ffmpeg/libavcodec/elbg.c

532 lines
17 KiB
C

/*
* Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* Codebook Generator using the ELBG algorithm
*/
#include <string.h>
#include "libavutil/avassert.h"
#include "libavutil/common.h"
#include "libavutil/lfg.h"
#include "libavutil/mem.h"
#include "elbg.h"
#define DELTA_ERR_MAX 0.1 ///< Precision of the ELBG algorithm (as percentage error)
/**
* In the ELBG jargon, a cell is the set of points that are closest to a
* codebook entry. Not to be confused with a RoQ Video cell. */
typedef struct cell_s {
int index;
struct cell_s *next;
} cell;
/**
* ELBG internal data
*/
typedef struct ELBGContext {
int error;
int dim;
int num_cb;
int *codebook;
cell **cells;
int *utility;
int *utility_inc;
int *nearest_cb;
int *points;
int *temp_points;
int *size_part;
AVLFG *rand_state;
int *scratchbuf;
cell *cell_buffer;
/* Sizes for the buffers above. Pointers without such a field
* are not allocated by us and only valid for the duration
* of a single call to avpriv_elbg_do(). */
unsigned utility_allocated;
unsigned utility_inc_allocated;
unsigned size_part_allocated;
unsigned cells_allocated;
unsigned scratchbuf_allocated;
unsigned cell_buffer_allocated;
unsigned temp_points_allocated;
} ELBGContext;
static inline int distance_limited(int *a, int *b, int dim, int limit)
{
int i, dist=0;
for (i=0; i<dim; i++) {
int64_t distance = a[i] - b[i];
distance *= distance;
if (dist >= limit - distance)
return limit;
dist += distance;
}
return dist;
}
static inline void vect_division(int *res, int *vect, int div, int dim)
{
int i;
if (div > 1)
for (i=0; i<dim; i++)
res[i] = ROUNDED_DIV(vect[i],div);
else if (res != vect)
memcpy(res, vect, dim*sizeof(int));
}
static int eval_error_cell(ELBGContext *elbg, int *centroid, cell *cells)
{
int error=0;
for (; cells; cells=cells->next) {
int distance = distance_limited(centroid, elbg->points + cells->index*elbg->dim, elbg->dim, INT_MAX);
if (error >= INT_MAX - distance)
return INT_MAX;
error += distance;
}
return error;
}
static int get_closest_codebook(ELBGContext *elbg, int index)
{
int pick = 0;
for (int i = 0, diff_min = INT_MAX; i < elbg->num_cb; i++)
if (i != index) {
int diff;
diff = distance_limited(elbg->codebook + i*elbg->dim, elbg->codebook + index*elbg->dim, elbg->dim, diff_min);
if (diff < diff_min) {
pick = i;
diff_min = diff;
}
}
return pick;
}
static int get_high_utility_cell(ELBGContext *elbg)
{
int i=0;
/* Using linear search, do binary if it ever turns to be speed critical */
uint64_t r;
if (elbg->utility_inc[elbg->num_cb - 1] < INT_MAX) {
r = av_lfg_get(elbg->rand_state) % (unsigned int)elbg->utility_inc[elbg->num_cb - 1] + 1;
} else {
r = av_lfg_get(elbg->rand_state);
r = (av_lfg_get(elbg->rand_state) + (r<<32)) % elbg->utility_inc[elbg->num_cb - 1] + 1;
}
while (elbg->utility_inc[i] < r) {
i++;
}
av_assert2(elbg->cells[i]);
return i;
}
/**
* Implementation of the simple LBG algorithm for just two codebooks
*/
static int simple_lbg(ELBGContext *elbg,
int dim,
int *centroid[3],
int newutility[3],
int *points,
cell *cells)
{
int i, idx;
int numpoints[2] = {0,0};
int *newcentroid[2] = {
elbg->scratchbuf + 3*dim,
elbg->scratchbuf + 4*dim
};
cell *tempcell;
memset(newcentroid[0], 0, 2 * dim * sizeof(*newcentroid[0]));
newutility[0] =
newutility[1] = 0;
for (tempcell = cells; tempcell; tempcell=tempcell->next) {
idx = distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX)>=
distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX);
numpoints[idx]++;
for (i=0; i<dim; i++)
newcentroid[idx][i] += points[tempcell->index*dim + i];
}
vect_division(centroid[0], newcentroid[0], numpoints[0], dim);
vect_division(centroid[1], newcentroid[1], numpoints[1], dim);
for (tempcell = cells; tempcell; tempcell=tempcell->next) {
int dist[2] = {distance_limited(centroid[0], points + tempcell->index*dim, dim, INT_MAX),
distance_limited(centroid[1], points + tempcell->index*dim, dim, INT_MAX)};
int idx = dist[0] > dist[1];
if (newutility[idx] >= INT_MAX - dist[idx])
newutility[idx] = INT_MAX;
else
newutility[idx] += dist[idx];
}
return (newutility[0] >= INT_MAX - newutility[1]) ? INT_MAX : newutility[0] + newutility[1];
}
static void get_new_centroids(ELBGContext *elbg, int huc, int *newcentroid_i,
int *newcentroid_p)
{
cell *tempcell;
int *min = newcentroid_i;
int *max = newcentroid_p;
int i;
for (i=0; i< elbg->dim; i++) {
min[i]=INT_MAX;
max[i]=0;
}
for (tempcell = elbg->cells[huc]; tempcell; tempcell = tempcell->next)
for(i=0; i<elbg->dim; i++) {
min[i]=FFMIN(min[i], elbg->points[tempcell->index*elbg->dim + i]);
max[i]=FFMAX(max[i], elbg->points[tempcell->index*elbg->dim + i]);
}
for (i=0; i<elbg->dim; i++) {
int ni = min[i] + (max[i] - min[i])/3;
int np = min[i] + (2*(max[i] - min[i]))/3;
newcentroid_i[i] = ni;
newcentroid_p[i] = np;
}
}
/**
* Add the points in the low utility cell to its closest cell. Split the high
* utility cell, putting the separated points in the (now empty) low utility
* cell.
*
* @param elbg Internal elbg data
* @param indexes {luc, huc, cluc}
* @param newcentroid A vector with the position of the new centroids
*/
static void shift_codebook(ELBGContext *elbg, int *indexes,
int *newcentroid[3])
{
cell *tempdata;
cell **pp = &elbg->cells[indexes[2]];
while(*pp)
pp= &(*pp)->next;
*pp = elbg->cells[indexes[0]];
elbg->cells[indexes[0]] = NULL;
tempdata = elbg->cells[indexes[1]];
elbg->cells[indexes[1]] = NULL;
while(tempdata) {
cell *tempcell2 = tempdata->next;
int idx = distance_limited(elbg->points + tempdata->index*elbg->dim,
newcentroid[0], elbg->dim, INT_MAX) >
distance_limited(elbg->points + tempdata->index*elbg->dim,
newcentroid[1], elbg->dim, INT_MAX);
tempdata->next = elbg->cells[indexes[idx]];
elbg->cells[indexes[idx]] = tempdata;
tempdata = tempcell2;
}
}
static void evaluate_utility_inc(ELBGContext *elbg)
{
int64_t inc=0;
for (int i = 0; i < elbg->num_cb; i++) {
if (elbg->num_cb * (int64_t)elbg->utility[i] > elbg->error)
inc += elbg->utility[i];
elbg->utility_inc[i] = FFMIN(inc, INT_MAX);
}
}
static void update_utility_and_n_cb(ELBGContext *elbg, int idx, int newutility)
{
cell *tempcell;
elbg->utility[idx] = newutility;
for (tempcell=elbg->cells[idx]; tempcell; tempcell=tempcell->next)
elbg->nearest_cb[tempcell->index] = idx;
}
/**
* Evaluate if a shift lower the error. If it does, call shift_codebooks
* and update elbg->error, elbg->utility and elbg->nearest_cb.
*
* @param elbg Internal elbg data
* @param idx {luc (low utility cell, huc (high utility cell), cluc (closest cell to low utility cell)}
*/
static void try_shift_candidate(ELBGContext *elbg, int idx[3])
{
int j, k, cont=0, tmp;
int64_t olderror=0, newerror;
int newutility[3];
int *newcentroid[3] = {
elbg->scratchbuf,
elbg->scratchbuf + elbg->dim,
elbg->scratchbuf + 2*elbg->dim
};
cell *tempcell;
for (j=0; j<3; j++)
olderror += elbg->utility[idx[j]];
memset(newcentroid[2], 0, elbg->dim*sizeof(int));
for (k=0; k<2; k++)
for (tempcell=elbg->cells[idx[2*k]]; tempcell; tempcell=tempcell->next) {
cont++;
for (j=0; j<elbg->dim; j++)
newcentroid[2][j] += elbg->points[tempcell->index*elbg->dim + j];
}
vect_division(newcentroid[2], newcentroid[2], cont, elbg->dim);
get_new_centroids(elbg, idx[1], newcentroid[0], newcentroid[1]);
newutility[2] = eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[0]]);
tmp = eval_error_cell(elbg, newcentroid[2], elbg->cells[idx[2]]);
newutility[2] = (tmp >= INT_MAX - newutility[2]) ? INT_MAX : newutility[2] + tmp;
newerror = newutility[2];
tmp = simple_lbg(elbg, elbg->dim, newcentroid, newutility, elbg->points,
elbg->cells[idx[1]]);
if (tmp >= INT_MAX - newerror)
newerror = INT_MAX;
else
newerror += tmp;
if (olderror > newerror) {
shift_codebook(elbg, idx, newcentroid);
elbg->error += newerror - olderror;
for (j=0; j<3; j++)
update_utility_and_n_cb(elbg, idx[j], newutility[j]);
evaluate_utility_inc(elbg);
}
}
/**
* Implementation of the ELBG block
*/
static void do_shiftings(ELBGContext *elbg)
{
int idx[3];
evaluate_utility_inc(elbg);
for (idx[0]=0; idx[0] < elbg->num_cb; idx[0]++)
if (elbg->num_cb * (int64_t)elbg->utility[idx[0]] < elbg->error) {
if (elbg->utility_inc[elbg->num_cb - 1] == 0)
return;
idx[1] = get_high_utility_cell(elbg);
idx[2] = get_closest_codebook(elbg, idx[0]);
if (idx[1] != idx[0] && idx[1] != idx[2])
try_shift_candidate(elbg, idx);
}
}
static void do_elbg(ELBGContext *restrict elbg, int *points, int numpoints,
int max_steps)
{
int *const size_part = elbg->size_part;
int i, j, steps = 0;
int best_idx = 0;
int last_error;
elbg->error = INT_MAX;
elbg->points = points;
do {
cell *free_cells = elbg->cell_buffer;
last_error = elbg->error;
steps++;
memset(elbg->utility, 0, elbg->num_cb * sizeof(*elbg->utility));
memset(elbg->cells, 0, elbg->num_cb * sizeof(*elbg->cells));
elbg->error = 0;
/* This loop evaluate the actual Voronoi partition. It is the most
costly part of the algorithm. */
for (i=0; i < numpoints; i++) {
int best_dist = distance_limited(elbg->points + i * elbg->dim,
elbg->codebook + best_idx * elbg->dim,
elbg->dim, INT_MAX);
for (int k = 0; k < elbg->num_cb; k++) {
int dist = distance_limited(elbg->points + i * elbg->dim,
elbg->codebook + k * elbg->dim,
elbg->dim, best_dist);
if (dist < best_dist) {
best_dist = dist;
best_idx = k;
}
}
elbg->nearest_cb[i] = best_idx;
elbg->error = (elbg->error >= INT_MAX - best_dist) ? INT_MAX : elbg->error + best_dist;
elbg->utility[elbg->nearest_cb[i]] = (elbg->utility[elbg->nearest_cb[i]] >= INT_MAX - best_dist) ?
INT_MAX : elbg->utility[elbg->nearest_cb[i]] + best_dist;
free_cells->index = i;
free_cells->next = elbg->cells[elbg->nearest_cb[i]];
elbg->cells[elbg->nearest_cb[i]] = free_cells;
free_cells++;
}
do_shiftings(elbg);
memset(size_part, 0, elbg->num_cb * sizeof(*size_part));
memset(elbg->codebook, 0, elbg->num_cb * elbg->dim * sizeof(*elbg->codebook));
for (i=0; i < numpoints; i++) {
size_part[elbg->nearest_cb[i]]++;
for (j=0; j < elbg->dim; j++)
elbg->codebook[elbg->nearest_cb[i]*elbg->dim + j] +=
elbg->points[i*elbg->dim + j];
}
for (int i = 0; i < elbg->num_cb; i++)
vect_division(elbg->codebook + i*elbg->dim,
elbg->codebook + i*elbg->dim, size_part[i], elbg->dim);
} while(((last_error - elbg->error) > DELTA_ERR_MAX*elbg->error) &&
(steps < max_steps));
}
#define BIG_PRIME 433494437LL
/**
* Initialize the codebook vector for the elbg algorithm.
* If numpoints <= 24 * num_cb this function fills codebook with random numbers.
* If not, it calls do_elbg for a (smaller) random sample of the points in
* points.
*/
static void init_elbg(ELBGContext *restrict elbg, int *points, int *temp_points,
int numpoints, int max_steps)
{
int dim = elbg->dim;
if (numpoints > 24LL * elbg->num_cb) {
/* ELBG is very costly for a big number of points. So if we have a lot
of them, get a good initial codebook to save on iterations */
for (int i = 0; i < numpoints / 8; i++) {
int k = (i*BIG_PRIME) % numpoints;
memcpy(temp_points + i*dim, points + k*dim, dim * sizeof(*temp_points));
}
/* If anything is changed in the recursion parameters,
* the allocated size of temp_points will also need to be updated. */
init_elbg(elbg, temp_points, temp_points + numpoints / 8 * dim,
numpoints / 8, 2 * max_steps);
do_elbg(elbg, temp_points, numpoints / 8, 2 * max_steps);
} else // If not, initialize the codebook with random positions
for (int i = 0; i < elbg->num_cb; i++)
memcpy(elbg->codebook + i * dim, points + ((i*BIG_PRIME)%numpoints)*dim,
dim * sizeof(*elbg->codebook));
}
int avpriv_elbg_do(ELBGContext **elbgp, int *points, int dim, int numpoints,
int *codebook, int num_cb, int max_steps,
int *closest_cb, AVLFG *rand_state, uintptr_t flags)
{
ELBGContext *const restrict elbg = *elbgp ? *elbgp : av_mallocz(sizeof(*elbg));
if (!elbg)
return AVERROR(ENOMEM);
*elbgp = elbg;
elbg->nearest_cb = closest_cb;
elbg->rand_state = rand_state;
elbg->codebook = codebook;
elbg->num_cb = num_cb;
elbg->dim = dim;
#define ALLOCATE_IF_NECESSARY(field, new_elements, multiplicator) \
if (elbg->field ## _allocated < new_elements) { \
av_freep(&elbg->field); \
elbg->field = av_malloc_array(new_elements, \
multiplicator * sizeof(*elbg->field)); \
if (!elbg->field) { \
elbg->field ## _allocated = 0; \
return AVERROR(ENOMEM); \
} \
elbg->field ## _allocated = new_elements; \
}
/* Allocating the buffers for do_elbg() here once relies
* on their size being always the same even when do_elbg()
* is called from init_elbg(). It also relies on do_elbg()
* never calling itself recursively. */
ALLOCATE_IF_NECESSARY(cells, num_cb, 1)
ALLOCATE_IF_NECESSARY(utility, num_cb, 1)
ALLOCATE_IF_NECESSARY(utility_inc, num_cb, 1)
ALLOCATE_IF_NECESSARY(size_part, num_cb, 1)
ALLOCATE_IF_NECESSARY(cell_buffer, numpoints, 1)
ALLOCATE_IF_NECESSARY(scratchbuf, dim, 5)
if (numpoints > 24LL * elbg->num_cb) {
/* The first step in the recursion in init_elbg() needs a buffer with
* (numpoints / 8) * dim elements; the next step needs numpoints / 8 / 8
* * dim elements etc. The geometric series leads to an upper bound of
* numpoints / 8 * 8 / 7 * dim elements. */
uint64_t prod = dim * (uint64_t)(numpoints / 7U);
if (prod > INT_MAX)
return AVERROR(ERANGE);
ALLOCATE_IF_NECESSARY(temp_points, prod, 1)
}
init_elbg(elbg, points, elbg->temp_points, numpoints, max_steps);
do_elbg (elbg, points, numpoints, max_steps);
return 0;
}
av_cold void avpriv_elbg_free(ELBGContext **elbgp)
{
ELBGContext *elbg = *elbgp;
if (!elbg)
return;
av_freep(&elbg->size_part);
av_freep(&elbg->utility);
av_freep(&elbg->cell_buffer);
av_freep(&elbg->cells);
av_freep(&elbg->utility_inc);
av_freep(&elbg->scratchbuf);
av_freep(&elbg->temp_points);
av_freep(elbgp);
}