2 * Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com>
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * Codebook Generator using the ELBG algorithm
28 #include "libavutil/lfg.h"
32 #define DELTA_ERR_MAX 0.1 ///< Precision of the ELBG algorithm (as percentual error)
35 * In the ELBG jargon, a cell is the set of points that are closest to a
36 * codebook entry. Not to be confused with a RoQ Video cell. */
37 typedef struct cell_s {
59 static inline int distance_limited(int *a, int *b, int dim, int limit)
62 for (i=0; i<dim; i++) {
63 dist += (a[i] - b[i])*(a[i] - b[i]);
71 static inline void vect_division(int *res, int *vect, int div, int dim)
76 res[i] = ROUNDED_DIV(vect[i],div);
78 memcpy(res, vect, dim*sizeof(int));
82 static int eval_error_cell(elbg_data *elbg, int *centroid, cell *cells)
85 for (; cells; cells=cells->next)
86 error += distance_limited(centroid, elbg->points + cells->index*elbg->dim, elbg->dim, INT_MAX);
91 static int get_closest_codebook(elbg_data *elbg, int index)
93 int i, pick=0, diff, diff_min = INT_MAX;
94 for (i=0; i<elbg->numCB; i++)
96 diff = distance_limited(elbg->codebook + i*elbg->dim, elbg->codebook + index*elbg->dim, elbg->dim, diff_min);
97 if (diff < diff_min) {
105 static int get_high_utility_cell(elbg_data *elbg)
108 /* Using linear search, do binary if it ever turns to be speed critical */
109 int r = av_lfg_get(elbg->rand_state)%elbg->utility_inc[elbg->numCB-1] + 1;
110 while (elbg->utility_inc[i] < r)
113 assert(!elbg->cells[i]);
119 * Implementation of the simple LBG algorithm for just two codebooks
121 static int simple_lbg(elbg_data *elbg,
129 int numpoints[2] = {0,0};
130 int *newcentroid[2] = {
131 elbg->scratchbuf + 3*dim,
132 elbg->scratchbuf + 4*dim
136 memset(newcentroid[0], 0, 2 * dim * sizeof(*newcentroid[0]));
141 for (tempcell = cells; tempcell; tempcell=tempcell->next) {
142 idx = distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX)>=
143 distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX);
145 for (i=0; i<dim; i++)
146 newcentroid[idx][i] += points[tempcell->index*dim + i];
149 vect_division(centroid[0], newcentroid[0], numpoints[0], dim);
150 vect_division(centroid[1], newcentroid[1], numpoints[1], dim);
152 for (tempcell = cells; tempcell; tempcell=tempcell->next) {
153 int dist[2] = {distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX),
154 distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX)};
155 int idx = dist[0] > dist[1];
156 newutility[idx] += dist[idx];
159 return newutility[0] + newutility[1];
162 static void get_new_centroids(elbg_data *elbg, int huc, int *newcentroid_i,
166 int *min = newcentroid_i;
167 int *max = newcentroid_p;
170 for (i=0; i< elbg->dim; i++) {
175 for (tempcell = elbg->cells[huc]; tempcell; tempcell = tempcell->next)
176 for(i=0; i<elbg->dim; i++) {
177 min[i]=FFMIN(min[i], elbg->points[tempcell->index*elbg->dim + i]);
178 max[i]=FFMAX(max[i], elbg->points[tempcell->index*elbg->dim + i]);
181 for (i=0; i<elbg->dim; i++) {
182 int ni = min[i] + (max[i] - min[i])/3;
183 int np = min[i] + (2*(max[i] - min[i]))/3;
184 newcentroid_i[i] = ni;
185 newcentroid_p[i] = np;
190 * Add the points in the low utility cell to its closest cell. Split the high
191 * utility cell, putting the separed points in the (now empty) low utility
194 * @param elbg Internal elbg data
195 * @param indexes {luc, huc, cluc}
196 * @param newcentroid A vector with the position of the new centroids
198 static void shift_codebook(elbg_data *elbg, int *indexes,
202 cell **pp = &elbg->cells[indexes[2]];
207 *pp = elbg->cells[indexes[0]];
209 elbg->cells[indexes[0]] = NULL;
210 tempdata = elbg->cells[indexes[1]];
211 elbg->cells[indexes[1]] = NULL;
214 cell *tempcell2 = tempdata->next;
215 int idx = distance_limited(elbg->points + tempdata->index*elbg->dim,
216 newcentroid[0], elbg->dim, INT_MAX) >
217 distance_limited(elbg->points + tempdata->index*elbg->dim,
218 newcentroid[1], elbg->dim, INT_MAX);
220 tempdata->next = elbg->cells[indexes[idx]];
221 elbg->cells[indexes[idx]] = tempdata;
222 tempdata = tempcell2;
226 static void evaluate_utility_inc(elbg_data *elbg)
230 for (i=0; i < elbg->numCB; i++) {
231 if (elbg->numCB*elbg->utility[i] > elbg->error)
232 inc += elbg->utility[i];
233 elbg->utility_inc[i] = inc;
238 static void update_utility_and_n_cb(elbg_data *elbg, int idx, int newutility)
242 elbg->utility[idx] = newutility;
243 for (tempcell=elbg->cells[idx]; tempcell; tempcell=tempcell->next)
244 elbg->nearest_cb[tempcell->index] = idx;
248 * Evaluate if a shift lower the error. If it does, call shift_codebooks
249 * and update elbg->error, elbg->utility and elbg->nearest_cb.
251 * @param elbg Internal elbg data
252 * @param idx {luc (low utility cell, huc (high utility cell), cluc (closest cell to low utility cell)}
254 static void try_shift_candidate(elbg_data *elbg, int idx[3])
256 int j, k, olderror=0, newerror, cont=0;
258 int *newcentroid[3] = {
260 elbg->scratchbuf + elbg->dim,
261 elbg->scratchbuf + 2*elbg->dim
266 olderror += elbg->utility[idx[j]];
268 memset(newcentroid[2], 0, elbg->dim*sizeof(int));
271 for (tempcell=elbg->cells[idx[2*k]]; tempcell; tempcell=tempcell->next) {
273 for (j=0; j<elbg->dim; j++)
274 newcentroid[2][j] += elbg->points[tempcell->index*elbg->dim + j];
277 vect_division(newcentroid[2], newcentroid[2], cont, elbg->dim);
279 get_new_centroids(elbg, idx[1], newcentroid[0], newcentroid[1]);
281 newutility[2] = eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[0]]);
282 newutility[2] += eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[2]]);
284 newerror = newutility[2];
286 newerror += simple_lbg(elbg, elbg->dim, newcentroid, newutility, elbg->points,
287 elbg->cells[idx[1]]);
289 if (olderror > newerror) {
290 shift_codebook(elbg, idx, newcentroid);
292 elbg->error += newerror - olderror;
295 update_utility_and_n_cb(elbg, idx[j], newutility[j]);
297 evaluate_utility_inc(elbg);
302 * Implementation of the ELBG block
304 static void do_shiftings(elbg_data *elbg)
308 evaluate_utility_inc(elbg);
310 for (idx[0]=0; idx[0] < elbg->numCB; idx[0]++)
311 if (elbg->numCB*elbg->utility[idx[0]] < elbg->error) {
312 if (elbg->utility_inc[elbg->numCB-1] == 0)
315 idx[1] = get_high_utility_cell(elbg);
316 idx[2] = get_closest_codebook(elbg, idx[0]);
318 if (idx[1] != idx[0] && idx[1] != idx[2])
319 try_shift_candidate(elbg, idx);
323 #define BIG_PRIME 433494437LL
325 void ff_init_elbg(int *points, int dim, int numpoints, int *codebook,
326 int numCB, int max_steps, int *closest_cb,
331 if (numpoints > 24*numCB) {
332 /* ELBG is very costly for a big number of points. So if we have a lot
333 of them, get a good initial codebook to save on iterations */
334 int *temp_points = av_malloc(dim*(numpoints/8)*sizeof(int));
335 for (i=0; i<numpoints/8; i++) {
336 k = (i*BIG_PRIME) % numpoints;
337 memcpy(temp_points + i*dim, points + k*dim, dim*sizeof(int));
340 ff_init_elbg(temp_points, dim, numpoints/8, codebook, numCB, 2*max_steps, closest_cb, rand_state);
341 ff_do_elbg(temp_points, dim, numpoints/8, codebook, numCB, 2*max_steps, closest_cb, rand_state);
343 av_free(temp_points);
345 } else // If not, initialize the codebook with random positions
346 for (i=0; i < numCB; i++)
347 memcpy(codebook + i*dim, points + ((i*BIG_PRIME)%numpoints)*dim,
352 void ff_do_elbg(int *points, int dim, int numpoints, int *codebook,
353 int numCB, int max_steps, int *closest_cb,
358 elbg_data *elbg = &elbg_d;
359 int i, j, k, last_error, steps=0;
360 int *dist_cb = av_malloc(numpoints*sizeof(int));
361 int *size_part = av_malloc(numCB*sizeof(int));
362 cell *list_buffer = av_malloc(numpoints*sizeof(cell));
364 int best_dist, best_idx = 0;
366 elbg->error = INT_MAX;
369 elbg->codebook = codebook;
370 elbg->cells = av_malloc(numCB*sizeof(cell *));
371 elbg->utility = av_malloc(numCB*sizeof(int));
372 elbg->nearest_cb = closest_cb;
373 elbg->points = points;
374 elbg->utility_inc = av_malloc(numCB*sizeof(int));
375 elbg->scratchbuf = av_malloc(5*dim*sizeof(int));
377 elbg->rand_state = rand_state;
380 free_cells = list_buffer;
381 last_error = elbg->error;
383 memset(elbg->utility, 0, numCB*sizeof(int));
384 memset(elbg->cells, 0, numCB*sizeof(cell *));
388 /* This loop evaluate the actual Voronoi partition. It is the most
389 costly part of the algorithm. */
390 for (i=0; i < numpoints; i++) {
391 best_dist = distance_limited(elbg->points + i*elbg->dim, elbg->codebook + best_idx*elbg->dim, dim, INT_MAX);
392 for (k=0; k < elbg->numCB; k++) {
393 dist = distance_limited(elbg->points + i*elbg->dim, elbg->codebook + k*elbg->dim, dim, best_dist);
394 if (dist < best_dist) {
399 elbg->nearest_cb[i] = best_idx;
400 dist_cb[i] = best_dist;
401 elbg->error += dist_cb[i];
402 elbg->utility[elbg->nearest_cb[i]] += dist_cb[i];
403 free_cells->index = i;
404 free_cells->next = elbg->cells[elbg->nearest_cb[i]];
405 elbg->cells[elbg->nearest_cb[i]] = free_cells;
411 memset(size_part, 0, numCB*sizeof(int));
413 memset(elbg->codebook, 0, elbg->numCB*dim*sizeof(int));
415 for (i=0; i < numpoints; i++) {
416 size_part[elbg->nearest_cb[i]]++;
417 for (j=0; j < elbg->dim; j++)
418 elbg->codebook[elbg->nearest_cb[i]*elbg->dim + j] +=
419 elbg->points[i*elbg->dim + j];
422 for (i=0; i < elbg->numCB; i++)
423 vect_division(elbg->codebook + i*elbg->dim,
424 elbg->codebook + i*elbg->dim, size_part[i], elbg->dim);
426 } while(((last_error - elbg->error) > DELTA_ERR_MAX*elbg->error) &&
427 (steps < max_steps));
431 av_free(elbg->utility);
432 av_free(list_buffer);
433 av_free(elbg->cells);
434 av_free(elbg->utility_inc);
435 av_free(elbg->scratchbuf);