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"
30 #include "libavutil/intreadwrite.h"
34 /* Reference a color and how much it's used */
40 /* Store a range of colors */
42 uint32_t color; // average color
43 int64_t variance; // overall variance of the box (how much the colors are spread)
44 int start; // index in PaletteGenContext->refs
45 int len; // number of referenced colors
46 int sorted_by; // whether range of colors is sorted by red (0), green (1) or blue (2)
50 struct color_ref *entries;
55 STATS_MODE_ALL_FRAMES,
56 STATS_MODE_DIFF_FRAMES,
57 STATS_MODE_SINGLE_FRAMES,
62 #define HIST_SIZE (1<<(3*NBITS))
64 typedef struct PaletteGenContext {
68 int reserve_transparent;
71 AVFrame *prev_frame; // previous frame used for the diff stats_mode
72 struct hist_node histogram[HIST_SIZE]; // histogram/hashtable of the colors
73 struct color_ref **refs; // references of all the colors used in the stream
74 int nb_refs; // number of color references (or number of different colors)
75 struct range_box boxes[256]; // define the segmentation of the colorspace (the final palette)
76 int nb_boxes; // number of boxes (increase will segmenting them)
77 int palette_pushed; // if the palette frame is pushed into the outlink or not
78 uint8_t transparency_color[4]; // background color for transparency
81 #define OFFSET(x) offsetof(PaletteGenContext, x)
82 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
83 static const AVOption palettegen_options[] = {
84 { "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 },
85 { "reserve_transparent", "reserve a palette entry for transparency", OFFSET(reserve_transparent), AV_OPT_TYPE_BOOL, {.i64=1}, 0, 1, FLAGS },
86 { "transparency_color", "set a background color for transparency", OFFSET(transparency_color), AV_OPT_TYPE_COLOR, {.str="lime"}, CHAR_MIN, CHAR_MAX, FLAGS },
87 { "stats_mode", "set statistics mode", OFFSET(stats_mode), AV_OPT_TYPE_INT, {.i64=STATS_MODE_ALL_FRAMES}, 0, NB_STATS_MODE-1, FLAGS, "mode" },
88 { "full", "compute full frame histograms", 0, AV_OPT_TYPE_CONST, {.i64=STATS_MODE_ALL_FRAMES}, INT_MIN, INT_MAX, FLAGS, "mode" },
89 { "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" },
90 { "single", "compute new histogram for each frame", 0, AV_OPT_TYPE_CONST, {.i64=STATS_MODE_SINGLE_FRAMES}, INT_MIN, INT_MAX, FLAGS, "mode" },
94 AVFILTER_DEFINE_CLASS(palettegen);
96 static int query_formats(AVFilterContext *ctx)
98 static const enum AVPixelFormat in_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
99 static const enum AVPixelFormat out_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
102 if ((ret = ff_formats_ref(ff_make_format_list(in_fmts) , &ctx->inputs[0]->out_formats)) < 0)
104 if ((ret = ff_formats_ref(ff_make_format_list(out_fmts), &ctx->outputs[0]->in_formats)) < 0)
109 typedef int (*cmp_func)(const void *, const void *);
111 #define DECLARE_CMP_FUNC(name, pos) \
112 static int cmp_##name(const void *pa, const void *pb) \
114 const struct color_ref * const *a = pa; \
115 const struct color_ref * const *b = pb; \
116 return ((*a)->color >> (8 * (2 - (pos))) & 0xff) \
117 - ((*b)->color >> (8 * (2 - (pos))) & 0xff); \
120 DECLARE_CMP_FUNC(r, 0)
121 DECLARE_CMP_FUNC(g, 1)
122 DECLARE_CMP_FUNC(b, 2)
124 static const cmp_func cmp_funcs[] = {cmp_r, cmp_g, cmp_b};
127 * Simple color comparison for sorting the final palette
129 static int cmp_color(const void *a, const void *b)
131 const struct range_box *box1 = a;
132 const struct range_box *box2 = b;
133 return FFDIFFSIGN(box1->color , box2->color);
136 static av_always_inline int diff(const uint32_t a, const uint32_t b)
138 const uint8_t c1[] = {a >> 16 & 0xff, a >> 8 & 0xff, a & 0xff};
139 const uint8_t c2[] = {b >> 16 & 0xff, b >> 8 & 0xff, b & 0xff};
140 const int dr = c1[0] - c2[0];
141 const int dg = c1[1] - c2[1];
142 const int db = c1[2] - c2[2];
143 return dr*dr + dg*dg + db*db;
147 * Find the next box to split: pick the one with the highest variance
149 static int get_next_box_id_to_split(PaletteGenContext *s)
151 int box_id, i, best_box_id = -1;
152 int64_t max_variance = -1;
154 if (s->nb_boxes == s->max_colors - s->reserve_transparent)
157 for (box_id = 0; box_id < s->nb_boxes; box_id++) {
158 struct range_box *box = &s->boxes[box_id];
160 if (s->boxes[box_id].len >= 2) {
162 if (box->variance == -1) {
163 int64_t variance = 0;
165 for (i = 0; i < box->len; i++) {
166 const struct color_ref *ref = s->refs[box->start + i];
167 variance += diff(ref->color, box->color) * ref->count;
169 box->variance = variance;
171 if (box->variance > max_variance) {
172 best_box_id = box_id;
173 max_variance = box->variance;
183 * Get the 32-bit average color for the range of RGB colors enclosed in the
184 * specified box. Takes into account the weight of each color.
186 static uint32_t get_avg_color(struct color_ref * const *refs,
187 const struct range_box *box)
190 const int n = box->len;
191 uint64_t r = 0, g = 0, b = 0, div = 0;
193 for (i = 0; i < n; i++) {
194 const struct color_ref *ref = refs[box->start + i];
195 r += (ref->color >> 16 & 0xff) * ref->count;
196 g += (ref->color >> 8 & 0xff) * ref->count;
197 b += (ref->color & 0xff) * ref->count;
205 return 0xffU<<24 | r<<16 | g<<8 | b;
209 * Split given box in two at position n. The original box becomes the left part
210 * of the split, and the new index box is the right part.
212 static void split_box(PaletteGenContext *s, struct range_box *box, int n)
214 struct range_box *new_box = &s->boxes[s->nb_boxes++];
215 new_box->start = n + 1;
216 new_box->len = box->start + box->len - new_box->start;
217 new_box->sorted_by = box->sorted_by;
218 box->len -= new_box->len;
220 av_assert0(box->len >= 1);
221 av_assert0(new_box->len >= 1);
223 box->color = get_avg_color(s->refs, box);
224 new_box->color = get_avg_color(s->refs, new_box);
226 new_box->variance = -1;
230 * Write the palette into the output frame.
232 static void write_palette(AVFilterContext *ctx, AVFrame *out)
234 const PaletteGenContext *s = ctx->priv;
235 int x, y, box_id = 0;
236 uint32_t *pal = (uint32_t *)out->data[0];
237 const int pal_linesize = out->linesize[0] >> 2;
238 uint32_t last_color = 0;
240 for (y = 0; y < out->height; y++) {
241 for (x = 0; x < out->width; x++) {
242 if (box_id < s->nb_boxes) {
243 pal[x] = s->boxes[box_id++].color;
244 if ((x || y) && pal[x] == last_color)
245 av_log(ctx, AV_LOG_WARNING, "Dupped color: %08"PRIX32"\n", pal[x]);
248 pal[x] = 0xff000000; // pad with black
254 if (s->reserve_transparent) {
255 av_assert0(s->nb_boxes < 256);
256 pal[out->width - pal_linesize - 1] = AV_RB32(&s->transparency_color) >> 8;
261 * Crawl the histogram to get all the defined colors, and create a linear list
262 * of them (each color reference entry is a pointer to the value in the
263 * histogram/hash table).
265 static struct color_ref **load_color_refs(const struct hist_node *hist, int nb_refs)
268 struct color_ref **refs = av_malloc_array(nb_refs, sizeof(*refs));
273 for (j = 0; j < HIST_SIZE; j++) {
274 const struct hist_node *node = &hist[j];
276 for (i = 0; i < node->nb_entries; i++)
277 refs[k++] = &node->entries[i];
283 static double set_colorquant_ratio_meta(AVFrame *out, int nb_out, int nb_in)
286 const double ratio = (double)nb_out / nb_in;
287 snprintf(buf, sizeof(buf), "%f", ratio);
288 av_dict_set(&out->metadata, "lavfi.color_quant_ratio", buf, 0);
293 * Main function implementing the Median Cut Algorithm defined by Paul Heckbert
294 * in Color Image Quantization for Frame Buffer Display (1982)
296 static AVFrame *get_palette_frame(AVFilterContext *ctx)
299 PaletteGenContext *s = ctx->priv;
300 AVFilterLink *outlink = ctx->outputs[0];
303 struct range_box *box;
305 /* reference only the used colors from histogram */
306 s->refs = load_color_refs(s->histogram, s->nb_refs);
308 av_log(ctx, AV_LOG_ERROR, "Unable to allocate references for %d different colors\n", s->nb_refs);
312 /* create the palette frame */
313 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
318 /* set first box for 0..nb_refs */
319 box = &s->boxes[box_id];
320 box->len = s->nb_refs;
322 box->color = get_avg_color(s->refs, box);
326 while (box && box->len > 1) {
327 int i, rr, gr, br, longest;
328 uint64_t median, box_weight = 0;
330 /* compute the box weight (sum all the weights of the colors in the
331 * range) and its boundings */
332 uint8_t min[3] = {0xff, 0xff, 0xff};
333 uint8_t max[3] = {0x00, 0x00, 0x00};
334 for (i = box->start; i < box->start + box->len; i++) {
335 const struct color_ref *ref = s->refs[i];
336 const uint32_t rgb = ref->color;
337 const uint8_t r = rgb >> 16 & 0xff, g = rgb >> 8 & 0xff, b = rgb & 0xff;
338 min[0] = FFMIN(r, min[0]), max[0] = FFMAX(r, max[0]);
339 min[1] = FFMIN(g, min[1]), max[1] = FFMAX(g, max[1]);
340 min[2] = FFMIN(b, min[2]), max[2] = FFMAX(b, max[2]);
341 box_weight += ref->count;
344 /* define the axis to sort by according to the widest range of colors */
345 rr = max[0] - min[0];
346 gr = max[1] - min[1];
347 br = max[2] - min[2];
348 longest = 1; // pick green by default (the color the eye is the most sensitive to)
349 if (br >= rr && br >= gr) longest = 2;
350 if (rr >= gr && rr >= br) longest = 0;
351 if (gr >= rr && gr >= br) longest = 1; // prefer green again
353 ff_dlog(ctx, "box #%02X [%6d..%-6d] (%6d) w:%-6"PRIu64" ranges:[%2x %2x %2x] sort by %c (already sorted:%c) ",
354 box_id, box->start, box->start + box->len - 1, box->len, box_weight,
355 rr, gr, br, "rgb"[longest], box->sorted_by == longest ? 'y':'n');
357 /* sort the range by its longest axis if it's not already sorted */
358 if (box->sorted_by != longest) {
359 cmp_func cmpf = cmp_funcs[longest];
360 AV_QSORT(&s->refs[box->start], box->len, const struct color_ref *, cmpf);
361 box->sorted_by = longest;
364 /* locate the median where to split */
365 median = (box_weight + 1) >> 1;
367 /* if you have 2 boxes, the maximum is actually #0: you must have at
368 * least 1 color on each side of the split, hence the -2 */
369 for (i = box->start; i < box->start + box->len - 2; i++) {
370 box_weight += s->refs[i]->count;
371 if (box_weight > median)
374 ff_dlog(ctx, "split @ i=%-6d with w=%-6"PRIu64" (target=%6"PRIu64")\n", i, box_weight, median);
375 split_box(s, box, i);
377 box_id = get_next_box_id_to_split(s);
378 box = box_id >= 0 ? &s->boxes[box_id] : NULL;
381 ratio = set_colorquant_ratio_meta(out, s->nb_boxes, s->nb_refs);
382 av_log(ctx, AV_LOG_INFO, "%d%s colors generated out of %d colors; ratio=%f\n",
383 s->nb_boxes, s->reserve_transparent ? "(+1)" : "", s->nb_refs, ratio);
385 qsort(s->boxes, s->nb_boxes, sizeof(*s->boxes), cmp_color);
387 write_palette(ctx, out);
393 * Hashing function for the color.
394 * It keeps the NBITS least significant bit of each component to make it
395 * "random" even if the scene doesn't have much different colors.
397 static inline unsigned color_hash(uint32_t color)
399 const uint8_t r = color >> 16 & ((1<<NBITS)-1);
400 const uint8_t g = color >> 8 & ((1<<NBITS)-1);
401 const uint8_t b = color & ((1<<NBITS)-1);
402 return r<<(NBITS*2) | g<<NBITS | b;
406 * Locate the color in the hash table and increment its counter.
408 static int color_inc(struct hist_node *hist, uint32_t color)
411 const unsigned hash = color_hash(color);
412 struct hist_node *node = &hist[hash];
415 for (i = 0; i < node->nb_entries; i++) {
416 e = &node->entries[i];
417 if (e->color == color) {
423 e = av_dynarray2_add((void**)&node->entries, &node->nb_entries,
424 sizeof(*node->entries), NULL);
426 return AVERROR(ENOMEM);
433 * Update histogram when pixels differ from previous frame.
435 static int update_histogram_diff(struct hist_node *hist,
436 const AVFrame *f1, const AVFrame *f2)
438 int x, y, ret, nb_diff_colors = 0;
440 for (y = 0; y < f1->height; y++) {
441 const uint32_t *p = (const uint32_t *)(f1->data[0] + y*f1->linesize[0]);
442 const uint32_t *q = (const uint32_t *)(f2->data[0] + y*f2->linesize[0]);
444 for (x = 0; x < f1->width; x++) {
447 ret = color_inc(hist, p[x]);
450 nb_diff_colors += ret;
453 return nb_diff_colors;
457 * Simple histogram of the frame.
459 static int update_histogram_frame(struct hist_node *hist, const AVFrame *f)
461 int x, y, ret, nb_diff_colors = 0;
463 for (y = 0; y < f->height; y++) {
464 const uint32_t *p = (const uint32_t *)(f->data[0] + y*f->linesize[0]);
466 for (x = 0; x < f->width; x++) {
467 ret = color_inc(hist, p[x]);
470 nb_diff_colors += ret;
473 return nb_diff_colors;
477 * Update the histogram for each passing frame. No frame will be pushed here.
479 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
481 AVFilterContext *ctx = inlink->dst;
482 PaletteGenContext *s = ctx->priv;
483 int ret = s->prev_frame ? update_histogram_diff(s->histogram, s->prev_frame, in)
484 : update_histogram_frame(s->histogram, in);
489 if (s->stats_mode == STATS_MODE_DIFF_FRAMES) {
490 av_frame_free(&s->prev_frame);
492 } else if (s->stats_mode == STATS_MODE_SINGLE_FRAMES) {
496 out = get_palette_frame(ctx);
499 ret = ff_filter_frame(ctx->outputs[0], out);
500 for (i = 0; i < HIST_SIZE; i++)
501 av_freep(&s->histogram[i].entries);
505 memset(s->boxes, 0, sizeof(s->boxes));
506 memset(s->histogram, 0, sizeof(s->histogram));
515 * Returns only one frame at the end containing the full palette.
517 static int request_frame(AVFilterLink *outlink)
519 AVFilterContext *ctx = outlink->src;
520 AVFilterLink *inlink = ctx->inputs[0];
521 PaletteGenContext *s = ctx->priv;
524 r = ff_request_frame(inlink);
525 if (r == AVERROR_EOF && !s->palette_pushed && s->nb_refs && s->stats_mode != STATS_MODE_SINGLE_FRAMES) {
526 r = ff_filter_frame(outlink, get_palette_frame(ctx));
527 s->palette_pushed = 1;
534 * The output is one simple 16x16 squared-pixels palette.
536 static int config_output(AVFilterLink *outlink)
538 outlink->w = outlink->h = 16;
539 outlink->sample_aspect_ratio = av_make_q(1, 1);
543 static av_cold void uninit(AVFilterContext *ctx)
546 PaletteGenContext *s = ctx->priv;
548 for (i = 0; i < HIST_SIZE; i++)
549 av_freep(&s->histogram[i].entries);
551 av_frame_free(&s->prev_frame);
554 static const AVFilterPad palettegen_inputs[] = {
557 .type = AVMEDIA_TYPE_VIDEO,
558 .filter_frame = filter_frame,
563 static const AVFilterPad palettegen_outputs[] = {
566 .type = AVMEDIA_TYPE_VIDEO,
567 .config_props = config_output,
568 .request_frame = request_frame,
573 AVFilter ff_vf_palettegen = {
574 .name = "palettegen",
575 .description = NULL_IF_CONFIG_SMALL("Find the optimal palette for a given stream."),
576 .priv_size = sizeof(PaletteGenContext),
578 .query_formats = query_formats,
579 .inputs = palettegen_inputs,
580 .outputs = palettegen_outputs,
581 .priv_class = &palettegen_class,