2 * Copyright (c) 2015 Stupeflix
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 * Generate one palette for a whole video stream.
26 #include "libavutil/avassert.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/opt.h"
29 #include "libavutil/qsort.h"
33 /* Reference a color and how much it's used */
39 /* Store a range of colors */
41 uint32_t color; // average color
42 int64_t variance; // overall variance of the box (how much the colors are spread)
43 int start; // index in PaletteGenContext->refs
44 int len; // number of referenced colors
45 int sorted_by; // whether range of colors is sorted by red (0), green (1) or blue (2)
49 struct color_ref *entries;
54 STATS_MODE_ALL_FRAMES,
55 STATS_MODE_DIFF_FRAMES,
60 #define HIST_SIZE (1<<(3*NBITS))
66 int reserve_transparent;
69 AVFrame *prev_frame; // previous frame used for the diff stats_mode
70 struct hist_node histogram[HIST_SIZE]; // histogram/hashtable of the colors
71 struct color_ref **refs; // references of all the colors used in the stream
72 int nb_refs; // number of color references (or number of different colors)
73 struct range_box boxes[256]; // define the segmentation of the colorspace (the final palette)
74 int nb_boxes; // number of boxes (increase will segmenting them)
75 int palette_pushed; // if the palette frame is pushed into the outlink or not
78 #define OFFSET(x) offsetof(PaletteGenContext, x)
79 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
80 static const AVOption palettegen_options[] = {
81 { "max_colors", "set the maximum number of colors to use in the palette", OFFSET(max_colors), AV_OPT_TYPE_INT, {.i64=256}, 4, 256, FLAGS },
82 { "reserve_transparent", "reserve a palette entry for transparency", OFFSET(reserve_transparent), AV_OPT_TYPE_BOOL, {.i64=1}, 0, 1, FLAGS },
83 { "stats_mode", "set statistics mode", OFFSET(stats_mode), AV_OPT_TYPE_INT, {.i64=STATS_MODE_ALL_FRAMES}, 0, NB_STATS_MODE, FLAGS, "mode" },
84 { "full", "compute full frame histograms", 0, AV_OPT_TYPE_CONST, {.i64=STATS_MODE_ALL_FRAMES}, INT_MIN, INT_MAX, FLAGS, "mode" },
85 { "diff", "compute histograms only for the part that differs from previous frame", 0, AV_OPT_TYPE_CONST, {.i64=STATS_MODE_DIFF_FRAMES}, INT_MIN, INT_MAX, FLAGS, "mode" },
89 AVFILTER_DEFINE_CLASS(palettegen);
91 static int query_formats(AVFilterContext *ctx)
93 static const enum AVPixelFormat in_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
94 static const enum AVPixelFormat out_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
95 AVFilterFormats *in = ff_make_format_list(in_fmts);
96 AVFilterFormats *out = ff_make_format_list(out_fmts);
100 return AVERROR(ENOMEM);
102 ff_formats_ref(in, &ctx->inputs[0]->out_formats);
103 ff_formats_ref(out, &ctx->outputs[0]->in_formats);
107 typedef int (*cmp_func)(const void *, const void *);
109 #define DECLARE_CMP_FUNC(name, pos) \
110 static int cmp_##name(const void *pa, const void *pb) \
112 const struct color_ref * const *a = pa; \
113 const struct color_ref * const *b = pb; \
114 return ((*a)->color >> (8 * (2 - (pos))) & 0xff) \
115 - ((*b)->color >> (8 * (2 - (pos))) & 0xff); \
118 DECLARE_CMP_FUNC(r, 0)
119 DECLARE_CMP_FUNC(g, 1)
120 DECLARE_CMP_FUNC(b, 2)
122 static const cmp_func cmp_funcs[] = {cmp_r, cmp_g, cmp_b};
125 * Simple color comparison for sorting the final palette
127 static int cmp_color(const void *a, const void *b)
129 const struct range_box *box1 = a;
130 const struct range_box *box2 = b;
131 return box1->color - box2->color;
134 static av_always_inline int diff(const uint32_t a, const uint32_t b)
136 const uint8_t c1[] = {a >> 16 & 0xff, a >> 8 & 0xff, a & 0xff};
137 const uint8_t c2[] = {b >> 16 & 0xff, b >> 8 & 0xff, b & 0xff};
138 const int dr = c1[0] - c2[0];
139 const int dg = c1[1] - c2[1];
140 const int db = c1[2] - c2[2];
141 return dr*dr + dg*dg + db*db;
145 * Find the next box to split: pick the one with the highest variance
147 static int get_next_box_id_to_split(PaletteGenContext *s)
149 int box_id, i, best_box_id = -1;
150 int64_t max_variance = -1;
152 if (s->nb_boxes == s->max_colors - s->reserve_transparent)
155 for (box_id = 0; box_id < s->nb_boxes; box_id++) {
156 struct range_box *box = &s->boxes[box_id];
158 if (s->boxes[box_id].len >= 2) {
160 if (box->variance == -1) {
161 int64_t variance = 0;
163 for (i = 0; i < box->len; i++) {
164 const struct color_ref *ref = s->refs[box->start + i];
165 variance += diff(ref->color, box->color) * ref->count;
167 box->variance = variance;
169 if (box->variance > max_variance) {
170 best_box_id = box_id;
171 max_variance = box->variance;
181 * Get the 32-bit average color for the range of RGB colors enclosed in the
182 * specified box. Takes into account the weight of each color.
184 static uint32_t get_avg_color(struct color_ref * const *refs,
185 const struct range_box *box)
188 const int n = box->len;
189 uint64_t r = 0, g = 0, b = 0, div = 0;
191 for (i = 0; i < n; i++) {
192 const struct color_ref *ref = refs[box->start + i];
193 r += (ref->color >> 16 & 0xff) * ref->count;
194 g += (ref->color >> 8 & 0xff) * ref->count;
195 b += (ref->color & 0xff) * ref->count;
203 return 0xffU<<24 | r<<16 | g<<8 | b;
207 * Split given box in two at position n. The original box becomes the left part
208 * of the split, and the new index box is the right part.
210 static void split_box(PaletteGenContext *s, struct range_box *box, int n)
212 struct range_box *new_box = &s->boxes[s->nb_boxes++];
213 new_box->start = n + 1;
214 new_box->len = box->start + box->len - new_box->start;
215 new_box->sorted_by = box->sorted_by;
216 box->len -= new_box->len;
218 av_assert0(box->len >= 1);
219 av_assert0(new_box->len >= 1);
221 box->color = get_avg_color(s->refs, box);
222 new_box->color = get_avg_color(s->refs, new_box);
224 new_box->variance = -1;
228 * Write the palette into the output frame.
230 static void write_palette(AVFilterContext *ctx, AVFrame *out)
232 const PaletteGenContext *s = ctx->priv;
233 int x, y, box_id = 0;
234 uint32_t *pal = (uint32_t *)out->data[0];
235 const int pal_linesize = out->linesize[0] >> 2;
236 uint32_t last_color = 0;
238 for (y = 0; y < out->height; y++) {
239 for (x = 0; x < out->width; x++) {
240 if (box_id < s->nb_boxes) {
241 pal[x] = s->boxes[box_id++].color;
242 if ((x || y) && pal[x] == last_color)
243 av_log(ctx, AV_LOG_WARNING, "Dupped color: %08X\n", pal[x]);
246 pal[x] = 0xff000000; // pad with black
252 if (s->reserve_transparent) {
253 av_assert0(s->nb_boxes < 256);
254 pal[out->width - pal_linesize - 1] = 0x0000ff00; // add a green transparent color
259 * Crawl the histogram to get all the defined colors, and create a linear list
260 * of them (each color reference entry is a pointer to the value in the
261 * histogram/hash table).
263 static struct color_ref **load_color_refs(const struct hist_node *hist, int nb_refs)
266 struct color_ref **refs = av_malloc_array(nb_refs, sizeof(*refs));
271 for (j = 0; j < HIST_SIZE; j++) {
272 const struct hist_node *node = &hist[j];
274 for (i = 0; i < node->nb_entries; i++)
275 refs[k++] = &node->entries[i];
281 static double set_colorquant_ratio_meta(AVFrame *out, int nb_out, int nb_in)
284 const double ratio = (double)nb_out / nb_in;
285 snprintf(buf, sizeof(buf), "%f", ratio);
286 av_dict_set(&out->metadata, "lavfi.color_quant_ratio", buf, 0);
291 * Main function implementing the Median Cut Algorithm defined by Paul Heckbert
292 * in Color Image Quantization for Frame Buffer Display (1982)
294 static AVFrame *get_palette_frame(AVFilterContext *ctx)
297 PaletteGenContext *s = ctx->priv;
298 AVFilterLink *outlink = ctx->outputs[0];
301 struct range_box *box;
303 /* reference only the used colors from histogram */
304 s->refs = load_color_refs(s->histogram, s->nb_refs);
306 av_log(ctx, AV_LOG_ERROR, "Unable to allocate references for %d different colors\n", s->nb_refs);
310 /* create the palette frame */
311 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
316 /* set first box for 0..nb_refs */
317 box = &s->boxes[box_id];
318 box->len = s->nb_refs;
320 box->color = get_avg_color(s->refs, box);
324 while (box && box->len > 1) {
325 int i, rr, gr, br, longest;
326 uint64_t median, box_weight = 0;
328 /* compute the box weight (sum all the weights of the colors in the
329 * range) and its boundings */
330 uint8_t min[3] = {0xff, 0xff, 0xff};
331 uint8_t max[3] = {0x00, 0x00, 0x00};
332 for (i = box->start; i < box->start + box->len; i++) {
333 const struct color_ref *ref = s->refs[i];
334 const uint32_t rgb = ref->color;
335 const uint8_t r = rgb >> 16 & 0xff, g = rgb >> 8 & 0xff, b = rgb & 0xff;
336 min[0] = FFMIN(r, min[0]), max[0] = FFMAX(r, max[0]);
337 min[1] = FFMIN(g, min[1]), max[1] = FFMAX(g, max[1]);
338 min[2] = FFMIN(b, min[2]), max[2] = FFMAX(b, max[2]);
339 box_weight += ref->count;
342 /* define the axis to sort by according to the widest range of colors */
343 rr = max[0] - min[0];
344 gr = max[1] - min[1];
345 br = max[2] - min[2];
346 longest = 1; // pick green by default (the color the eye is the most sensitive to)
347 if (br >= rr && br >= gr) longest = 2;
348 if (rr >= gr && rr >= br) longest = 0;
349 if (gr >= rr && gr >= br) longest = 1; // prefer green again
351 ff_dlog(ctx, "box #%02X [%6d..%-6d] (%6d) w:%-6"PRIu64" ranges:[%2x %2x %2x] sort by %c (already sorted:%c) ",
352 box_id, box->start, box->start + box->len - 1, box->len, box_weight,
353 rr, gr, br, "rgb"[longest], box->sorted_by == longest ? 'y':'n');
355 /* sort the range by its longest axis if it's not already sorted */
356 if (box->sorted_by != longest) {
357 cmp_func cmpf = cmp_funcs[longest];
358 AV_QSORT(&s->refs[box->start], box->len, const struct color_ref *, cmpf);
359 box->sorted_by = longest;
362 /* locate the median where to split */
363 median = (box_weight + 1) >> 1;
365 /* if you have 2 boxes, the maximum is actually #0: you must have at
366 * least 1 color on each side of the split, hence the -2 */
367 for (i = box->start; i < box->start + box->len - 2; i++) {
368 box_weight += s->refs[i]->count;
369 if (box_weight > median)
372 ff_dlog(ctx, "split @ i=%-6d with w=%-6"PRIu64" (target=%6"PRIu64")\n", i, box_weight, median);
373 split_box(s, box, i);
375 box_id = get_next_box_id_to_split(s);
376 box = box_id >= 0 ? &s->boxes[box_id] : NULL;
379 ratio = set_colorquant_ratio_meta(out, s->nb_boxes, s->nb_refs);
380 av_log(ctx, AV_LOG_INFO, "%d%s colors generated out of %d colors; ratio=%f\n",
381 s->nb_boxes, s->reserve_transparent ? "(+1)" : "", s->nb_refs, ratio);
383 qsort(s->boxes, s->nb_boxes, sizeof(*s->boxes), cmp_color);
385 write_palette(ctx, out);
391 * Hashing function for the color.
392 * It keeps the NBITS least significant bit of each component to make it
393 * "random" even if the scene doesn't have much different colors.
395 static inline unsigned color_hash(uint32_t color)
397 const uint8_t r = color >> 16 & ((1<<NBITS)-1);
398 const uint8_t g = color >> 8 & ((1<<NBITS)-1);
399 const uint8_t b = color & ((1<<NBITS)-1);
400 return r<<(NBITS*2) | g<<NBITS | b;
404 * Locate the color in the hash table and increment its counter.
406 static int color_inc(struct hist_node *hist, uint32_t color)
409 const unsigned hash = color_hash(color);
410 struct hist_node *node = &hist[hash];
413 for (i = 0; i < node->nb_entries; i++) {
414 e = &node->entries[i];
415 if (e->color == color) {
421 e = av_dynarray2_add((void**)&node->entries, &node->nb_entries,
422 sizeof(*node->entries), NULL);
424 return AVERROR(ENOMEM);
431 * Update histogram when pixels differ from previous frame.
433 static int update_histogram_diff(struct hist_node *hist,
434 const AVFrame *f1, const AVFrame *f2)
436 int x, y, ret, nb_diff_colors = 0;
438 for (y = 0; y < f1->height; y++) {
439 const uint32_t *p = (const uint32_t *)(f1->data[0] + y*f1->linesize[0]);
440 const uint32_t *q = (const uint32_t *)(f2->data[0] + y*f2->linesize[0]);
442 for (x = 0; x < f1->width; x++) {
445 ret = color_inc(hist, p[x]);
448 nb_diff_colors += ret;
451 return nb_diff_colors;
455 * Simple histogram of the frame.
457 static int update_histogram_frame(struct hist_node *hist, const AVFrame *f)
459 int x, y, ret, nb_diff_colors = 0;
461 for (y = 0; y < f->height; y++) {
462 const uint32_t *p = (const uint32_t *)(f->data[0] + y*f->linesize[0]);
464 for (x = 0; x < f->width; x++) {
465 ret = color_inc(hist, p[x]);
468 nb_diff_colors += ret;
471 return nb_diff_colors;
475 * Update the histogram for each passing frame. No frame will be pushed here.
477 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
479 AVFilterContext *ctx = inlink->dst;
480 PaletteGenContext *s = ctx->priv;
481 const int ret = s->prev_frame ? update_histogram_diff(s->histogram, s->prev_frame, in)
482 : update_histogram_frame(s->histogram, in);
487 if (s->stats_mode == STATS_MODE_DIFF_FRAMES) {
488 av_frame_free(&s->prev_frame);
498 * Returns only one frame at the end containing the full palette.
500 static int request_frame(AVFilterLink *outlink)
502 AVFilterContext *ctx = outlink->src;
503 AVFilterLink *inlink = ctx->inputs[0];
504 PaletteGenContext *s = ctx->priv;
507 r = ff_request_frame(inlink);
508 if (r == AVERROR_EOF && !s->palette_pushed && s->nb_refs) {
509 r = ff_filter_frame(outlink, get_palette_frame(ctx));
510 s->palette_pushed = 1;
517 * The output is one simple 16x16 squared-pixels palette.
519 static int config_output(AVFilterLink *outlink)
521 outlink->w = outlink->h = 16;
522 outlink->sample_aspect_ratio = av_make_q(1, 1);
523 outlink->flags |= FF_LINK_FLAG_REQUEST_LOOP;
527 static av_cold void uninit(AVFilterContext *ctx)
530 PaletteGenContext *s = ctx->priv;
532 for (i = 0; i < HIST_SIZE; i++)
533 av_freep(&s->histogram[i].entries);
535 av_frame_free(&s->prev_frame);
538 static const AVFilterPad palettegen_inputs[] = {
541 .type = AVMEDIA_TYPE_VIDEO,
542 .filter_frame = filter_frame,
547 static const AVFilterPad palettegen_outputs[] = {
550 .type = AVMEDIA_TYPE_VIDEO,
551 .config_props = config_output,
552 .request_frame = request_frame,
557 AVFilter ff_vf_palettegen = {
558 .name = "palettegen",
559 .description = NULL_IF_CONFIG_SMALL("Find the optimal palette for a given stream."),
560 .priv_size = sizeof(PaletteGenContext),
562 .query_formats = query_formats,
563 .inputs = palettegen_inputs,
564 .outputs = palettegen_outputs,
565 .priv_class = &palettegen_class,