dist_rho5D.c

Go to the documentation of this file.

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "../ascot5.h"
#include "../consts.h"
#include "../physlib.h"
#include "dist_rho5D.h"
#include "../particle.h"

size_t dist_rho5D_index(int i_rho, int i_theta, int i_phi, int i_ppara,
                        int i_pperp, int i_time, int i_q, size_t step_6,
                        size_t step_5, size_t step_4, size_t step_3,
                        size_t step_2, size_t step_1) {
    return (size_t)(i_rho)   * step_6
         + (size_t)(i_theta) * step_5
         + (size_t)(i_phi)   * step_4
         + (size_t)(i_ppara) * step_3
         + (size_t)(i_pperp) * step_2
         + (size_t)(i_time)  * step_1
         + (size_t)(i_q);
}

int dist_rho5D_init(dist_rho5D_data* data) {
 
    size_t n_q     = (size_t)(data->n_q);
    size_t n_time  = (size_t)(data->n_time);
    size_t n_pperp = (size_t)(data->n_pperp);
    size_t n_ppara = (size_t)(data->n_ppara);
    size_t n_phi   = (size_t)(data->n_phi);
    size_t n_theta = (size_t)(data->n_theta);
    data->step_6 = n_q * n_time * n_pperp * n_ppara * n_phi * n_theta;
    data->step_5 = n_q * n_time * n_pperp * n_ppara * n_phi;
    data->step_4 = n_q * n_time * n_pperp * n_ppara;
    data->step_3 = n_q * n_time * n_pperp;
    data->step_2 = n_q * n_time;
    data->step_1 = n_q;
 
    data->histogram = calloc(data->step_6 * (size_t)data->n_rho, sizeof(real));
    return data->histogram == NULL;
}

void dist_rho5D_free(dist_rho5D_data* data) {
    free(data->histogram);
}

void dist_rho5D_offload(dist_rho5D_data* data) {
    GPU_MAP_TO_DEVICE(
        data->histogram[0:data->n_rho*data->n_theta*data->n_phi*data->n_ppara*data->n_pperp*data->n_time*data->n_q]
    )
}

void dist_rho5D_onload(dist_rho5D_data* data) {
    GPU_UPDATE_FROM_DEVICE(
        data->histogram[0:data->n_rho*data->n_theta*data->n_phi*data->n_ppara*data->n_pperp*data->n_time*data->n_q]
        )
}

void dist_rho5D_update_fo(dist_rho5D_data* dist, particle_simd_fo* p_f,
                          particle_simd_fo* p_i) {
 
#ifdef GPU
    size_t index;
    real weight;
#else
    size_t index[NSIMD];
    real weight[NSIMD];
    int valid[NSIMD] = {0};
#endif
 
    GPU_PARALLEL_LOOP_ALL_LEVELS
    for(int i = 0; i < p_f->n_mrk; i++) {
        if(p_f->running[i]) {
 
            int i_rho = floor((p_f->rho[i] - dist->min_rho)
                             / ((dist->max_rho - dist->min_rho)/dist->n_rho));
 
            real phi = fmod(p_f->phi[i], 2*CONST_PI);
            if(phi < 0) {
                phi += 2*CONST_PI;
            }
            int i_phi = floor((phi - dist->min_phi)
                             / ((dist->max_phi - dist->min_phi)/dist->n_phi));
 
            real theta = fmod(p_f->theta[i], 2*CONST_PI);
            if(theta < 0) {
                theta += 2*CONST_PI;
            }
            int i_theta = floor((theta - dist->min_theta)
                               / ( (dist->max_theta - dist->min_theta)
                                   / dist->n_theta) );
 
            real ppara = (  p_f->p_r[i]   * p_f->B_r[i]
                        + p_f->p_phi[i] * p_f->B_phi[i]
                        + p_f->p_z[i]   * p_f->B_z[i])
                       / sqrt(  p_f->B_r[i]  * p_f->B_r[i]
                              + p_f->B_phi[i]* p_f->B_phi[i]
                              + p_f->B_z[i]  * p_f->B_z[i]);
            int i_ppara = floor((ppara - dist->min_ppara)
                               / ((dist->max_ppara - dist->min_ppara)
                                  / dist->n_ppara));
 
            real pperp = sqrt(
                  p_f->p_r[i]   * p_f->p_r[i]
                + p_f->p_phi[i] * p_f->p_phi[i]
                + p_f->p_z[i]   * p_f->p_z[i]
                - ppara * ppara);
            int i_pperp = floor((pperp - dist->min_pperp)
                               / ((dist->max_pperp - dist->min_pperp)
                                  / dist->n_pperp));
 
            int i_time = floor((p_f->time[i] - dist->min_time)
                          / ((dist->max_time - dist->min_time) / dist->n_time));
 
            int i_q = floor((p_f->charge[i]/CONST_E - dist->min_q)
                           / ((dist->max_q - dist->min_q) / dist->n_q));
 
            if(i_rho   >= 0 && i_rho   <= dist->n_rho - 1   &&
               i_phi   >= 0 && i_phi   <= dist->n_phi - 1   &&
               i_theta >= 0 && i_theta <= dist->n_theta - 1 &&
               i_ppara >= 0 && i_ppara <= dist->n_ppara - 1 &&
               i_pperp >= 0 && i_pperp <= dist->n_pperp - 1 &&
               i_time  >= 0 && i_time  <= dist->n_time - 1  &&
               i_q     >= 0 && i_q     <= dist->n_q - 1       ) {
#ifdef GPU
                index = dist_rho5D_index(
                    i_rho, i_theta, i_phi, i_ppara, i_pperp,
                    i_time, i_q, dist->step_6, dist->step_5, dist->step_4,
                    dist->step_3, dist->step_2, dist->step_1);
                weight = p_f->weight[i] * (p_f->time[i] - p_i->time[i]);
                GPU_ATOMIC
                dist->histogram[index] += weight;
#else
                index[i] = dist_rho5D_index(
                    i_rho, i_theta, i_phi, i_ppara, i_pperp,
                    i_time, i_q, dist->step_6, dist->step_5, dist->step_4,
                    dist->step_3, dist->step_2, dist->step_1);
                weight[i] = p_f->weight[i] * (p_f->time[i] - p_i->time[i]);
                valid[i] = 1;
#endif
            }
        }
    }
#ifndef GPU
    for(int i = 0; i < p_f->n_mrk; i++) {
        if(p_f->running[i] && valid[i] == 1) {
            GPU_ATOMIC
            dist->histogram[index[i]] += weight[i];
        }
    }
#endif
}

void dist_rho5D_update_gc(dist_rho5D_data* dist, particle_simd_gc* p_f,
                          particle_simd_gc* p_i) {
 
    GPU_PARALLEL_LOOP_ALL_LEVELS
    for(int i = 0; i < p_f->n_mrk; i++) {
        if(p_f->running[i]) {
 
            int i_rho = floor((p_f->rho[i] - dist->min_rho)
                             / ((dist->max_rho - dist->min_rho)/dist->n_rho));
 
            real phi = fmod(p_f->phi[i], 2*CONST_PI);
            if(phi < 0) {
                phi = phi + 2*CONST_PI;
            }
            int i_phi = floor((phi - dist->min_phi)
                             / ((dist->max_phi - dist->min_phi)/dist->n_phi));
 
            real theta = fmod(p_f->theta[i], 2*CONST_PI);
            if(theta < 0) {
                theta = theta + 2*CONST_PI;
            }
            int i_theta = floor((theta - dist->min_theta)
                               / ((dist->max_theta - dist->min_theta)
                                  / dist->n_theta));
 
            int i_ppara = floor((p_f->ppar[i] - dist->min_ppara)
                       / ((dist->max_ppara - dist->min_ppara) / dist->n_ppara));
 
            real pperp = sqrt(2 * sqrt(  p_f->B_r[i]   * p_f->B_r[i]
                                     + p_f->B_phi[i] * p_f->B_phi[i]
                                     + p_f->B_z[i]   * p_f->B_z[i] )
                            * p_f->mu[i] * p_f->mass[i]);
            int i_pperp = floor((pperp - dist->min_pperp)
                               / ((dist->max_pperp - dist->min_pperp)
                                  / dist->n_pperp));
 
            int i_time = floor((p_f->time[i] - dist->min_time)
                          / ((dist->max_time - dist->min_time) / dist->n_time));
 
            int i_q = floor((p_f->charge[i]/CONST_E - dist->min_q)
                           / ((dist->max_q - dist->min_q) / dist->n_q));
 
            if(i_rho   >= 0 && i_rho   <= dist->n_rho - 1   &&
               i_phi   >= 0 && i_phi   <= dist->n_phi - 1   &&
               i_theta >= 0 && i_theta <= dist->n_theta - 1 &&
               i_ppara >= 0 && i_ppara <= dist->n_ppara - 1 &&
               i_pperp >= 0 && i_pperp <= dist->n_pperp - 1 &&
               i_time  >= 0 && i_time  <= dist->n_time - 1  &&
               i_q     >= 0 && i_q     <= dist->n_q - 1       ) {
          real weight = p_f->weight[i] * (p_f->time[i] - p_i->time[i]);
          size_t index = dist_rho5D_index(
                i_rho, i_theta, i_phi, i_ppara, i_pperp,
                i_time, i_q, dist->step_6, dist->step_5, dist->step_4,
                dist->step_3, dist->step_2, dist->step_1);
          GPU_ATOMIC
          dist->histogram[index] += weight;
            }
        }
    }
}