yading@10: /*
yading@10:  * Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com>
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * Codebook Generator using the ELBG algorithm
yading@10:  */
yading@10: 
yading@10: #include <string.h>
yading@10: 
yading@10: #include "libavutil/avassert.h"
yading@10: #include "libavutil/common.h"
yading@10: #include "libavutil/lfg.h"
yading@10: #include "elbg.h"
yading@10: #include "avcodec.h"
yading@10: 
yading@10: #define DELTA_ERR_MAX 0.1  ///< Precision of the ELBG algorithm (as percentual error)
yading@10: 
yading@10: /**
yading@10:  * In the ELBG jargon, a cell is the set of points that are closest to a
yading@10:  * codebook entry. Not to be confused with a RoQ Video cell. */
yading@10: typedef struct cell_s {
yading@10:     int index;
yading@10:     struct cell_s *next;
yading@10: } cell;
yading@10: 
yading@10: /**
yading@10:  * ELBG internal data
yading@10:  */
yading@10: typedef struct{
yading@10:     int error;
yading@10:     int dim;
yading@10:     int numCB;
yading@10:     int *codebook;
yading@10:     cell **cells;
yading@10:     int *utility;
yading@10:     int *utility_inc;
yading@10:     int *nearest_cb;
yading@10:     int *points;
yading@10:     AVLFG *rand_state;
yading@10:     int *scratchbuf;
yading@10: } elbg_data;
yading@10: 
yading@10: static inline int distance_limited(int *a, int *b, int dim, int limit)
yading@10: {
yading@10:     int i, dist=0;
yading@10:     for (i=0; i<dim; i++) {
yading@10:         dist += (a[i] - b[i])*(a[i] - b[i]);
yading@10:         if (dist > limit)
yading@10:             return INT_MAX;
yading@10:     }
yading@10: 
yading@10:     return dist;
yading@10: }
yading@10: 
yading@10: static inline void vect_division(int *res, int *vect, int div, int dim)
yading@10: {
yading@10:     int i;
yading@10:     if (div > 1)
yading@10:         for (i=0; i<dim; i++)
yading@10:             res[i] = ROUNDED_DIV(vect[i],div);
yading@10:     else if (res != vect)
yading@10:         memcpy(res, vect, dim*sizeof(int));
yading@10: 
yading@10: }
yading@10: 
yading@10: static int eval_error_cell(elbg_data *elbg, int *centroid, cell *cells)
yading@10: {
yading@10:     int error=0;
yading@10:     for (; cells; cells=cells->next)
yading@10:         error += distance_limited(centroid, elbg->points + cells->index*elbg->dim, elbg->dim, INT_MAX);
yading@10: 
yading@10:     return error;
yading@10: }
yading@10: 
yading@10: static int get_closest_codebook(elbg_data *elbg, int index)
yading@10: {
yading@10:     int i, pick=0, diff, diff_min = INT_MAX;
yading@10:     for (i=0; i<elbg->numCB; i++)
yading@10:         if (i != index) {
yading@10:             diff = distance_limited(elbg->codebook + i*elbg->dim, elbg->codebook + index*elbg->dim, elbg->dim, diff_min);
yading@10:             if (diff < diff_min) {
yading@10:                 pick = i;
yading@10:                 diff_min = diff;
yading@10:             }
yading@10:         }
yading@10:     return pick;
yading@10: }
yading@10: 
yading@10: static int get_high_utility_cell(elbg_data *elbg)
yading@10: {
yading@10:     int i=0;
yading@10:     /* Using linear search, do binary if it ever turns to be speed critical */
yading@10:     int r = av_lfg_get(elbg->rand_state)%elbg->utility_inc[elbg->numCB-1] + 1;
yading@10:     while (elbg->utility_inc[i] < r)
yading@10:         i++;
yading@10: 
yading@10:     av_assert2(elbg->cells[i]);
yading@10: 
yading@10:     return i;
yading@10: }
yading@10: 
yading@10: /**
yading@10:  * Implementation of the simple LBG algorithm for just two codebooks
yading@10:  */
yading@10: static int simple_lbg(elbg_data *elbg,
yading@10:                       int dim,
yading@10:                       int *centroid[3],
yading@10:                       int newutility[3],
yading@10:                       int *points,
yading@10:                       cell *cells)
yading@10: {
yading@10:     int i, idx;
yading@10:     int numpoints[2] = {0,0};
yading@10:     int *newcentroid[2] = {
yading@10:         elbg->scratchbuf + 3*dim,
yading@10:         elbg->scratchbuf + 4*dim
yading@10:     };
yading@10:     cell *tempcell;
yading@10: 
yading@10:     memset(newcentroid[0], 0, 2 * dim * sizeof(*newcentroid[0]));
yading@10: 
yading@10:     newutility[0] =
yading@10:     newutility[1] = 0;
yading@10: 
yading@10:     for (tempcell = cells; tempcell; tempcell=tempcell->next) {
yading@10:         idx = distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX)>=
yading@10:               distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX);
yading@10:         numpoints[idx]++;
yading@10:         for (i=0; i<dim; i++)
yading@10:             newcentroid[idx][i] += points[tempcell->index*dim + i];
yading@10:     }
yading@10: 
yading@10:     vect_division(centroid[0], newcentroid[0], numpoints[0], dim);
yading@10:     vect_division(centroid[1], newcentroid[1], numpoints[1], dim);
yading@10: 
yading@10:     for (tempcell = cells; tempcell; tempcell=tempcell->next) {
yading@10:         int dist[2] = {distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX),
yading@10:                        distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX)};
yading@10:         int idx = dist[0] > dist[1];
yading@10:         newutility[idx] += dist[idx];
yading@10:     }
yading@10: 
yading@10:     return newutility[0] + newutility[1];
yading@10: }
yading@10: 
yading@10: static void get_new_centroids(elbg_data *elbg, int huc, int *newcentroid_i,
yading@10:                               int *newcentroid_p)
yading@10: {
yading@10:     cell *tempcell;
yading@10:     int *min = newcentroid_i;
yading@10:     int *max = newcentroid_p;
yading@10:     int i;
yading@10: 
yading@10:     for (i=0; i< elbg->dim; i++) {
yading@10:         min[i]=INT_MAX;
yading@10:         max[i]=0;
yading@10:     }
yading@10: 
yading@10:     for (tempcell = elbg->cells[huc]; tempcell; tempcell = tempcell->next)
yading@10:         for(i=0; i<elbg->dim; i++) {
yading@10:             min[i]=FFMIN(min[i], elbg->points[tempcell->index*elbg->dim + i]);
yading@10:             max[i]=FFMAX(max[i], elbg->points[tempcell->index*elbg->dim + i]);
yading@10:         }
yading@10: 
yading@10:     for (i=0; i<elbg->dim; i++) {
yading@10:         int ni = min[i] + (max[i] - min[i])/3;
yading@10:         int np = min[i] + (2*(max[i] - min[i]))/3;
yading@10:         newcentroid_i[i] = ni;
yading@10:         newcentroid_p[i] = np;
yading@10:     }
yading@10: }
yading@10: 
yading@10: /**
yading@10:  * Add the points in the low utility cell to its closest cell. Split the high
yading@10:  * utility cell, putting the separate points in the (now empty) low utility
yading@10:  * cell.
yading@10:  *
yading@10:  * @param elbg         Internal elbg data
yading@10:  * @param indexes      {luc, huc, cluc}
yading@10:  * @param newcentroid  A vector with the position of the new centroids
yading@10:  */
yading@10: static void shift_codebook(elbg_data *elbg, int *indexes,
yading@10:                            int *newcentroid[3])
yading@10: {
yading@10:     cell *tempdata;
yading@10:     cell **pp = &elbg->cells[indexes[2]];
yading@10: 
yading@10:     while(*pp)
yading@10:         pp= &(*pp)->next;
yading@10: 
yading@10:     *pp = elbg->cells[indexes[0]];
yading@10: 
yading@10:     elbg->cells[indexes[0]] = NULL;
yading@10:     tempdata = elbg->cells[indexes[1]];
yading@10:     elbg->cells[indexes[1]] = NULL;
yading@10: 
yading@10:     while(tempdata) {
yading@10:         cell *tempcell2 = tempdata->next;
yading@10:         int idx = distance_limited(elbg->points + tempdata->index*elbg->dim,
yading@10:                            newcentroid[0], elbg->dim, INT_MAX) >
yading@10:                   distance_limited(elbg->points + tempdata->index*elbg->dim,
yading@10:                            newcentroid[1], elbg->dim, INT_MAX);
yading@10: 
yading@10:         tempdata->next = elbg->cells[indexes[idx]];
yading@10:         elbg->cells[indexes[idx]] = tempdata;
yading@10:         tempdata = tempcell2;
yading@10:     }
yading@10: }
yading@10: 
yading@10: static void evaluate_utility_inc(elbg_data *elbg)
yading@10: {
yading@10:     int i, inc=0;
yading@10: 
yading@10:     for (i=0; i < elbg->numCB; i++) {
yading@10:         if (elbg->numCB*elbg->utility[i] > elbg->error)
yading@10:             inc += elbg->utility[i];
yading@10:         elbg->utility_inc[i] = inc;
yading@10:     }
yading@10: }
yading@10: 
yading@10: 
yading@10: static void update_utility_and_n_cb(elbg_data *elbg, int idx, int newutility)
yading@10: {
yading@10:     cell *tempcell;
yading@10: 
yading@10:     elbg->utility[idx] = newutility;
yading@10:     for (tempcell=elbg->cells[idx]; tempcell; tempcell=tempcell->next)
yading@10:         elbg->nearest_cb[tempcell->index] = idx;
yading@10: }
yading@10: 
yading@10: /**
yading@10:  * Evaluate if a shift lower the error. If it does, call shift_codebooks
yading@10:  * and update elbg->error, elbg->utility and elbg->nearest_cb.
yading@10:  *
yading@10:  * @param elbg  Internal elbg data
yading@10:  * @param idx   {luc (low utility cell, huc (high utility cell), cluc (closest cell to low utility cell)}
yading@10:  */
yading@10: static void try_shift_candidate(elbg_data *elbg, int idx[3])
yading@10: {
yading@10:     int j, k, olderror=0, newerror, cont=0;
yading@10:     int newutility[3];
yading@10:     int *newcentroid[3] = {
yading@10:         elbg->scratchbuf,
yading@10:         elbg->scratchbuf + elbg->dim,
yading@10:         elbg->scratchbuf + 2*elbg->dim
yading@10:     };
yading@10:     cell *tempcell;
yading@10: 
yading@10:     for (j=0; j<3; j++)
yading@10:         olderror += elbg->utility[idx[j]];
yading@10: 
yading@10:     memset(newcentroid[2], 0, elbg->dim*sizeof(int));
yading@10: 
yading@10:     for (k=0; k<2; k++)
yading@10:         for (tempcell=elbg->cells[idx[2*k]]; tempcell; tempcell=tempcell->next) {
yading@10:             cont++;
yading@10:             for (j=0; j<elbg->dim; j++)
yading@10:                 newcentroid[2][j] += elbg->points[tempcell->index*elbg->dim + j];
yading@10:         }
yading@10: 
yading@10:     vect_division(newcentroid[2], newcentroid[2], cont, elbg->dim);
yading@10: 
yading@10:     get_new_centroids(elbg, idx[1], newcentroid[0], newcentroid[1]);
yading@10: 
yading@10:     newutility[2]  = eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[0]]);
yading@10:     newutility[2] += eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[2]]);
yading@10: 
yading@10:     newerror = newutility[2];
yading@10: 
yading@10:     newerror += simple_lbg(elbg, elbg->dim, newcentroid, newutility, elbg->points,
yading@10:                            elbg->cells[idx[1]]);
yading@10: 
yading@10:     if (olderror > newerror) {
yading@10:         shift_codebook(elbg, idx, newcentroid);
yading@10: 
yading@10:         elbg->error += newerror - olderror;
yading@10: 
yading@10:         for (j=0; j<3; j++)
yading@10:             update_utility_and_n_cb(elbg, idx[j], newutility[j]);
yading@10: 
yading@10:         evaluate_utility_inc(elbg);
yading@10:     }
yading@10:  }
yading@10: 
yading@10: /**
yading@10:  * Implementation of the ELBG block
yading@10:  */
yading@10: static void do_shiftings(elbg_data *elbg)
yading@10: {
yading@10:     int idx[3];
yading@10: 
yading@10:     evaluate_utility_inc(elbg);
yading@10: 
yading@10:     for (idx[0]=0; idx[0] < elbg->numCB; idx[0]++)
yading@10:         if (elbg->numCB*elbg->utility[idx[0]] < elbg->error) {
yading@10:             if (elbg->utility_inc[elbg->numCB-1] == 0)
yading@10:                 return;
yading@10: 
yading@10:             idx[1] = get_high_utility_cell(elbg);
yading@10:             idx[2] = get_closest_codebook(elbg, idx[0]);
yading@10: 
yading@10:             if (idx[1] != idx[0] && idx[1] != idx[2])
yading@10:                 try_shift_candidate(elbg, idx);
yading@10:         }
yading@10: }
yading@10: 
yading@10: #define BIG_PRIME 433494437LL
yading@10: 
yading@10: void ff_init_elbg(int *points, int dim, int numpoints, int *codebook,
yading@10:                   int numCB, int max_steps, int *closest_cb,
yading@10:                   AVLFG *rand_state)
yading@10: {
yading@10:     int i, k;
yading@10: 
yading@10:     if (numpoints > 24*numCB) {
yading@10:         /* ELBG is very costly for a big number of points. So if we have a lot
yading@10:            of them, get a good initial codebook to save on iterations       */
yading@10:         int *temp_points = av_malloc(dim*(numpoints/8)*sizeof(int));
yading@10:         for (i=0; i<numpoints/8; i++) {
yading@10:             k = (i*BIG_PRIME) % numpoints;
yading@10:             memcpy(temp_points + i*dim, points + k*dim, dim*sizeof(int));
yading@10:         }
yading@10: 
yading@10:         ff_init_elbg(temp_points, dim, numpoints/8, codebook, numCB, 2*max_steps, closest_cb, rand_state);
yading@10:         ff_do_elbg(temp_points, dim, numpoints/8, codebook, numCB, 2*max_steps, closest_cb, rand_state);
yading@10: 
yading@10:         av_free(temp_points);
yading@10: 
yading@10:     } else  // If not, initialize the codebook with random positions
yading@10:         for (i=0; i < numCB; i++)
yading@10:             memcpy(codebook + i*dim, points + ((i*BIG_PRIME)%numpoints)*dim,
yading@10:                    dim*sizeof(int));
yading@10: 
yading@10: }
yading@10: 
yading@10: void ff_do_elbg(int *points, int dim, int numpoints, int *codebook,
yading@10:                 int numCB, int max_steps, int *closest_cb,
yading@10:                 AVLFG *rand_state)
yading@10: {
yading@10:     int dist;
yading@10:     elbg_data elbg_d;
yading@10:     elbg_data *elbg = &elbg_d;
yading@10:     int i, j, k, last_error, steps=0;
yading@10:     int *dist_cb = av_malloc(numpoints*sizeof(int));
yading@10:     int *size_part = av_malloc(numCB*sizeof(int));
yading@10:     cell *list_buffer = av_malloc(numpoints*sizeof(cell));
yading@10:     cell *free_cells;
yading@10:     int best_dist, best_idx = 0;
yading@10: 
yading@10:     elbg->error = INT_MAX;
yading@10:     elbg->dim = dim;
yading@10:     elbg->numCB = numCB;
yading@10:     elbg->codebook = codebook;
yading@10:     elbg->cells = av_malloc(numCB*sizeof(cell *));
yading@10:     elbg->utility = av_malloc(numCB*sizeof(int));
yading@10:     elbg->nearest_cb = closest_cb;
yading@10:     elbg->points = points;
yading@10:     elbg->utility_inc = av_malloc(numCB*sizeof(int));
yading@10:     elbg->scratchbuf = av_malloc(5*dim*sizeof(int));
yading@10: 
yading@10:     elbg->rand_state = rand_state;
yading@10: 
yading@10:     do {
yading@10:         free_cells = list_buffer;
yading@10:         last_error = elbg->error;
yading@10:         steps++;
yading@10:         memset(elbg->utility, 0, numCB*sizeof(int));
yading@10:         memset(elbg->cells, 0, numCB*sizeof(cell *));
yading@10: 
yading@10:         elbg->error = 0;
yading@10: 
yading@10:         /* This loop evaluate the actual Voronoi partition. It is the most
yading@10:            costly part of the algorithm. */
yading@10:         for (i=0; i < numpoints; i++) {
yading@10:             best_dist = distance_limited(elbg->points + i*elbg->dim, elbg->codebook + best_idx*elbg->dim, dim, INT_MAX);
yading@10:             for (k=0; k < elbg->numCB; k++) {
yading@10:                 dist = distance_limited(elbg->points + i*elbg->dim, elbg->codebook + k*elbg->dim, dim, best_dist);
yading@10:                 if (dist < best_dist) {
yading@10:                     best_dist = dist;
yading@10:                     best_idx = k;
yading@10:                 }
yading@10:             }
yading@10:             elbg->nearest_cb[i] = best_idx;
yading@10:             dist_cb[i] = best_dist;
yading@10:             elbg->error += dist_cb[i];
yading@10:             elbg->utility[elbg->nearest_cb[i]] += dist_cb[i];
yading@10:             free_cells->index = i;
yading@10:             free_cells->next = elbg->cells[elbg->nearest_cb[i]];
yading@10:             elbg->cells[elbg->nearest_cb[i]] = free_cells;
yading@10:             free_cells++;
yading@10:         }
yading@10: 
yading@10:         do_shiftings(elbg);
yading@10: 
yading@10:         memset(size_part, 0, numCB*sizeof(int));
yading@10: 
yading@10:         memset(elbg->codebook, 0, elbg->numCB*dim*sizeof(int));
yading@10: 
yading@10:         for (i=0; i < numpoints; i++) {
yading@10:             size_part[elbg->nearest_cb[i]]++;
yading@10:             for (j=0; j < elbg->dim; j++)
yading@10:                 elbg->codebook[elbg->nearest_cb[i]*elbg->dim + j] +=
yading@10:                     elbg->points[i*elbg->dim + j];
yading@10:         }
yading@10: 
yading@10:         for (i=0; i < elbg->numCB; i++)
yading@10:             vect_division(elbg->codebook + i*elbg->dim,
yading@10:                           elbg->codebook + i*elbg->dim, size_part[i], elbg->dim);
yading@10: 
yading@10:     } while(((last_error - elbg->error) > DELTA_ERR_MAX*elbg->error) &&
yading@10:             (steps < max_steps));
yading@10: 
yading@10:     av_free(dist_cb);
yading@10:     av_free(size_part);
yading@10:     av_free(elbg->utility);
yading@10:     av_free(list_buffer);
yading@10:     av_free(elbg->cells);
yading@10:     av_free(elbg->utility_inc);
yading@10:     av_free(elbg->scratchbuf);
yading@10: }