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 * Use a palette to downsample an input video stream.
26 #include "libavutil/bprint.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/opt.h"
29 #include "libavutil/qsort.h"
32 #include "framesync.h"
39 DITHERING_FLOYD_STEINBERG,
45 enum color_search_method {
46 COLOR_SEARCH_NNS_ITERATIVE,
47 COLOR_SEARCH_NNS_RECURSIVE,
48 COLOR_SEARCH_BRUTEFORCE,
62 int left_id, right_id;
66 #define CACHE_SIZE (1<<(3*NBITS))
74 struct cached_color *entries;
78 struct PaletteUseContext;
80 typedef int (*set_frame_func)(struct PaletteUseContext *s, AVFrame *out, AVFrame *in,
81 int x_start, int y_start, int width, int height);
83 typedef struct PaletteUseContext {
86 struct cache_node cache[CACHE_SIZE]; /* lookup cache */
87 struct color_node map[AVPALETTE_COUNT]; /* 3D-Tree (KD-Tree with K=3) for reverse colormap */
88 uint32_t palette[AVPALETTE_COUNT];
89 int transparency_index; /* index in the palette of transparency. -1 if there is no transparency in the palette. */
94 set_frame_func set_frame;
96 int ordered_dither[8*8];
103 int color_search_method;
105 uint64_t total_mean_err;
109 #define OFFSET(x) offsetof(PaletteUseContext, x)
110 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
111 static const AVOption paletteuse_options[] = {
112 { "dither", "select dithering mode", OFFSET(dither), AV_OPT_TYPE_INT, {.i64=DITHERING_SIERRA2_4A}, 0, NB_DITHERING-1, FLAGS, "dithering_mode" },
113 { "bayer", "ordered 8x8 bayer dithering (deterministic)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_BAYER}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
114 { "heckbert", "dithering as defined by Paul Heckbert in 1982 (simple error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_HECKBERT}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
115 { "floyd_steinberg", "Floyd and Steingberg dithering (error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_FLOYD_STEINBERG}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
116 { "sierra2", "Frankie Sierra dithering v2 (error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_SIERRA2}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
117 { "sierra2_4a", "Frankie Sierra dithering v2 \"Lite\" (error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_SIERRA2_4A}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
118 { "bayer_scale", "set scale for bayer dithering", OFFSET(bayer_scale), AV_OPT_TYPE_INT, {.i64=2}, 0, 5, FLAGS },
119 { "diff_mode", "set frame difference mode", OFFSET(diff_mode), AV_OPT_TYPE_INT, {.i64=DIFF_MODE_NONE}, 0, NB_DIFF_MODE-1, FLAGS, "diff_mode" },
120 { "rectangle", "process smallest different rectangle", 0, AV_OPT_TYPE_CONST, {.i64=DIFF_MODE_RECTANGLE}, INT_MIN, INT_MAX, FLAGS, "diff_mode" },
121 { "new", "take new palette for each output frame", OFFSET(new), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
122 { "alpha_threshold", "set the alpha threshold for transparency", OFFSET(trans_thresh), AV_OPT_TYPE_INT, {.i64=128}, 0, 255, FLAGS },
124 /* following are the debug options, not part of the official API */
125 { "debug_kdtree", "save Graphviz graph of the kdtree in specified file", OFFSET(dot_filename), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
126 { "color_search", "set reverse colormap color search method", OFFSET(color_search_method), AV_OPT_TYPE_INT, {.i64=COLOR_SEARCH_NNS_ITERATIVE}, 0, NB_COLOR_SEARCHES-1, FLAGS, "search" },
127 { "nns_iterative", "iterative search", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_NNS_ITERATIVE}, INT_MIN, INT_MAX, FLAGS, "search" },
128 { "nns_recursive", "recursive search", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_NNS_RECURSIVE}, INT_MIN, INT_MAX, FLAGS, "search" },
129 { "bruteforce", "brute-force into the palette", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_BRUTEFORCE}, INT_MIN, INT_MAX, FLAGS, "search" },
130 { "mean_err", "compute and print mean error", OFFSET(calc_mean_err), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
131 { "debug_accuracy", "test color search accuracy", OFFSET(debug_accuracy), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
135 AVFILTER_DEFINE_CLASS(paletteuse);
137 static int load_apply_palette(FFFrameSync *fs);
139 static int query_formats(AVFilterContext *ctx)
141 static const enum AVPixelFormat in_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
142 static const enum AVPixelFormat inpal_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
143 static const enum AVPixelFormat out_fmts[] = {AV_PIX_FMT_PAL8, AV_PIX_FMT_NONE};
145 if ((ret = ff_formats_ref(ff_make_format_list(in_fmts),
146 &ctx->inputs[0]->outcfg.formats)) < 0 ||
147 (ret = ff_formats_ref(ff_make_format_list(inpal_fmts),
148 &ctx->inputs[1]->outcfg.formats)) < 0 ||
149 (ret = ff_formats_ref(ff_make_format_list(out_fmts),
150 &ctx->outputs[0]->incfg.formats)) < 0)
155 static av_always_inline int dither_color(uint32_t px, int er, int eg, int eb, int scale, int shift)
157 return av_clip_uint8( px >> 24 ) << 24
158 | av_clip_uint8((px >> 16 & 0xff) + ((er * scale) / (1<<shift))) << 16
159 | av_clip_uint8((px >> 8 & 0xff) + ((eg * scale) / (1<<shift))) << 8
160 | av_clip_uint8((px & 0xff) + ((eb * scale) / (1<<shift)));
163 static av_always_inline int diff(const uint8_t *c1, const uint8_t *c2, const int trans_thresh)
165 // XXX: try L*a*b with CIE76 (dL*dL + da*da + db*db)
166 const int dr = c1[1] - c2[1];
167 const int dg = c1[2] - c2[2];
168 const int db = c1[3] - c2[3];
170 if (c1[0] < trans_thresh && c2[0] < trans_thresh) {
172 } else if (c1[0] >= trans_thresh && c2[0] >= trans_thresh) {
173 return dr*dr + dg*dg + db*db;
175 return 255*255 + 255*255 + 255*255;
179 static av_always_inline uint8_t colormap_nearest_bruteforce(const uint32_t *palette, const uint8_t *argb, const int trans_thresh)
181 int i, pal_id = -1, min_dist = INT_MAX;
183 for (i = 0; i < AVPALETTE_COUNT; i++) {
184 const uint32_t c = palette[i];
186 if (c >> 24 >= trans_thresh) { // ignore transparent entry
187 const uint8_t palargb[] = {
188 palette[i]>>24 & 0xff,
189 palette[i]>>16 & 0xff,
190 palette[i]>> 8 & 0xff,
193 const int d = diff(palargb, argb, trans_thresh);
203 /* Recursive form, simpler but a bit slower. Kept for reference. */
204 struct nearest_color {
209 static void colormap_nearest_node(const struct color_node *map,
211 const uint8_t *target,
212 const int trans_thresh,
213 struct nearest_color *nearest)
215 const struct color_node *kd = map + node_pos;
216 const int s = kd->split;
217 int dx, nearer_kd_id, further_kd_id;
218 const uint8_t *current = kd->val;
219 const int current_to_target = diff(target, current, trans_thresh);
221 if (current_to_target < nearest->dist_sqd) {
222 nearest->node_pos = node_pos;
223 nearest->dist_sqd = current_to_target;
226 if (kd->left_id != -1 || kd->right_id != -1) {
227 dx = target[s] - current[s];
229 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
230 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
232 if (nearer_kd_id != -1)
233 colormap_nearest_node(map, nearer_kd_id, target, trans_thresh, nearest);
235 if (further_kd_id != -1 && dx*dx < nearest->dist_sqd)
236 colormap_nearest_node(map, further_kd_id, target, trans_thresh, nearest);
240 static av_always_inline uint8_t colormap_nearest_recursive(const struct color_node *node, const uint8_t *rgb, const int trans_thresh)
242 struct nearest_color res = {.dist_sqd = INT_MAX, .node_pos = -1};
243 colormap_nearest_node(node, 0, rgb, trans_thresh, &res);
244 return node[res.node_pos].palette_id;
252 static av_always_inline uint8_t colormap_nearest_iterative(const struct color_node *root, const uint8_t *target, const int trans_thresh)
254 int pos = 0, best_node_id = -1, best_dist = INT_MAX, cur_color_id = 0;
255 struct stack_node nodes[16];
256 struct stack_node *node = &nodes[0];
260 const struct color_node *kd = &root[cur_color_id];
261 const uint8_t *current = kd->val;
262 const int current_to_target = diff(target, current, trans_thresh);
264 /* Compare current color node to the target and update our best node if
265 * it's actually better. */
266 if (current_to_target < best_dist) {
267 best_node_id = cur_color_id;
268 if (!current_to_target)
269 goto end; // exact match, we can return immediately
270 best_dist = current_to_target;
273 /* Check if it's not a leaf */
274 if (kd->left_id != -1 || kd->right_id != -1) {
275 const int split = kd->split;
276 const int dx = target[split] - current[split];
277 int nearer_kd_id, further_kd_id;
279 /* Define which side is the most interesting. */
280 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
281 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
283 if (nearer_kd_id != -1) {
284 if (further_kd_id != -1) {
285 /* Here, both paths are defined, so we push a state for
286 * when we are going back. */
287 node->color_id = further_kd_id;
292 /* We can now update current color with the most probable path
293 * (no need to create a state since there is nothing to save
295 cur_color_id = nearer_kd_id;
297 } else if (dx*dx < best_dist) {
298 /* The nearest path isn't available, so there is only one path
299 * possible and it's the least probable. We enter it only if the
300 * distance from the current point to the hyper rectangle is
301 * less than our best distance. */
302 cur_color_id = further_kd_id;
307 /* Unstack as much as we can, typically as long as the least probable
308 * branch aren't actually probable. */
313 } while (node->dx2 >= best_dist);
315 /* We got a node where the least probable branch might actually contain
316 * a relevant color. */
317 cur_color_id = node->color_id;
321 return root[best_node_id].palette_id;
324 #define COLORMAP_NEAREST(search, palette, root, target, trans_thresh) \
325 search == COLOR_SEARCH_NNS_ITERATIVE ? colormap_nearest_iterative(root, target, trans_thresh) : \
326 search == COLOR_SEARCH_NNS_RECURSIVE ? colormap_nearest_recursive(root, target, trans_thresh) : \
327 colormap_nearest_bruteforce(palette, target, trans_thresh)
330 * Check if the requested color is in the cache already. If not, find it in the
331 * color tree and cache it.
332 * Note: a, r, g, and b are the components of color, but are passed as well to avoid
333 * recomputing them (they are generally computed by the caller for other uses).
335 static av_always_inline int color_get(PaletteUseContext *s, uint32_t color,
336 uint8_t a, uint8_t r, uint8_t g, uint8_t b,
337 const enum color_search_method search_method)
340 const uint8_t argb_elts[] = {a, r, g, b};
341 const uint8_t rhash = r & ((1<<NBITS)-1);
342 const uint8_t ghash = g & ((1<<NBITS)-1);
343 const uint8_t bhash = b & ((1<<NBITS)-1);
344 const unsigned hash = rhash<<(NBITS*2) | ghash<<NBITS | bhash;
345 struct cache_node *node = &s->cache[hash];
346 struct cached_color *e;
348 // first, check for transparency
349 if (a < s->trans_thresh && s->transparency_index >= 0) {
350 return s->transparency_index;
353 for (i = 0; i < node->nb_entries; i++) {
354 e = &node->entries[i];
355 if (e->color == color)
359 e = av_dynarray2_add((void**)&node->entries, &node->nb_entries,
360 sizeof(*node->entries), NULL);
362 return AVERROR(ENOMEM);
364 e->pal_entry = COLORMAP_NEAREST(search_method, s->palette, s->map, argb_elts, s->trans_thresh);
369 static av_always_inline int get_dst_color_err(PaletteUseContext *s,
370 uint32_t c, int *er, int *eg, int *eb,
371 const enum color_search_method search_method)
373 const uint8_t a = c >> 24 & 0xff;
374 const uint8_t r = c >> 16 & 0xff;
375 const uint8_t g = c >> 8 & 0xff;
376 const uint8_t b = c & 0xff;
378 const int dstx = color_get(s, c, a, r, g, b, search_method);
381 dstc = s->palette[dstx];
382 *er = r - (dstc >> 16 & 0xff);
383 *eg = g - (dstc >> 8 & 0xff);
384 *eb = b - (dstc & 0xff);
388 static av_always_inline int set_frame(PaletteUseContext *s, AVFrame *out, AVFrame *in,
389 int x_start, int y_start, int w, int h,
390 enum dithering_mode dither,
391 const enum color_search_method search_method)
394 const int src_linesize = in ->linesize[0] >> 2;
395 const int dst_linesize = out->linesize[0];
396 uint32_t *src = ((uint32_t *)in ->data[0]) + y_start*src_linesize;
397 uint8_t *dst = out->data[0] + y_start*dst_linesize;
402 for (y = y_start; y < h; y++) {
403 for (x = x_start; x < w; x++) {
406 if (dither == DITHERING_BAYER) {
407 const int d = s->ordered_dither[(y & 7)<<3 | (x & 7)];
408 const uint8_t a8 = src[x] >> 24 & 0xff;
409 const uint8_t r8 = src[x] >> 16 & 0xff;
410 const uint8_t g8 = src[x] >> 8 & 0xff;
411 const uint8_t b8 = src[x] & 0xff;
412 const uint8_t r = av_clip_uint8(r8 + d);
413 const uint8_t g = av_clip_uint8(g8 + d);
414 const uint8_t b = av_clip_uint8(b8 + d);
415 const int color = color_get(s, src[x], a8, r, g, b, search_method);
421 } else if (dither == DITHERING_HECKBERT) {
422 const int right = x < w - 1, down = y < h - 1;
423 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
429 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 3, 3);
430 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 3, 3);
431 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 2, 3);
433 } else if (dither == DITHERING_FLOYD_STEINBERG) {
434 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
435 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
441 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 7, 4);
442 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 3, 4);
443 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 5, 4);
444 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 1, 4);
446 } else if (dither == DITHERING_SIERRA2) {
447 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
448 const int right2 = x < w - 2, left2 = x > x_start + 1;
449 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
455 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 4, 4);
456 if (right2) src[ x + 2] = dither_color(src[ x + 2], er, eg, eb, 3, 4);
459 if (left2) src[ src_linesize + x - 2] = dither_color(src[ src_linesize + x - 2], er, eg, eb, 1, 4);
460 if (left) src[ src_linesize + x - 1] = dither_color(src[ src_linesize + x - 1], er, eg, eb, 2, 4);
461 if (1) src[ src_linesize + x ] = dither_color(src[ src_linesize + x ], er, eg, eb, 3, 4);
462 if (right) src[ src_linesize + x + 1] = dither_color(src[ src_linesize + x + 1], er, eg, eb, 2, 4);
463 if (right2) src[ src_linesize + x + 2] = dither_color(src[ src_linesize + x + 2], er, eg, eb, 1, 4);
466 } else if (dither == DITHERING_SIERRA2_4A) {
467 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
468 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
474 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 2, 2);
475 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 1, 2);
476 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 1, 2);
479 const uint8_t a = src[x] >> 24 & 0xff;
480 const uint8_t r = src[x] >> 16 & 0xff;
481 const uint8_t g = src[x] >> 8 & 0xff;
482 const uint8_t b = src[x] & 0xff;
483 const int color = color_get(s, src[x], a, r, g, b, search_method);
497 static void disp_node(AVBPrint *buf,
498 const struct color_node *map,
499 int parent_id, int node_id,
502 const struct color_node *node = &map[node_id];
503 const uint32_t fontcolor = node->val[1] > 0x50 &&
504 node->val[2] > 0x50 &&
505 node->val[3] > 0x50 ? 0 : 0xffffff;
506 const int rgb_comp = node->split - 1;
507 av_bprintf(buf, "%*cnode%d ["
508 "label=\"%c%02X%c%02X%c%02X%c\" "
509 "fillcolor=\"#%02x%02x%02x\" "
510 "fontcolor=\"#%06"PRIX32"\"]\n",
511 depth*INDENT, ' ', node->palette_id,
512 "[ "[rgb_comp], node->val[1],
513 "][ "[rgb_comp], node->val[2],
514 " ]["[rgb_comp], node->val[3],
516 node->val[1], node->val[2], node->val[3],
519 av_bprintf(buf, "%*cnode%d -> node%d\n", depth*INDENT, ' ',
520 map[parent_id].palette_id, node->palette_id);
521 if (node->left_id != -1) disp_node(buf, map, node_id, node->left_id, depth + 1);
522 if (node->right_id != -1) disp_node(buf, map, node_id, node->right_id, depth + 1);
525 // debug_kdtree=kdtree.dot -> dot -Tpng kdtree.dot > kdtree.png
526 static int disp_tree(const struct color_node *node, const char *fname)
529 FILE *f = av_fopen_utf8(fname, "w");
532 int ret = AVERROR(errno);
533 av_log(NULL, AV_LOG_ERROR, "Cannot open file '%s' for writing: %s\n",
534 fname, av_err2str(ret));
538 av_bprint_init(&buf, 0, AV_BPRINT_SIZE_UNLIMITED);
540 av_bprintf(&buf, "digraph {\n");
541 av_bprintf(&buf, " node [style=filled fontsize=10 shape=box]\n");
542 disp_node(&buf, node, -1, 0, 0);
543 av_bprintf(&buf, "}\n");
545 fwrite(buf.str, 1, buf.len, f);
547 av_bprint_finalize(&buf, NULL);
551 static int debug_accuracy(const struct color_node *node, const uint32_t *palette, const int trans_thresh,
552 const enum color_search_method search_method)
554 int r, g, b, ret = 0;
556 for (r = 0; r < 256; r++) {
557 for (g = 0; g < 256; g++) {
558 for (b = 0; b < 256; b++) {
559 const uint8_t argb[] = {0xff, r, g, b};
560 const int r1 = COLORMAP_NEAREST(search_method, palette, node, argb, trans_thresh);
561 const int r2 = colormap_nearest_bruteforce(palette, argb, trans_thresh);
563 const uint32_t c1 = palette[r1];
564 const uint32_t c2 = palette[r2];
565 const uint8_t palargb1[] = { 0xff, c1>>16 & 0xff, c1>> 8 & 0xff, c1 & 0xff };
566 const uint8_t palargb2[] = { 0xff, c2>>16 & 0xff, c2>> 8 & 0xff, c2 & 0xff };
567 const int d1 = diff(palargb1, argb, trans_thresh);
568 const int d2 = diff(palargb2, argb, trans_thresh);
570 av_log(NULL, AV_LOG_ERROR,
571 "/!\\ %02X%02X%02X: %d ! %d (%06"PRIX32" ! %06"PRIX32") / dist: %d ! %d\n",
572 r, g, b, r1, r2, c1 & 0xffffff, c2 & 0xffffff, d1, d2);
592 typedef int (*cmp_func)(const void *, const void *);
594 #define DECLARE_CMP_FUNC(name, pos) \
595 static int cmp_##name(const void *pa, const void *pb) \
597 const struct color *a = pa; \
598 const struct color *b = pb; \
599 return (a->value >> (8 * (3 - (pos))) & 0xff) \
600 - (b->value >> (8 * (3 - (pos))) & 0xff); \
603 DECLARE_CMP_FUNC(a, 0)
604 DECLARE_CMP_FUNC(r, 1)
605 DECLARE_CMP_FUNC(g, 2)
606 DECLARE_CMP_FUNC(b, 3)
608 static const cmp_func cmp_funcs[] = {cmp_a, cmp_r, cmp_g, cmp_b};
610 static int get_next_color(const uint8_t *color_used, const uint32_t *palette,
611 const int trans_thresh,
612 int *component, const struct color_rect *box)
616 unsigned nb_color = 0;
617 struct color_rect ranges;
618 struct color tmp_pal[256];
621 ranges.min[0] = ranges.min[1] = ranges.min[2] = 0xff;
622 ranges.max[0] = ranges.max[1] = ranges.max[2] = 0x00;
624 for (i = 0; i < AVPALETTE_COUNT; i++) {
625 const uint32_t c = palette[i];
626 const uint8_t a = c >> 24 & 0xff;
627 const uint8_t r = c >> 16 & 0xff;
628 const uint8_t g = c >> 8 & 0xff;
629 const uint8_t b = c & 0xff;
631 if (a < trans_thresh) {
635 if (color_used[i] || (a != 0xff) ||
636 r < box->min[0] || g < box->min[1] || b < box->min[2] ||
637 r > box->max[0] || g > box->max[1] || b > box->max[2])
640 if (r < ranges.min[0]) ranges.min[0] = r;
641 if (g < ranges.min[1]) ranges.min[1] = g;
642 if (b < ranges.min[2]) ranges.min[2] = b;
644 if (r > ranges.max[0]) ranges.max[0] = r;
645 if (g > ranges.max[1]) ranges.max[1] = g;
646 if (b > ranges.max[2]) ranges.max[2] = b;
648 tmp_pal[nb_color].value = c;
649 tmp_pal[nb_color].pal_id = i;
657 /* define longest axis that will be the split component */
658 wr = ranges.max[0] - ranges.min[0];
659 wg = ranges.max[1] - ranges.min[1];
660 wb = ranges.max[2] - ranges.min[2];
661 if (wr >= wg && wr >= wb) longest = 1;
662 if (wg >= wr && wg >= wb) longest = 2;
663 if (wb >= wr && wb >= wg) longest = 3;
664 cmpf = cmp_funcs[longest];
665 *component = longest;
667 /* sort along this axis to get median */
668 AV_QSORT(tmp_pal, nb_color, struct color, cmpf);
670 return tmp_pal[nb_color >> 1].pal_id;
673 static int colormap_insert(struct color_node *map,
676 const uint32_t *palette,
677 const int trans_thresh,
678 const struct color_rect *box)
681 int component, cur_id;
682 int node_left_id = -1, node_right_id = -1;
683 struct color_node *node;
684 struct color_rect box1, box2;
685 const int pal_id = get_next_color(color_used, palette, trans_thresh, &component, box);
690 /* create new node with that color */
691 cur_id = (*nb_used)++;
694 node->split = component;
695 node->palette_id = pal_id;
696 node->val[0] = c>>24 & 0xff;
697 node->val[1] = c>>16 & 0xff;
698 node->val[2] = c>> 8 & 0xff;
699 node->val[3] = c & 0xff;
701 color_used[pal_id] = 1;
703 /* get the two boxes this node creates */
705 box1.max[component-1] = node->val[component];
706 box2.min[component-1] = node->val[component] + 1;
708 node_left_id = colormap_insert(map, color_used, nb_used, palette, trans_thresh, &box1);
710 if (box2.min[component-1] <= box2.max[component-1])
711 node_right_id = colormap_insert(map, color_used, nb_used, palette, trans_thresh, &box2);
713 node->left_id = node_left_id;
714 node->right_id = node_right_id;
719 static int cmp_pal_entry(const void *a, const void *b)
721 const int c1 = *(const uint32_t *)a & 0xffffff;
722 const int c2 = *(const uint32_t *)b & 0xffffff;
726 static void load_colormap(PaletteUseContext *s)
729 uint8_t color_used[AVPALETTE_COUNT] = {0};
730 uint32_t last_color = 0;
731 struct color_rect box;
733 /* disable transparent colors and dups */
734 qsort(s->palette, AVPALETTE_COUNT, sizeof(*s->palette), cmp_pal_entry);
735 // update transparency index:
736 if (s->transparency_index >= 0) {
737 for (i = 0; i < AVPALETTE_COUNT; i++) {
738 if ((s->palette[i]>>24 & 0xff) == 0) {
739 s->transparency_index = i; // we are assuming at most one transparent color in palette
745 for (i = 0; i < AVPALETTE_COUNT; i++) {
746 const uint32_t c = s->palette[i];
747 if (i != 0 && c == last_color) {
752 if (c >> 24 < s->trans_thresh) {
753 color_used[i] = 1; // ignore transparent color(s)
758 box.min[0] = box.min[1] = box.min[2] = 0x00;
759 box.max[0] = box.max[1] = box.max[2] = 0xff;
761 colormap_insert(s->map, color_used, &nb_used, s->palette, s->trans_thresh, &box);
764 disp_tree(s->map, s->dot_filename);
766 if (s->debug_accuracy) {
767 if (!debug_accuracy(s->map, s->palette, s->trans_thresh, s->color_search_method))
768 av_log(NULL, AV_LOG_INFO, "Accuracy check passed\n");
772 static void debug_mean_error(PaletteUseContext *s, const AVFrame *in1,
773 const AVFrame *in2, int frame_count)
776 const uint32_t *palette = s->palette;
777 uint32_t *src1 = (uint32_t *)in1->data[0];
778 uint8_t *src2 = in2->data[0];
779 const int src1_linesize = in1->linesize[0] >> 2;
780 const int src2_linesize = in2->linesize[0];
781 const float div = in1->width * in1->height * 3;
782 unsigned mean_err = 0;
784 for (y = 0; y < in1->height; y++) {
785 for (x = 0; x < in1->width; x++) {
786 const uint32_t c1 = src1[x];
787 const uint32_t c2 = palette[src2[x]];
788 const uint8_t argb1[] = {0xff, c1 >> 16 & 0xff, c1 >> 8 & 0xff, c1 & 0xff};
789 const uint8_t argb2[] = {0xff, c2 >> 16 & 0xff, c2 >> 8 & 0xff, c2 & 0xff};
790 mean_err += diff(argb1, argb2, s->trans_thresh);
792 src1 += src1_linesize;
793 src2 += src2_linesize;
796 s->total_mean_err += mean_err;
798 av_log(NULL, AV_LOG_INFO, "MEP:%.3f TotalMEP:%.3f\n",
799 mean_err / div, s->total_mean_err / (div * frame_count));
802 static void set_processing_window(enum diff_mode diff_mode,
803 const AVFrame *prv_src, const AVFrame *cur_src,
804 const AVFrame *prv_dst, AVFrame *cur_dst,
805 int *xp, int *yp, int *wp, int *hp)
807 int x_start = 0, y_start = 0;
808 int width = cur_src->width;
809 int height = cur_src->height;
811 if (prv_src->data[0] && diff_mode == DIFF_MODE_RECTANGLE) {
813 int x_end = cur_src->width - 1,
814 y_end = cur_src->height - 1;
815 const uint32_t *prv_srcp = (const uint32_t *)prv_src->data[0];
816 const uint32_t *cur_srcp = (const uint32_t *)cur_src->data[0];
817 const uint8_t *prv_dstp = prv_dst->data[0];
818 uint8_t *cur_dstp = cur_dst->data[0];
820 const int prv_src_linesize = prv_src->linesize[0] >> 2;
821 const int cur_src_linesize = cur_src->linesize[0] >> 2;
822 const int prv_dst_linesize = prv_dst->linesize[0];
823 const int cur_dst_linesize = cur_dst->linesize[0];
825 /* skip common lines */
826 while (y_start < y_end && !memcmp(prv_srcp + y_start*prv_src_linesize,
827 cur_srcp + y_start*cur_src_linesize,
828 cur_src->width * 4)) {
829 memcpy(cur_dstp + y_start*cur_dst_linesize,
830 prv_dstp + y_start*prv_dst_linesize,
834 while (y_end > y_start && !memcmp(prv_srcp + y_end*prv_src_linesize,
835 cur_srcp + y_end*cur_src_linesize,
836 cur_src->width * 4)) {
837 memcpy(cur_dstp + y_end*cur_dst_linesize,
838 prv_dstp + y_end*prv_dst_linesize,
843 height = y_end + 1 - y_start;
845 /* skip common columns */
846 while (x_start < x_end) {
848 for (y = y_start; y <= y_end; y++) {
849 if (prv_srcp[y*prv_src_linesize + x_start] != cur_srcp[y*cur_src_linesize + x_start]) {
858 while (x_end > x_start) {
860 for (y = y_start; y <= y_end; y++) {
861 if (prv_srcp[y*prv_src_linesize + x_end] != cur_srcp[y*cur_src_linesize + x_end]) {
870 width = x_end + 1 - x_start;
873 for (y = y_start; y <= y_end; y++)
874 memcpy(cur_dstp + y*cur_dst_linesize,
875 prv_dstp + y*prv_dst_linesize, x_start);
877 if (x_end != cur_src->width - 1) {
878 const int copy_len = cur_src->width - 1 - x_end;
879 for (y = y_start; y <= y_end; y++)
880 memcpy(cur_dstp + y*cur_dst_linesize + x_end + 1,
881 prv_dstp + y*prv_dst_linesize + x_end + 1,
891 static int apply_palette(AVFilterLink *inlink, AVFrame *in, AVFrame **outf)
894 AVFilterContext *ctx = inlink->dst;
895 PaletteUseContext *s = ctx->priv;
896 AVFilterLink *outlink = inlink->dst->outputs[0];
898 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
901 return AVERROR(ENOMEM);
903 av_frame_copy_props(out, in);
905 set_processing_window(s->diff_mode, s->last_in, in,
906 s->last_out, out, &x, &y, &w, &h);
907 av_frame_unref(s->last_in);
908 av_frame_unref(s->last_out);
909 if ((ret = av_frame_ref(s->last_in, in)) < 0 ||
910 (ret = av_frame_ref(s->last_out, out)) < 0 ||
911 (ret = av_frame_make_writable(s->last_in)) < 0) {
917 ff_dlog(ctx, "%dx%d rect: (%d;%d) -> (%d,%d) [area:%dx%d]\n",
918 w, h, x, y, x+w, y+h, in->width, in->height);
920 ret = s->set_frame(s, out, in, x, y, w, h);
926 memcpy(out->data[1], s->palette, AVPALETTE_SIZE);
927 if (s->calc_mean_err)
928 debug_mean_error(s, in, out, inlink->frame_count_out);
933 static int config_output(AVFilterLink *outlink)
936 AVFilterContext *ctx = outlink->src;
937 PaletteUseContext *s = ctx->priv;
939 ret = ff_framesync_init_dualinput(&s->fs, ctx);
942 s->fs.opt_repeatlast = 1; // only 1 frame in the palette
943 s->fs.in[1].before = s->fs.in[1].after = EXT_INFINITY;
944 s->fs.on_event = load_apply_palette;
946 outlink->w = ctx->inputs[0]->w;
947 outlink->h = ctx->inputs[0]->h;
949 outlink->time_base = ctx->inputs[0]->time_base;
950 if ((ret = ff_framesync_configure(&s->fs)) < 0)
955 static int config_input_palette(AVFilterLink *inlink)
957 AVFilterContext *ctx = inlink->dst;
959 if (inlink->w * inlink->h != AVPALETTE_COUNT) {
960 av_log(ctx, AV_LOG_ERROR,
961 "Palette input must contain exactly %d pixels. "
962 "Specified input has %dx%d=%d pixels\n",
963 AVPALETTE_COUNT, inlink->w, inlink->h,
964 inlink->w * inlink->h);
965 return AVERROR(EINVAL);
970 static void load_palette(PaletteUseContext *s, const AVFrame *palette_frame)
973 const uint32_t *p = (const uint32_t *)palette_frame->data[0];
974 const int p_linesize = palette_frame->linesize[0] >> 2;
976 s->transparency_index = -1;
979 memset(s->palette, 0, sizeof(s->palette));
980 memset(s->map, 0, sizeof(s->map));
981 for (i = 0; i < CACHE_SIZE; i++)
982 av_freep(&s->cache[i].entries);
983 memset(s->cache, 0, sizeof(s->cache));
987 for (y = 0; y < palette_frame->height; y++) {
988 for (x = 0; x < palette_frame->width; x++) {
989 s->palette[i] = p[x];
990 if (p[x]>>24 < s->trans_thresh) {
991 s->transparency_index = i; // we are assuming at most one transparent color in palette
1001 s->palette_loaded = 1;
1004 static int load_apply_palette(FFFrameSync *fs)
1006 AVFilterContext *ctx = fs->parent;
1007 AVFilterLink *inlink = ctx->inputs[0];
1008 PaletteUseContext *s = ctx->priv;
1009 AVFrame *master, *second, *out = NULL;
1012 // writable for error diffusal dithering
1013 ret = ff_framesync_dualinput_get_writable(fs, &master, &second);
1016 if (!master || !second) {
1017 av_frame_free(&master);
1020 if (!s->palette_loaded) {
1021 load_palette(s, second);
1023 ret = apply_palette(inlink, master, &out);
1024 av_frame_free(&master);
1027 return ff_filter_frame(ctx->outputs[0], out);
1030 #define DEFINE_SET_FRAME(color_search, name, value) \
1031 static int set_frame_##name(PaletteUseContext *s, AVFrame *out, AVFrame *in, \
1032 int x_start, int y_start, int w, int h) \
1034 return set_frame(s, out, in, x_start, y_start, w, h, value, color_search); \
1037 #define DEFINE_SET_FRAME_COLOR_SEARCH(color_search, color_search_macro) \
1038 DEFINE_SET_FRAME(color_search_macro, color_search##_##none, DITHERING_NONE) \
1039 DEFINE_SET_FRAME(color_search_macro, color_search##_##bayer, DITHERING_BAYER) \
1040 DEFINE_SET_FRAME(color_search_macro, color_search##_##heckbert, DITHERING_HECKBERT) \
1041 DEFINE_SET_FRAME(color_search_macro, color_search##_##floyd_steinberg, DITHERING_FLOYD_STEINBERG) \
1042 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2, DITHERING_SIERRA2) \
1043 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2_4a, DITHERING_SIERRA2_4A) \
1045 DEFINE_SET_FRAME_COLOR_SEARCH(nns_iterative, COLOR_SEARCH_NNS_ITERATIVE)
1046 DEFINE_SET_FRAME_COLOR_SEARCH(nns_recursive, COLOR_SEARCH_NNS_RECURSIVE)
1047 DEFINE_SET_FRAME_COLOR_SEARCH(bruteforce, COLOR_SEARCH_BRUTEFORCE)
1049 #define DITHERING_ENTRIES(color_search) { \
1050 set_frame_##color_search##_none, \
1051 set_frame_##color_search##_bayer, \
1052 set_frame_##color_search##_heckbert, \
1053 set_frame_##color_search##_floyd_steinberg, \
1054 set_frame_##color_search##_sierra2, \
1055 set_frame_##color_search##_sierra2_4a, \
1058 static const set_frame_func set_frame_lut[NB_COLOR_SEARCHES][NB_DITHERING] = {
1059 DITHERING_ENTRIES(nns_iterative),
1060 DITHERING_ENTRIES(nns_recursive),
1061 DITHERING_ENTRIES(bruteforce),
1064 static int dither_value(int p)
1066 const int q = p ^ (p >> 3);
1067 return (p & 4) >> 2 | (q & 4) >> 1 \
1068 | (p & 2) << 1 | (q & 2) << 2 \
1069 | (p & 1) << 4 | (q & 1) << 5;
1072 static av_cold int init(AVFilterContext *ctx)
1074 PaletteUseContext *s = ctx->priv;
1076 s->last_in = av_frame_alloc();
1077 s->last_out = av_frame_alloc();
1078 if (!s->last_in || !s->last_out) {
1079 av_frame_free(&s->last_in);
1080 av_frame_free(&s->last_out);
1081 return AVERROR(ENOMEM);
1084 s->set_frame = set_frame_lut[s->color_search_method][s->dither];
1086 if (s->dither == DITHERING_BAYER) {
1088 const int delta = 1 << (5 - s->bayer_scale); // to avoid too much luma
1090 for (i = 0; i < FF_ARRAY_ELEMS(s->ordered_dither); i++)
1091 s->ordered_dither[i] = (dither_value(i) >> s->bayer_scale) - delta;
1097 static int activate(AVFilterContext *ctx)
1099 PaletteUseContext *s = ctx->priv;
1100 return ff_framesync_activate(&s->fs);
1103 static av_cold void uninit(AVFilterContext *ctx)
1106 PaletteUseContext *s = ctx->priv;
1108 ff_framesync_uninit(&s->fs);
1109 for (i = 0; i < CACHE_SIZE; i++)
1110 av_freep(&s->cache[i].entries);
1111 av_frame_free(&s->last_in);
1112 av_frame_free(&s->last_out);
1115 static const AVFilterPad paletteuse_inputs[] = {
1118 .type = AVMEDIA_TYPE_VIDEO,
1121 .type = AVMEDIA_TYPE_VIDEO,
1122 .config_props = config_input_palette,
1127 static const AVFilterPad paletteuse_outputs[] = {
1130 .type = AVMEDIA_TYPE_VIDEO,
1131 .config_props = config_output,
1136 AVFilter ff_vf_paletteuse = {
1137 .name = "paletteuse",
1138 .description = NULL_IF_CONFIG_SMALL("Use a palette to downsample an input video stream."),
1139 .priv_size = sizeof(PaletteUseContext),
1140 .query_formats = query_formats,
1143 .activate = activate,
1144 .inputs = paletteuse_inputs,
1145 .outputs = paletteuse_outputs,
1146 .priv_class = &paletteuse_class,