++)
ws_x[i] = gsl_fft_complex_workspace_alloc (nx1);
for (i = 0 ; i < nx1 ; i++)
ws_y[i] = gsl_fft_complex_workspace_alloc (ny1);
//calculate fft of psf
psf_f = cpa2d_alloc(nx1, ny1);
zero_padding(psf, nx_psf, ny_psf, psf_f);
fft_2D_omp(psf_f);
//and finally allocated padded image with which we will work
pim_global = cpa2d_alloc(psf_f->nx, psf_f->ny);
//real psf we will not need anymore
free_square_double(psf, nx_psf);
}
else
{
//we just check that noting have changed after initialisation
const extern struct g_observ O;
if (nx != nx_init || ny != ny_init || scale != scale_init)
{
fprintf(stderr, "Error | d_seeing_omp.c/d_seeing_omp | nx, ny or scale have been changed from first run");
exit(EXIT_FAILURE);
}
if (O.setseeing != O_psf_f.setseeing ||
O.seeing != O_psf_f.seeing ||
O.seeing_a != O_psf_f.seeing_a ||
O.seeing_b != O_psf_f.seeing_b ||
O.seeing_angle != O_psf_f.seeing_angle )
{
fprintf(stderr, "Error | d_seeing_omp.c/d_seeing_omp | O structure have been changed from first run");
exit(EXIT_FAILURE);
}
}
//Now we can start main part of algorim
//psf_f->nx and psf_f->ny keep dimension of padded image
struct cpa2d* pim = pim_global; //just for simplicity
zero_padding(im, nx, ny, pim);
fft_2D_omp(pim);
//simple multiplication of pim and psf_f
//(see basic theory of convolution)
#pragma omp parallel for schedule(static)
for (i = 0 ; i < pim->nx * pim->ny * 2 ; i += 2)
{
double rr = pim->v[i] * psf_f->v[i] - pim->v[i + 1] * psf_f->v[i + 1];
double ii = pim->v[i] * psf_f->v[i + 1] + pim->v[i + 1] * psf_f->v[i];
pim->v[i] = rr;
pim->v[i + 1] = ii;
}
fft_2D_inv_omp(pim);
split_unpadding(im, nx, ny, pim);
}
c4science · Help